Advances in Intelligent Robotics

and Collaborative Automation

RIVER PUBLISHERS SERIES IN AUTOMATION,
CONTROL AND ROBOTICS
Volume 1

Series Editors
Tarek Sobh André Veltman
University of Bridgeport, PIAK and TU Eindhoven,
USA The Netherlands

The “River Publishers Series in Automation, Control and Robotics” is a series of

comprehensive academic and professional books which focus on the theory and
applications of automation, control and robotics. The series focuses on topics ranging
from the theory and use of control systems, automation engineering, robotics and
intelligent machines.
Books published in the series include research monographs, edited volumes,
handbooks and textbooks. The books provide professionals, researchers, educators,
and advanced students in the field with an invaluable insight into the latest research
and developments.
Topics covered in the series include, but are by no means restricted to the
following:

• Robots and Intelligent Machines

• Robotics
• Control Systems
• Control Theory
• Automation Engineering

For a list of other books in this series, visit www.riverpublishers.com

http://www.riverpublishers.com/series.php?msg=Automation, Control and Robotics
The NEC and You Perfect
Advances Together: Robotics
in Intelligent
A Comprehensive Study of the
and Collaborative Automation
National Electrical Code

Editors

Richard Duro
Yuriy Kondratenko
Gregory P. Bierals
Electrical Design Institute, USA

River Publishers
 Published 2015 by River Publishers
River Publishers
Alsbjergvej 10, 9260 Gistrup, Denmark
www.riverpublishers.com

Distributed exclusively by Routledge

4 Park Square, Milton Park, Abingdon, Oxon OX14 4RN
605 Third Avenue, New York, NY 10017, USA

Advances in Intelligent Robotics and Collaborative Automation / by Richard Duro, Yuriy

Kondratenko.

© River Publishers 2015. This book is published open access.

Open Access
This book is distributed under the terms of the Creative Commons Attribution-Non-Commercial
4.0 International License, CC-BY-NC 4.0) (http://creativecommons.org/licenses/by/4.0/), which
permits use, duplication, adaptation, distribution and reproduction in any medium or format, as
long as you give appropriate credit to the original author(s) and the source, a link is provided to
the Creative Commons license and any changes made are indicated. The images or other third
party material in this book are included in the work’s Creative Commons license, unless indicated
otherwise in the credit line; if such material is not included in the work’s Creative Commons
license and the respective action is not permitted by statutory regulation, users will need to obtain
permission from the license holder to duplicate, adapt, or reproduce the material.
The use of general descriptive names, registered names, trademarks, service marks, etc. in this
publication does not imply, even in the absence of a specific statement, that such names are exempt
from the relevant protective laws and regulations and therefore free for general use.
The publisher, the authors and the editors are safe to assume that the advice and information in
this book are believed to be true and accurate at the date of publication. Neither the publisher nor
the authors or the editors give a warranty, express or implied, with respect to the material contained
herein or for any errors or omissions that may have been made.
Printed on acid-free paper.

Routledge is an imprint of the Taylor & Francis Group, an informa business

ISBN 978-87-93237-03-2 (print)

While every effort is made to provide dependable information, the publisher, authors, and editors
cannot be held responsible for any errors or omissions.
 Contents

Preface xiii

List of Figures xix

List of Tables xxix

List of Abbreviations xxxi

1 A Modular Architecture for Developing Robots for Industrial

Applications 1
A. Faíña, F. Orjales, D. Souto, F. Bellas and R. J. Duro
1.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 2
1.2 Main Characteristics for Industrial Operation and Design
Decisions . . . . . . . . . . . . . . . . . . . . . . . . . . . 4
1.3 Implementation of a Heterogeneous Modular Architecture
Prototype . . . . . . . . . . . . . . . . . . . . . . . . . . . 9
1.3.1 Actuator Modules . . . . . . . . . . . . . . . . . . . 10
1.3.1.1 Slider module . . . . . . . . . . . . . . . 11
1.3.1.2 Telescopic module . . . . . . . . . . . . . 11
1.3.1.3 Rotational module . . . . . . . . . . . . . 11
1.3.1.4 Hinge module . . . . . . . . . . . . . . . 12
1.3.2 Connection Mechanism . . . . . . . . . . . . . . . 12
1.3.3 Energy . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3.4 Sensors . . . . . . . . . . . . . . . . . . . . . . . . 13
1.3.5 Communications . . . . . . . . . . . . . . . . . . . 14
1.3.6 Control . . . . . . . . . . . . . . . . . . . . . . . . 15
1.4 Some Configurations for Practical Applications . . . . . . . 16
1.4.1 Manipulators . . . . . . . . . . . . . . . . . . . . . 16
1.4.2 Climber and Walker Robots . . . . . . . . . . . . . 18
1.5 Towards Industrial Applications . . . . . . . . . . . . . . . 20
1.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 21

v
vi Contents

2 The Dynamic Characteristics of a Manipulator with Parallel

Kinematic Structure Based on Experimental Data 27
S. Osadchy, V. Zozulya and A. Timoshenko
2.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 27
2.2 Purpose and Task of Research . . . . . . . . . . . . . . . . 29
2.3 Algorithm for the Structural Identification of the
Multivariable Dynamic Object with the Help of the
Complete Data . . . . . . . . . . . . . . . . . . . . . . . . 30
2.4 Algorithm for the Structural Identification of the
Multivariable Dynamic Object with the Help of
Incomplete Data . . . . . . . . . . . . . . . . . . . . . . . . 33
2.5 The Dynamics of the Mechanism with a Parallel
Structure Obtained by Means of the Complete
Data Identification . . . . . . . . . . . . . . . . . . . . . . 36
2.6 The Dynamics of the Mechanism with a Parallel
Structure Obtained by Means of the Incomplete Data
Identification . . . . . . . . . . . . . . . . . . . . . . . . . 41
2.7 Verification of the Structural Identification Results . . . . . . 45
2.8 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 47

3 An Autonomous Scale Ship Model for Parametric Rolling Towing

Tank Testing 49
M. Míguez González, A. Deibe, F. Orjales, B. Priego
and F. López Peña
3.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 49
3.2 System Architecture . . . . . . . . . . . . . . . . . . . . . . 51
3.2.1 Data Acquisition . . . . . . . . . . . . . . . . . . . 52
3.2.2 Software Systems . . . . . . . . . . . . . . . . . . . 53
3.2.3 Speed Control . . . . . . . . . . . . . . . . . . . . 55
3.2.4 Track-Keeping Control . . . . . . . . . . . . . . . . 58
3.2.5 Other Components . . . . . . . . . . . . . . . . . . 58
3.3 Testing . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 59
3.3.1 Prediction System . . . . . . . . . . . . . . . . . . 60
3.3.2 Prevention System . . . . . . . . . . . . . . . . . . 61
3.3.3 Towing Tank Tests and Results . . . . . . . . . . . . 62
3.3.3.1 Mathematical model validation . . . . . . 66
3.3.3.2 Validation of stability diagrams . . . . . . 67
3.3.3.3 Prediction system tests . . . . . . . . . . 68
3.4 Conclusions and Future Work . . . . . . . . . . . . . . . . . 69
 Contents vii

4 Autonomous Knowledge Discovery Based on Artificial

Curiosity-Driven Learning by Interaction 73
K. Madani, D. M. Ramik and C. Sabourin
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 73
4.2 Proposed System and Role of Curiosity . . . . . . . . . . . 76
4.2.1 Interpretation from Observation . . . . . . . . . . . 78
4.2.2 Search for the Most Coherent Interpretation . . . . . 81
4.2.3 Human-Robot Interaction . . . . . . . . . . . . . . 82
4.3 Validation Results by Simulation . . . . . . . . . . . . . . . 83
4.4 Implementation on Real Robot and Validation Results . . . . 85
4.4.1 Implementation . . . . . . . . . . . . . . . . . . . . 85
4.4.2 Validation Results . . . . . . . . . . . . . . . . . . 88
4.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 92

5 Information Technology for Interactive Robot Task Training

Through Demonstration of Movement 95
F. Kulakov and S. Chernakova
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 95
5.2 Conception and Principles of Motion Modeling . . . . . . . 97
5.2.1 Generalized Model of Motion . . . . . . . . . . . . 97
5.2.2 Algorithm for Robot Task Training by
Demonstration . . . . . . . . . . . . . . . . . . . . 99
5.2.3 Algorithm for Motion Reproduction after Task
Training by Demonstration . . . . . . . . . . . . . . 102
5.2.4 Verification of Results for the Task of Training the
Telecontrolled (Remote Controlled) Robot . . . . . . 104
5.2.5 Major Advantages of Task Training by
Demonstration . . . . . . . . . . . . . . . . . . . . 105
5.3 Algorithms and Models for Teaching Movements . . . . . . 106
5.3.1 Task Training by Demonstration of Movement among
the Objects of the Environment . . . . . . . . . . . 106
5.3.2 Basic Algorithms for Robot Task Training by
Demonstration . . . . . . . . . . . . . . . . . . . . 110
5.3.3 Training Algorithm for the Environmental Survey
Motion . . . . . . . . . . . . . . . . . . . . . . . . 110
5.3.4 Training Algorithm for Grabbing a Single
Object . . . . . . . . . . . . . . . . . . . . . . . . . 111
5.3.5 Special Features of the Algorithm for Reproduction
of Movements . . . . . . . . . . . . . . . . . . . . 112
viii Contents

5.3.6 Some Results of Experimental Studies . . . . . . . . 113

5.3.7 Overview of the Environment for Task Training by
Demonstration of the Movements of the Human
Head . . . . . . . . . . . . . . . . . . . . . . . . . 115
5.3.8 Training the Robot to Grab Objects by Demonstration
of Operator Hand Movements . . . . . . . . . . . . 116
5.4 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 119

6 A Multi-Agent Reinforcement Learning Approach for the

Efficient Control of Mobile Robots 123
U. Dziomin, A. Kabysh, R. Stetter and V. Golovko
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 123
6.2 Holonic Homogenous Multi-Agent Systems . . . . . . . . . 125
6.2.1 Holonic, Multi-Agent Systems . . . . . . . . . . . . 126
6.2.2 Homogenous, Multi-Agent Systems . . . . . . . . . 127
6.2.3 Approach to Commitment and Coordination
in H2 MAS . . . . . . . . . . . . . . . . . . . . . . 128
6.2.4 Learning to Coordinate Through Interaction . . . . . 129
6.3 Vehicle Steering Module . . . . . . . . . . . . . . . . . . . 130
6.4 A Decomposition of Mobile Platform . . . . . . . . . . . . . 132
6.5 The Robot Control System Learning . . . . . . . . . . . . . 134
6.5.1 Learning of the Turning of a Module-Agent . . . . . 135
6.5.1.1 Simulation . . . . . . . . . . . . . . . . . 138
6.5.1.2 Verification . . . . . . . . . . . . . . . . . 139
6.5.2 Learning of the Turning of a Module-Agent . . . . . 140
6.5.2.1 Simulation . . . . . . . . . . . . . . . . . 140
6.5.2.2 Verification . . . . . . . . . . . . . . . . . 141
6.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 142

7 Underwater Robot Intelligent Control Based on Multilayer

Neural Network 147
D. A. Oskin, A. A. Dyda, S. Longhi and A. Monteriù
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 147
7.2 Underwater Robot Model . . . . . . . . . . . . . . . . . . . 149
7.3 Intelligent NN Controller and Learning Algorithm
Derivation . . . . . . . . . . . . . . . . . . . . . . . . . . . 149
7.4 Simulation Results of the Intelligent NN Controller . . . . . 153
7.5 Modification of NN Control . . . . . . . . . . . . . . . . . 154
7.6 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 165
 Contents ix

8 Advanced Trends in Design of Slip Displacement Sensors

for Intelligent Robots 167
Y. P. Kondratenko and V. Y. Kondratenko
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 168
8.2 Analysis of Robot Task Solving Based on Slip Displacement
Signals Detection . . . . . . . . . . . . . . . . . . . . . . . 168
8.3 Analysis of Methods for Slip Displacement Sensors
Design . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 172
8.4 Mathematical Model of Magnetic Slip Displacement
Sensor . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 174
8.4.1 SDS Based on “Permanent Magnet/Hall Sensor”
Sensitive Element and Its Mathematical Model . . . 174
8.4.2 Simulation Results . . . . . . . . . . . . . . . . . . 177
8.5 Advanced Approaches for Increasing the Efficiency
of Slip Displacement Sensors . . . . . . . . . . . . . . . . . 179
8.6 Advances in Development of Smart Grippers for Intelligent
Robots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 184
8.6.1 Self-Clamping Grippers of Intelligent Robots . . . . 184
8.6.2 Slip Displacement Signal Processing in Real
Time . . . . . . . . . . . . . . . . . . . . . . . . . 186
8.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 188

9 Distributed Data Acquisition and Control Systems for a Sized

Autonomous Vehicle 193
T. Happek, U. Lang, T. Bockmeier, D. Neubauer and A. Kuznietsov
9.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 193
9.2 The Testing Environment . . . . . . . . . . . . . . . . . . . 194
9.3 Description of the System . . . . . . . . . . . . . . . . . . . 195
9.4 Lane Detection . . . . . . . . . . . . . . . . . . . . . . . . 199
9.4.1 In-Range Filter . . . . . . . . . . . . . . . . . . . . 199
9.4.2 Hough-Transformation . . . . . . . . . . . . . . . . 200
9.4.3 Lane Marks . . . . . . . . . . . . . . . . . . . . . . 201
9.4.4 Polynomial . . . . . . . . . . . . . . . . . . . . . . 201
9.4.5 Driving Lane . . . . . . . . . . . . . . . . . . . . . 201
9.4.6 Stop Line . . . . . . . . . . . . . . . . . . . . . . . 203
9.4.7 Coordinate Transformation . . . . . . . . . . . . . . 204
9.5 Control of the Vehicle . . . . . . . . . . . . . . . . . . . . . 206
9.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 207
9.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 207
x Contents

10 Polymetric Sensing in Intelligent Systems 211

Yu. Zhukov, B. Gordeev, A. Zivenko and A. Nakonechniy
10.1 Topicality of Polymetric Sensing . . . . . . . . . . . . . . . 211
10.2 Advanced Perception Components of Intelligent Systems or
Robots . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 213
10.2.1 Comparison of the Basics of Classical and Polymetric
Sensing . . . . . . . . . . . . . . . . . . . . . . . . 213
10.2.2 Advanced Structure of Multi-Agent Intelligent
Systems . . . . . . . . . . . . . . . . . . . . . . . . 214
10.3 Practical Example of Polymetric Sensing . . . . . . . . . . . 216
10.3.1 Adding the Time Scale . . . . . . . . . . . . . . . . 217
10.3.2 Adding the Information about the Velocity of the
Electromagnetic Wave . . . . . . . . . . . . . . . . 219
10.4 Efficiency of Industrial Polymetric Systems . . . . . . . . . 221
10.4.1 Naval Application . . . . . . . . . . . . . . . . . . 221
10.4.1.1 Sensory monitoring agency SMA . . . . . 224
10.4.1.2 Information Environment Agency
INE . . . . . . . . . . . . . . . . . . . . 226
10.4.1.3 Operator Interface Agency OPI . . . . . . 226
10.4.1.4 Advantages of the polymetric sensing . . . 227
10.4.1.5 Floating dock operation control
system . . . . . . . . . . . . . . . . . . . 230
10.4.1.6 Onshore applications . . . . . . . . . . . 231
10.4.1.7 Special applications . . . . . . . . . . . . 231
10.5 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 232

11 Design and Implementation of Wireless Sensor Network Based

on Multilevel Femtocells for Home Monitoring 235
D. Popescu, G. Stamatescu, A. Măciucă and M. Struţu
11.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 235
11.2 Network Architecture and Femtocell Structure . . . . . . . . 237
11.2.1 Body Sensor Network . . . . . . . . . . . . . . . . 238
11.2.2 Ambient Sensor Network . . . . . . . . . . . . . . . 239
11.2.3 Emergency Sensor Network . . . . . . . . . . . . . 239
11.2.4 Higher-level Architecture and Functional
Overview . . . . . . . . . . . . . . . . . . . . . . . 240
11.3 Data Processing . . . . . . . . . . . . . . . . . . . . . . . . 243
11.4 Experimental Results . . . . . . . . . . . . . . . . . . . . . 245
11.5 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . 253
 Contents xi

12 Common Framework Model for Multi-Purpose Underwater

Data Collection Devices Deployed with Remotely Operated
Vehicles 257
M.C. Caraivan, V. Dache and V. Sgarciu
12.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 257
12.2 Research Challenges . . . . . . . . . . . . . . . . . . . . . 260
12.2.1 Power Supply . . . . . . . . . . . . . . . . . . . . . 260
12.2.2 Communications . . . . . . . . . . . . . . . . . . . 263
12.2.3 Maintenance . . . . . . . . . . . . . . . . . . . . . 268
12.2.4 Law and Finance . . . . . . . . . . . . . . . . . . . 268
12.2.5 Possible Applications . . . . . . . . . . . . . . . . . 270
12.3 Mathematical Model . . . . . . . . . . . . . . . . . . . . . 270
12.3.1 System Definition . . . . . . . . . . . . . . . . . . 272
12.3.2 Actuator Definition . . . . . . . . . . . . . . . . . . 273
12.3.3 Sensor Definition . . . . . . . . . . . . . . . . . . . 275
12.4 ROV . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 277
12.4.1 ROV Manipulator Systems . . . . . . . . . . . . . . 279
12.4.2 Types of Offshore Constructions . . . . . . . . . . . 282
12.5 ROV Simulator . . . . . . . . . . . . . . . . . . . . . . . . 282
12.6 Common Modular Framework . . . . . . . . . . . . . . . . 287
12.7 Conclusions . . . . . . . . . . . . . . . . . . . . . . . . . . 290

13 M2M in Agriculture – Business Models and Security Issues 295

S. Gansemer, J. Sell, U. Grossmann, E. Eren, B. Horster,
T. Horster-Möller and C. Rusch
13.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . 295
13.2 Related Work . . . . . . . . . . . . . . . . . . . . . . . . . 296
13.3 Communication and Standardization . . . . . . . . . . . . . 296
13.4 Business Cases . . . . . . . . . . . . . . . . . . . . . . . . 300
13.4.1 Process Transparency (PT) . . . . . . . . . . . . . . 300
13.4.2 Operations Data Acquisition (ODA) . . . . . . . . . 302
13.4.3 Remote Software Update (RSU) . . . . . . . . . . . 302
13.5 Business Models . . . . . . . . . . . . . . . . . . . . . . . 302
13.6 Economic Analysis . . . . . . . . . . . . . . . . . . . . . . 304
13.7 Communication Security . . . . . . . . . . . . . . . . . . . 307
13.7.1 CA . . . . . . . . . . . . . . . . . . . . . . . . . . 307
13.7.2 Communicating On-the-Go . . . . . . . . . . . . . 307
13.7.3 Covering Dead Spots . . . . . . . . . . . . . . . . . 308
13.7.4 Securing WLAN Infrastructures . . . . . . . . . . . 308
xii Contents

13.7.5 Firmware Update . . . . . . . . . . . . . . . . . . . 309

13.8 Resume . . . . . . . . . . . . . . . . . . . . . . . . . . . . 309
13.9 Acknowledgement . . . . . . . . . . . . . . . . . . . . . . 311

Index 313

Editor’s Biographies 317

Author’s Biographies 319

 Preface

This book provides an overview of a series of advanced trends in robotics as

well as of design and development methodologies for intelligent robots and
their intelligent components.
All the contributionswere discussed at the International IEEE Conference
IDAACS-2013 (Berlin, Germany, 12–14 June, 2013). The IDAACS workshop
series is established as a forum for high quality reports on state-of-the-art
theory, technology and applications of intelligent data acquisition and
advanced computer systems. All of these techniques and applications have
experienced a rapid expansion in the last few years that has resulted in more
intelligent, sensitive, and accurate methods for the acquisition of data and its
processing applied to manufacturing process control and inspection, environ-
mental and medical monitoring and diagnostics. One of the most interesting
paradigms that encompass much of the research presented at IDAACs is
Intelligent Robotic Systems, and this is the area this book concentrates on.
The success of IDAACS arises not only from the importance of the
topics it focuses on, but also because of its nature as a unique forum for
establishing scientific contacts between research teams and scientists from
different countries. This purpose has become one of the main reasons for the
rapid success of IDAACS, as it turns out to be one of the few events in this
area of research where Western and former Eastern European scientists can
discuss and exchange ideas and information, allowing them to characterize
common and articulated research activities and creating the environment for
establishing joint research collaborations. It provides an opportunity for all the
participants to discuss topics with colleagues from different spheres such as
academia, industry, and public and private research institutions. Even though
this book concentrates on providing insights into what is being done in the area
of robotics and intelligent systems, the papers that were selected reflect the
variety of research presented during the workshop as well as the very diverse
fields that may benefit from these techniques.
In terms of structure, the 13 chapters of the book are grouped into
four sections: Robots, Control and Intelligence, Sensing and Collaborative

xiii
xiv Preface

Automation. The chapters have been thought out to provide an easy to follow
introduction to the topics that are addressed, including the most relevant
references, so that anyone interested in them can start their introduction to
the topic through these references. At the same time, all of them correspond to
different aspects of work in progress being carried out in various laboratories
throughout the world and, therefore, provide information on the state of the
art of some of these topics.
The first part, “Robots”, includes three contributions:
“A Modular Architecture for Developing Robots for Industrial Applica-
tions”, by A. Faíña, F. Orjales, D. Souto, F. Bellas and R. J. Duro, considers
ways to make feasible the use of robots in many sectors characterized by
dynamic and unstructured environments. The authors propose a new approach,
based on modular robotics, to allow the fast deployment of robots to solve
specific tasks. In this approach, the authors start by defining the industrial
settings the architecture is aimed at and then extract the main features that
would be required from a modular robotic architecture to operate successfully
in this context. Finally, a particular heterogeneous modular robotic architecture
is designed from these requirements and a laboratory implementation of it is
built in order to test its capabilities and show its versatility using a set of
different configurations including manipulators, climbers and walkers.
S. Osadchy, V. Zozulya and A. Timoshenko, in “The Dynamic Char-
acteristics of a Manipulator with Parallel Kinematic Structure Based on
Experimental Data”, studies two identification techniques which the authors
found most useful in examining the dynamic characteristics of a manipulator
with a parallel kinematic structure as an object of control. These techniques
emphasize a frequency domain approach. If all input/output signals of an
object can be measured then the first one of such techniques may be used
for identification. In the case when all disturbances cannot be measured the
second identification technique may be used.
In “An Autonomous Scale Ship Model for Parametric Rolling Towing
Tank Testing”, M. Míguez González, A. Deibe, F. Orjales, B. Priego and
F. López Peña analyze a special kind of robotic system model, in particular, a
self-propelled scale ship model for model testing, with the main characteristic
of not having any material link to a towing device to carry out the tests. This
model has been fully instrumented in order to acquire all the significant raw
data, process them onboard and communicate with an inshore station.
The second part “Control and Intelligence” includes four contributions:
In “Autonomous Knowledge Discovery Based on Artificial Curiosity Driven
Learning by Interaction”, K. Madani, D. M. Ramik and C. Sabourin investigate
 Preface xv

the development of a real-time intelligent system allowing a robot to discover

its surrounding world and to learn autonomously new knowledge about it by
semantically interacting with humans. The learning is performed by obser-
vation and by interaction with a human. The authors provide experimental
results both using simulated environments and implementing the approach on
a humanoid robot in a real-world environment including everyday objects.
The proposed approach allows a humanoid robot to learn without negative
input and from a small number of samples.
F. Kulakov and S. Chernakova, in “Information Technology for Interactive
Robot Task Training through Demonstration of Movement”, consider the
problem of remote robot control, which includes the solution of the following
routine problems: surveillance of the remote working area, remote operation
of the robot situated in the remote working area, as well as pre-training of
the robot. Authors propose a new technique for robot control using intelligent
multimodal human-machine interfaces (HMI). The application of the new
training technology is very promising for space robots as well as for modern
assembly plants, including the use of micro-and nanorobots.
In “A Multi-Agent Reinforcement Learning Approach for the Effi-
cient Control of Mobile Robots”, U. Dziomin, A. Kabysh, R. Stetter and
V. Golovko present a multi-agent control architecture for the efficient control
of a multi-wheeled mobile platform. The proposed control architecture is
based on the decomposition of a platform into a holonic, homogenous, multi-
agent system. The multi-agent system incorporates multiple Q-learning agents,
which permits them to effectively control every wheel relative to other wheels.
The learning process consists module positioning—where the agents learn
to minimize the error of orientation, and cooperative movement—where the
agents learn to adjust the desired velocity in order to conform to the desired
position in formation. Experiments with a simulation model and the real robot
are discussed in details.
D. Oskin, A. Dyda, S. Longhi and A. Monteriù, in “Underwater Robot
Intelligent Control Based on Multilayer Neural Network”, analyse the design
of an intelligent neural network based control system for underwater robots.
A new algorithm for intelligent controller learning is derived using the speed
gradient method. The proposed systems provide robot dynamics close to
the reference ones. Simulation results of neural network control systems for
underwater robot dynamics with parameter and partial structural uncertainty
have confirmed the perspectives and effectiveness of the developed approach.
The third part “Sensing” includes four contributions:
xvi Preface

“Advanced Trends in Design of Slip Displacement Sensors for Intelligent

Robots”, by Y. Kondratenko and V. Kondratenko, discusses advanced trends in
the design of modern tactile sensors andsensor systems for intelligent robots.
The detection of slip displacement signals provides information on three
approaches for using slip displacement signals, in particular, for the correction
of the clamping force, the identification of manipulated object mass and the
correction of the robot control algorithm.The chapter presents the analysis of
different methods for the detectionof slip displacement signals, as well as new
sensor schemes, mathematical models and correction methods.
T. Happek, U. Lang, T. Bockmeier, D. Neubauer and A. Kuznietsov, in
“Distributed Data Acquisition and Control Systems for a Sized Autonomous
Vehicle”, present an autonomous car with distributed data processing. The
car is controlled by a multitude of independent sensors. For lane detection, a
camera is used, which detects the lane marks using a Hough transformation.
Once the camera detects these, one of them is selected to be followed by the
car. This lane is verified by the other sensors of the car. These sensors check
the route for obstructions or allow the car to scan a parking space and to park
on the roadside if the gap is large enough.
In “Polymetric Sensing in Intelligent Systems”, Yu. Zhukov, B. Gordeev,
A. Zivenko and A. Nakonechniy examine the up-to-date relationship between
the theory of polymetric measurements and the state of the art in intelligent
system sensing. The chapter discusses concepts of polymetric measurements,
corresponding to monitoring information systems used in different technolo-
gies and some prospects for polymetric sensing in intelligent systems and
robots. The application of the described concepts in technological processes
ready to be controlled by intelligent systems is illustrated.
D. Popescu, G. Stamatescu, A. Maciuca and M. Strutu, in “Design and
Implementation of Wireless SensorNetwork Based on Multilevel Femtocells
for Home Monitoring”, propose an intelligent femtocell-based sensor network
for home monitoring of elderly or people with chronic diseases. The femtocell
is defined as a small sensor network which is placed into the patient’s house
and consists of both mobile and fixed sensors disposed on three layers. The first
layer contains body sensors attached to the patient that monitor different health
parameters, patient location, position and possible falls. The second layer is
dedicated for ambient sensors and routing inside the cell. The third layer
contains emergency ambient sensors that cover burglary events or toxic gas
concentration, distributed by necessities. Cell implementation is based on The
IRIS family of motes running the embedded software for resource constrained
 Preface xvii

devices, TinyOS. Experimental results within the system architecture are

presented for a detailed analysis and validation.
The fourth part “Collaborative automation” includes two contributions:
In “Common Framework Model for Multi-purpose Underwater Data Col-
lection Devices Deployed with Remotely Operated Vehicles”, M. Caraivan,
V. Dache and V. Sgarciu presents a common framework model for multi-
purpose underwater sensors used for offshore exploration. The development
of real-time applications for marine operations focusing on modern modeling
and simulation methods are discussed with addressing deployment challenges
of underwater sensor networks “Safe-Nets” by using Remotely Operated
Vehicles.
Finally, S. Gansemer, J. Sell, U. Grossmann, E. Eren, B. Horster, T.
Horster-Möller and C. Rusch, in “M2M in Agriculture - Business Models and
Security Issues” consider the machine-to-machine communication (M2M)
as one of the major ICT innovations. A concept for process optimization
in agricultural business using M2M technologies is presented using three
application scenarios. Within that concept standardization and communication
as well as security aspects are discussed.
The papers selected for this book are extended and improved versions of
those presented at the workshop and as such are significantly expanded with
respect to the original ones presented at the workshop. Obviously, this set of
papers are just a sample of the dozens of presentations and results that were
seen at IDAACS 2011, but we do believe that they provide an overview of some
of the problems in the area of robotics systems and intelligent automation and
the approaches and techniques that relevant research groups within this area
are employing to try to solve them. We would like to express our appreciation
to all authors for their contributions as well as to reviewers for their timely and
interesting comments and suggestions. We certainly look forward to working
with you again.

Yuriy Kondratenko
Richard Duro
 List of Figures

Figure 1.1 Diagram of the selected missions, tasks and sub-

tasks considered, and the required actuators and
effectors. . . . . . . . . . . . . . . . . . . . . . . 7
Figure 1.2 Different types of modules developed in this project:
three effectors on the left part, a linker on the top,
a slider on the right, and in the middle there is a
rotational module, a hinge module and a telescopic
module. . . . . . . . . . . . . . . . . . . . . . . . 10
Figure 1.3 Control board for the slider module and its main
components. . . . . . . . . . . . . . . . . . . . . . 15
Figure 1.4 Spherical manipulator moving a load from one place
to another. . . . . . . . . . . . . . . . . . . . . . . 17
Figure 1.5 Cartesian manipulators for Static missions. . . . . 18
Figure 1.6 A snake Robot that can inspect inside a pipe. . . . . 19
Figure 1.7 Climber and Walker Robots for linear and surface
missions. . . . . . . . . . . . . . . . . . . . . . . 19
Figure 1.8 A Biped Robot able to overpass obstacles. . . . . . 20
Figure 2.1 Kinematic diagram of single-section
mechanism. . . . . . . . . . . . . . . . . . . . . . 28
Figure 2.2 Block diagram of the mechanism with parallel
kinematics. . . . . . . . . . . . . . . . . . . . . . 29
Figure 2.3 Graphs of changes in the length of the rods. . . . . 36
Figure 2.4 Graphs of changes in the projections of the resis-
tance moments. . . . . . . . . . . . . . . . . . . . 37
Figure 2.5 Graphs of chances in the coordinates of the plat-
form’s center of mass. . . . . . . . . . . . . . . . 37
Figure 2.6 Bode diagrams of the mechanism with a parallel
structure. . . . . . . . . . . . . . . . . . . . . . . 40
Figure 2.7 Block diagram of the mechanism with a parallel
kinematics. . . . . . . . . . . . . . . . . . . . . . 41

xix
xx List of Figures

Figure 2.8 Manipulator with a controlled diode motor-operated

drive. . . . . . . . . . . . . . . . . . . . . . . . . 42
Figure 2.9 Grapas of the vector u componente changes. . . . . 43
Figure 2.10 Grapas of vector x component changes. . . . . . . 43
Figure 2.11 The scheme simulation model of the mechanism
with parallel kinematics. . . . . . . . . . . . . . . 46
Figure 2.12 Graphs of the change of coordinates X of the center
of mass of the platform. . . . . . . . . . . . . . . . 46
Figure 2.13 Graphs of thee changes of Y coordinates of the center
of mass of the platform. . . . . . . . . . . . . . . . 47
Figure 3.1 Ship scale model overview. . . . . . . . . . . . . . 53
Figure 3.2 Connectivity between Mini-PC and sensors
onboard. . . . . . . . . . . . . . . . . . . . . . . . 54
Figure 3.3 Graphical user interface to monitor/control the
model from an external workstation. . . . . . . . . 55
Figure 3.4 Double PID Speed control. . . . . . . . . . . . . . 56
Figure 3.5 Speed control. complementary filter. . . . . . . . . 57
Figure 3.6 Propulsion system block. . . . . . . . . . . . . . . 59
Figure 3.7 Roll and pitch motions in parametric roll resonance.
Conventional carriage-towed model. . . . . . . . . 63
Figure 3.8 Conventional carriage-towed model during
testing. . . . . . . . . . . . . . . . . . . . . . . . . 63
Figure 3.9 Proposed model during testing. . . . . . . . . . . . 65
Figure 3.10 Roll and pitch motions in parametric roll resonance.
Proposed model. . . . . . . . . . . . . . . . . . . 65
Figure 3.11 Comparison between experimental and numerical
data. Resonant case. . . . . . . . . . . . . . . . . . 66
Figure 3.12 Comparison between experimental and numerical
data. Non-Resonant case. . . . . . . . . . . . . . . 66
Figure 3.13 Experimental stability diagrams. Fn = 0.1. (Left), Fn
= 0.15 (Right), GM = 0.370 m. . . . . . . . . . . . 67
Figure 3.14 Comparison between experimental and numerical
stability diagrams. Fn = 0.1. (Left), Fn = 0.15
(Right), GM = 0.370 m. . . . . . . . . . . . . . . . 67
Figure 3.15 Forecast results. 30 neuron, 3 layer MP. 10 seconds
prediction. Resonant case. . . . . . . . . . . . . . 69
Figure 3.16 Forecast results. 30 neuron, 3 layer mp. 10 Seconds
prediction. Non Resonant case. . . . . . . . . . . . 69
 List of Figures xxi

Figure 4.1 General Bloc-diagram of the proposed curios-

ity driven architecture (left) and principle of
curiosity-based stimulation-satisfaction mechanism
for knowledge acquisition (right). . . . . . . . . . 77
Figure 4.2 A Human would describe this Apple as “Red” in
spite of the fact, that this is not the only visible
color. . . . . . . . . . . . . . . . . . . . . . . . . 78
Figure 4.3 A Human would describe this Toy-frog as green in
spite of the fact, that this is not the only visible
color. . . . . . . . . . . . . . . . . . . . . . . . . 79
Figure 4.4 Bloc-diagram of relations between observations,
features, beliefs and utterances in sense of terms
defined in the text. . . . . . . . . . . . . . . . . . . 80
Figure 4.5 Upper: the WCS color table. lower: the WCS color
table interpreted by robot taught to distinguish warm
(marked by red), cool (blue) and neutral (white)
colors. . . . . . . . . . . . . . . . . . . . . . . . . 83
Figure 4.6 Evolution of number of correctly described objects
with increasing number of exposures of each color
to the simulated robot. . . . . . . . . . . . . . . . 84
Figure 4.7 Examples of obtained visual colors’ interpretations
(lower images) and corresponding original images
(upper images) for several testing objects from
COIL database. . . . . . . . . . . . . . . . . . . . 84
Figure 4.8 Block diagram the implementation’s
architecture. . . . . . . . . . . . . . . . . . . . . . 86
Figure 4.9 Example of English phrase and the corresponding
syntactic analysis output generated by
treetagger. . . . . . . . . . . . . . . . . . . . . . . 87
Figure 4.10 Flow diagram of communication between a robot
and a human which is used in this work. . . . . . . 88
Figure 4.11 Everyday objects used in the experiments in this
work. . . . . . . . . . . . . . . . . . . . . . . . . 89
Figure 4.12 Tutor pointing an aid-kit detected by robot describ-
ing its name and color to the robot (left-side picture).
Pointing, in the same way, another visible objects
detected by robot, tutor describes them to the robot
(Right-Side picture). . . . . . . . . . . . . . . . . 90
xxii List of Figures

Figure 4.13 Tutor pointing a yellow chocolate box which has

been seen, interpreted and learned (by the robot) in
terms of colors then asking the robot to describe the
chosen object (left-side picture). Tutor pointing an
unseen white teddy-bear asking the robot to describe
the chosen object (right-side picture). . . . . . . . 91
Figure 4.14 Images from a video sequence showing the robot
searching for the book (left-side picture) and robot’s
camera view and visualization of color interpreta-
tion of the searched object (right-side picture). . . . 91
Figure 5.1 Images of the Space Station for two positions:
“Convenient for observation” and “Convenient for
grabbing” objects with the virtual manipulator . . . 98
Figure 5.2 “Sensitized Glove” with a camera and the process
of training the robot by means of demonstration . . 99
Figure 5.3 Formation of images of 3 characteristic points of the
object. . . . . . . . . . . . . . . . . . . . . . . . . 108
Figure 5.4 Functional diagram of robot task training regarding
survey motions and object grabbing motions using
THM and RCS. . . . . . . . . . . . . . . . . . . . 115
Figure 5.5 Robot task training to execute survey movements,
based on the movements of the operator’s head:
Training the robot to execute survey motions to
insect surroundings (a); Training process (b); Repro-
duction of earlier trained movements (c). . . . . . . 116
Figure 5.6 Variations of the “Sensitized Glove”
construction. . . . . . . . . . . . . . . . . . . . . . 117
Figure 5.7 Using the special glove for training the robot manip-
ulator. . . . . . . . . . . . . . . . . . . . . . . . . 117
Figure 5.8 Stand for teaching robots to execute motions of
surveillance and grabbing objects. . . . . . . . . . 117
Figure 5.9 Training of motion coordination of two robot manip-
ulators by natural movements of human head and
hand. . . . . . . . . . . . . . . . . . . . . . . . . . 118
Figure 5.10 Training with the use of a system for the recognition
of hand movements and gestures without “Sensi-
tized Gloves” against the real background of the
operator’s work station. . . . . . . . . . . . . . . . 118
 List of Figures xxiii

Figure 6.1 Production mobile robot: Production mobile plat-

form (a); Driving module (b). . . . . . . . . . . . . 125
Figure 6.2 Organizational structure of a Holonic Multi-Agent
System. Lines indicate the communication
channels. . . . . . . . . . . . . . . . . . . . . . . 127
Figure 6.3 Model of Influence Based Multi-Agent Reinforce-
ment Learning in the Case of a Holonic Homoge-
nous Multi-Agent System. . . . . . . . . . . . . . 130
Figure 6.4 The Maneuverability of one module. . . . . . . . . 131
Figure 6.5 Mobile Robot Trajectory Decomposition. . . . . . 131
Figure 6.6 Holonic Decomposition of the Mobile Platform.
Dashed lines represent the boundary of a Multi-
Agent System (the Holon). Introduction of the Head
Agent Leads to a reduction of communication
costs. . . . . . . . . . . . . . . . . . . . . . . . . 132
Figure 6.7 Virtual Structure with a Virtual Coordinate Frame
composed of Four Modules with a known Virtual
Center. . . . . . . . . . . . . . . . . . . . . . . . . 133
Figure 6.8 A unified view of the control architecture for a
Mobile Platform. . . . . . . . . . . . . . . . . . . 135
Figure 6.9 State of the Module with Respect to Reference
Beacon. . . . . . . . . . . . . . . . . . . . . . . . 136
Figure 6.10 A Decision tree of the reward function. . . . . . . . 137
Figure 6.11 Result Topology of the Q-Function. . . . . . . . . 138
Figure 6.12 Initial and Final Agente Positions. . . . . . . . . . 139
Figure 6.13 Execution of a Learned Control System to turn
modules to the center, which is placed on the rear
right relative to the platform. . . . . . . . . . . . . 140
Figure 6.14 Agents Team Driving Process. . . . . . . . . . . . 141
Figure 6.15 The Experiment of modules turning as in the Car
Kinematics Scheme (1–6 screenshots) and move-
ment around a White Beacon (7–9). . . . . . . . . 142
Figure 6.16 The Experiment shows that the radius doesn’t
change during movement. . . . . . . . . . . . . . . 142
Figure 7.1 Neural network structure. . . . . . . . . . . . . . . 151
Figure 7.2 Transient processes in NN control system (α = 0.01,
γ = 250). . . . . . . . . . . . . . . . . . . . . . . 155
Figure 7.3 Forces and Torque in NN control system (α = 0.01,
γ = 250). . . . . . . . . . . . . . . . . . . . . . . 155
xxiv List of Figures

Figure 7.4 Examples of hidden layer weight coefficients evo-

lution (α = 0.01,γ = 250). . . . . . . . . . . . . . . 156
Figure 7.5 Examples of output layer weight coefficients evolu-
tion (α = 0.01, γ = 250). . . . . . . . . . . . . . . 156
Figure 7.6 Transient processes in NN control system (α = 0.01,
γ = 200). . . . . . . . . . . . . . . . . . . . . . . 157
Figure 7.7 Forces ant Torque in NN control system (α = 0.01,
γ = 200). . . . . . . . . . . . . . . . . . . . . . . 157
Figure 7.8 Examples of hidden layer weight coefficients evo-
lution (α = 0.01, γ = 200). . . . . . . . . . . . . . 158
Figure 7.9 Examples of output layer weight coefficients evolu-
tion (α = 0.01, γ = 200). . . . . . . . . . . . . . . 158
Figure 7.10 Scheme of the NN control system. . . . . . . . . . 160
Figure 7.11 Transient processes with modified NN-control
(α = 0, γ = 200). . . . . . . . . . . . . . . . . . . 161
Figure 7.12 Forces and torque with modified NN control (α = 0,
γ = 200). . . . . . . . . . . . . . . . . . . . . . . 161
Figure 7.13 Examples of hidden layer weight coefficients evo-
lution (α = 0, γ = 200). . . . . . . . . . . . . . . . 162
Figure 7.14 Examples of output layer weight coefficients evolu-
tion (α = 0, γ = 200). . . . . . . . . . . . . . . . . 162
Figure 7.15 Transient processes with modified NN control
(α =0.001, γ = 200). . . . . . . . . . . . . . . . . 163
Figure 7.16 Forces and Torque with modified NN control
(α = 0.001, γ = 200). . . . . . . . . . . . . . . . . 163
Figure 7.17 Examples of hidden layer weight coefficients evo-
lution (α =0.001, γ =200). . . . . . . . . . . . . . 164
Figure 7.18 Examples of output layer weight coefficients evolu-
tion (α =0.001, γ = 200). . . . . . . . . . . . . . . 164
Figure 8.1 Grasping and lifting an object with the robot’s arm:
Initial positions of the gripper fingers (1,2) and
object (3) (a); Creating the required clamping force
Fob by the gripper fingers during object slippage in
the lifting process (b). . . . . . . . . . . . . . . . . 169
Figure 8.2 Series of trial motions with increasing clamping
force F of gripper fingers based on object
slippage. . . . . . . . . . . . . . . . . . . . . . . . 170
Figure 8.3 The algorithm for solving different robot tasks based
on slip signal detection. . . . . . . . . . . . . . . . 171
 List of Figures xxv

Figure 8.4 Magnetic SDS: 1– Rod; 2– Head; 3– Permanent

magnet; 4– Hall sensor. . . . . . . . . . . . . . . . 175
Figure 8.5 Model of magnetic sensitive element. . . . . . . . 175
Figure 8.6 Simulation results for By (P) based on the mathe-
matical model (8.2). . . . . . . . . . . . . . . . . . 178
Figure 8.7 Comparative analysis of modeling and experimental
results. . . . . . . . . . . . . . . . . . . . . . . . . 179
Figure 8.8 The ball as sensitive element of SDS: 1– Finger of
robot’s gripper; 2– Cavity for SDS instalation; 3–
Guides; 4– Sensitive element (a ball); 5– Spring; 6–
Conductive rubber; 7, 8– Fiber optic light guides; 9–
a Sleeve; 10– Light; 11– Photodetector; 13–Cover;
14–Screw; 15–Hole. . . . . . . . . . . . . . . . . 180
Figure 8.9 Light-reflecting surface of the sensitive ball with
reflecting and absorbing portions (12) for light
signal. . . . . . . . . . . . . . . . . . . . . . . . . 180
Figure 8.10 Capacitated SDS for the detection of object slippage
in different directions: 1– Main cavity of robot’s
gripper; 2– Additional cavity; 3– Gripper’s finger;
4– Rod; 5– Tip; 6– Elastic working surface; 7–
Spring; 8– Resilient element; 9, 10– Capacitor
plates. . . . . . . . . . . . . . . . . . . . . . . . . 181
Figure 8.11 Intelligent sensor system for identification of object
slippage direction: 3– Gripper’s finger; 4– Rod; 9,
10– Capacitor plates; 11– Converter “capacitance-
voltage”; 12– Delay element; 13, 18, 23– Adders;
14, 15, 21, 26– Threshold elements; 16– Multi-
Inputs element OR; 17– Computer information-
control system; 19, 20, 24, 25– Channels for sensor
information processing; 22, 27– Elements NOT;
28–39– Elements AND. . . . . . . . . . . . . . . . 181
Figure 8.12 Self-adjusting gripper of an intelligent robot with
angle movement of clamping rollers: 1, 2– Finger;
3, 4– Guide groove; 5, 6– Roller; 7, 8– Roller axis; 9,
15, 20– Spring; 10– Object; 11, 18– Elastic working
surface; 12– Clamping force sensor; 13, 14– Ele
ctroconductive contacts; 16, 19– Fixator; 17– Stock;
21– Adjusting screw; 22– Deepening; 23– Finger’s
drive. . . . . . . . . . . . . . . . . . . . . . . . . 185
xxvi List of Figures

Figure 8.13 Self-clamping gripper of an intelligent robot with

plane-parallel displacement of the clamping roller:
Front view (a); Top view (b). . . . . . . . . . . . . 186
Figure 8.14 Experimental self-clamping gripper with plane-
parallel displacement of the clamping rollers. . . . 187
Figure 8.15 Intelligent robot with 4 degrees of freedom for
experemental investigations of SDS. . . . . . . . . 187
Figure 9.1 Dimensions of the test track. . . . . . . . . . . . . 194
Figure 9.2 Schematic base of the model car. . . . . . . . . . . 195
Figure 9.3 The infrared sensors distributed alongside the car. . 196
Figure 9.4 Position of the camera. . . . . . . . . . . . . . . . 197
Figure 9.5 Schematic signal flow of the vehicle. . . . . . . . . 198
Figure 9.6 Overview of the data processing system. . . . . . . 199
Figure 9.7 Comparison between original and in-range
image. . . . . . . . . . . . . . . . . . . . . . . . . 200
Figure 9.8 Original image without and with Hough-lines. . . . 200
Figure 9.9 Hough-Lines and sorted points along the
Hough-Lines. . . . . . . . . . . . . . . . . . . . . 201
Figure 9.10 Sorted Points and Least-Square Parable. . . . . . . 202
Figure 9.11 Parables and driving lane. . . . . . . . . . . . . . 203
Figure 9.12 Detection of stop lines. . . . . . . . . . . . . . . . 204
Figure 9.13 Driving along a set path: Track model (a); Lateral
deviation and heading angle (b). . . . . . . . . . . 206
Figure 10.1 The main idea of the replacement of the distributed
multi-Sensor system by a polymetric perceptive
agent. . . . . . . . . . . . . . . . . . . . . . . . . 215
Figure 10.2 TDR Coaxial probe immersed into the liquid and the
corresponding polymetric signal. . . . . . . . . . . 216
Figure 10.3 Time diagrams for polymetric signal formation
using additional reference time intervals. . . . . . . 219
Figure 10.4 Disposition of the measuring probe in the tank,
position of the reflector and corresponding signals
for the cases with air and vapour. . . . . . . . . . . 220
Figure 10.5 Example of the general structure of lascos hardware
components and elements. . . . . . . . . . . . . . 222
Figure 10.6 The general structure of lascos software elements
and functions. . . . . . . . . . . . . . . . . . . . . 223
Figure 10.7 The General Structure of LASCOS Holonic agen-
cies functions. . . . . . . . . . . . . . . . . . . . . 225
 List of Figures xxvii

Figure 10.8 Safe Storming Diagram Interface of LASCOS. . . 227

Figure 10.9 The Main Window of Ballast System Operations
Interface. . . . . . . . . . . . . . . . . . . . . . . 228
Figure 11.1 Hybrid femtocell configuration. . . . . . . . . . . 238
Figure 11.2 Data and information flow within and outside the
femtocell. . . . . . . . . . . . . . . . . . . . . . . 240
Figure 11.3 System architecture. . . . . . . . . . . . . . . . . 241
Figure 11.4 Sensor deployment example. . . . . . . . . . . . . 242
Figure 11.5 Embedded carbon dioxide sensor node
module. . . . . . . . . . . . . . . . . . . . . . . . 243
Figure 11.6 Accelerometer node placement on the patient’s
body. . . . . . . . . . . . . . . . . . . . . . . . . . 245
Figure 11.7 X Axis experiment acceleration with
thresholding. . . . . . . . . . . . . . . . . . . . . 246
Figure 11.8 Y Axis experiment acceleration with
thresholding. . . . . . . . . . . . . . . . . . . . . 247
Figure 11.9 Measurement data: humidity (a); temperature (b). . 250
Figure 11.10 Measurement data: barometric pressure (a);
light (b). . . . . . . . . . . . . . . . . . . . . . . . 251
Figure 12.1 Pelamis wave converter Orkney, U.K. . . . . . . . 261
Figure 12.2 Axial symmetric absorption buoy. . . . . . . . . . 262
Figure 12.3 Wave Dragon - Overtopping devices principle. . . 262
Figure 12.4 Archimedes Waveswing (AWS). . . . . . . . . . . 262
Figure 12.5 Wind farms in North Sea. . . . . . . . . . . . . . . 263
Figure 12.6 Possible underwater sensor network deployment
nearby Jack-up rig. . . . . . . . . . . . . . . . . . 265
Figure 12.7 Sound pressure diagram:1– Equilibrium; 2– Sound;
3– Environment Pressure; 4– Instantaneous Pressure
of Sound. . . . . . . . . . . . . . . . . . . . . . . 265
Figure 12.8 Signal-Noise Ratio (SNR). . . . . . . . . . . . . . 267
Figure 12.9 Most important underwater data & voice cables
(2008). . . . . . . . . . . . . . . . . . . . . . . . . 271
Figure 12.10 Graphical representation of actuators’ supports . . 274
Figure 12.11 Illustration of the geometrical support and spatial
distribution of an actuator. . . . . . . . . . . . . . 275
Figure 12.12 Graphical representation of the sensor supports. . . 275
Figure 12.13 Titan 4 Manipulator 7-F. . . . . . . . . . . . . . . 279
Figure 12.14 Master Arm Control. . . . . . . . . . . . . . . . . 280
Figure 12.15 Titan 4 – Stow dimensions. . . . . . . . . . . . . . 280
xxviii List of Figures

Figure 12.16 RigMaster 5-F. . . . . . . . . . . . . . . . . . . . 281

Figure 12.17 RigMaster range of motion – Side view. . . . . . . 281
Figure 12.18 Triton XLS ROV in simulation scenario. . . . . . . 284
Figure 12.19 Triton XLS schilling robotics 7-Function arm
in scenario. . . . . . . . . . . . . . . . . . . . . . 284
Figure 12.20 Typical hierarchical file layout. . . . . . . . . . . . 285
Figure 12.21 Spherical-shaped model designed for common
framework; a >= b; c is Tether/Cable entry point
diameter. . . . . . . . . . . . . . . . . . . . . . . . 288
Figure 12.22 Underwater multi-purpose devices prototypes
01 – 05. . . . . . . . . . . . . . . . . . . . . . . . 289
Figure 12.23 Grouping method for multiple simultaneous sensor
deployment. . . . . . . . . . . . . . . . . . . . . . 289
Figure 12.24 Basic sensor device holder designed for
simulation. . . . . . . . . . . . . . . . . . . . . . 290
Figure 13.1 Synchronization of standards. . . . . . . . . . . . 297
Figure 13.2 M2M teledesk framework. . . . . . . . . . . . . . 298
Figure 13.3 Femtocell communication in agriculture. . . . . . . 299
Figure 13.4 Information- and controlflow of scenario
harvesting process. . . . . . . . . . . . . . . . . . 301
Figure 13.5 Value chain of M2M-Teledesk. . . . . . . . . . . . 302
Figure 13.6 Service-links between delivering and receiving ser-
vices. . . . . . . . . . . . . . . . . . . . . . . . . 305
Figure 13.7 Relation matrix A, transfer price vector B and sales
price vector D. . . . . . . . . . . . . . . . . . . . 306
Figure 13.8 Vector of variable costs C and vector of marginal
return per unit M. . . . . . . . . . . . . . . . . . . 306
Figure 13.9 Large file encryption. . . . . . . . . . . . . . . . . 310
 List of Tables

Table 1.1 Actuator Modules . . . . . . . . . . . . . . . . . . 11

Table 3.1 Measured Parameters . . . . . . . . . . . . . . . . 52
Table 6.1 The Environment Information . . . . . . . . . . . 136
Table 6.2 Agent Actions . . . . . . . . . . . . . . . . . . . . 137
Table 8.1 The base of production rules “IF-THEN” for
indetification of the slip displacement direction . . 183
Table 10.1 Quantity and Sensor Types for the Traditional Cargo
Sensory System (Example) . . . . . . . . . . . . . 229
Table 11.1 Main characteristics of IEEE 802.15.1 and
802.15.4 . . . . . . . . . . . . . . . . . . . . . . . 237
Table 11.2 XMesh power configuration matrix . . . . . . . . 248
Table 11.3 XMesh performance summary table . . . . . . . . 248
Table 11.4 Ambiental Monitoring Summary . . . . . . . . . . 252
Table 12.1 Symbols Definition and Corresponding I.S.
Measurement Units . . . . . . . . . . . . . . . . . 266
Table 13.1 Revenue, costs and business potential for partners
along M2M value chain . . . . . . . . . . . . . . . 303

xxix
 List of Abbreviations

ABS Acrylonitrile butadiene styrene

ABS American Bureau of Shipping
ADC analog to digital converter
AES Advanced Encryption Standard
AI Artificial Intelligence
ANN Artificial Neural Network
AquaRET Aquatic Renewable Energy Technologies
ASN Ambient sensor network
AUV Autonomous Underwater Vehicle
AWS Archimedes Waveswing
BCU Behavior Control Unit
BOP Blow-Out Preventer
BSN Body sensor network
CAN Controller Area Network
CCD charge-coupled device
CCF Conscious Cognitive Functions
CCMP Counter Mode/CBC-MAC Protocol
CCSDBS computer-aided floating dock ballasting process control and
monitoring system
CMF continuous max-flow algorithm
CO Carbon monoxide
CO2 Carbon dioxide
COIL Columbia Object Image Library
CP characteristic points
CRL Certficate Revocation List
CS(SC) system of coordinates
CSMA/CA Carrier sense multiple access/collision avoidance
CU Communication Unit
DB database
DC Direct Current
DHCP Dynamic Host Configuration Protocol
DHSS Direct hopping spread spectrum
DMA decision-making agency
DMP decision-making person

xxxi
xxxii List of Abbreviations

DNS Domain Name System

ECG Electrocardiogram
EE external environment
ESC Electronic Speed Control
ESN Emergency sensor network
EU European Union
FHSS Frequency hopping spread spectrum
FIM Fisher Information Matrix
FMN Femtocell management node
FPSO Floating production storage and offloading
FSM frame-structured model
GIS geographic information systems
GM graphical model
GM Ship Metacentric Height
GND Ground
GPS Global Positioning System
GSM Global System for Mobile Communications
GUI Graphical User Interface
H2 MAS holonic homogenous multi-agent system
HLAU High-level Knowledge Acquisition Unit
HMI human-machine interfaces
HTPS hand position tracking system
HTTPS Secure Hypertext Transfer Protocol
ICT Information and Communication Technology
IE internal environment
IEEE Institute of electrical and electronics engineers
IMI intelligent multimodal interface
IMM intellectual multy modal system
IMU Inertial Measurement Unit
INE information environment agency
IP Internet Protocol
IRR ideal rational robot
IS International System (Measuring)
ISM Industrial Scientific Medical
ISO International Standardization Organization
ITTC International Towing Tank Conference
JRE Java Runtime Environment
KB knowledge base
LASCOS loading and safety control system
LHP left half-plane
LKAU Low-level Knowledge Acquisition Unit
LNG liquefied natural gas
 List of Abbreviations xxxiii

LPG liquefied petroleum gas

M2M Machine To Machine
MAS multi-agent system
MEE combination of the EE
MEMS Microelectromechanical Systems
MFM motion shape models
MIE particular IE
MIT Massachusetts Institute of Technology
MiWi Wireless protocol designed by Microchip Technology
MM mathematical model
MMI man-machine interface
MOE model of the OE
MP Multilayer Perceptron
MPNN Multilayer Perceptron Neural Network
NEM navigation environment monitoring
NN Neural Network
NU Navigation Unit
OCSP Online Certificate Status Protocol
ODA Operations Data Aquisition
OE objects of the environment
OpenCV Open Computer Vision
OPI operator interface agency
OWC OscillatingWater Column
PC Personal computer
PCB printed circuit board
PCBs Printed Circuit Boards
PID Proportional-Integral-Derivative
PPA polymetric perceptive agency
PT Process Transparency
RADIUS Remote Authentication Dias-in User Service
RC Radio Control
RCS robot control system
RF Radio Frequency
RHP right half-plane
RMS Root Mean Square
ROV Remotely (Underwater) Operated Vehicle
RSN Robust Secure Network
RSU Remote Software Update
SADCO systems for automated distant (remote) control
SDS slip displacement sensor
SIFT Scale-invariant feature transform
SIM Subscriber Identity Module
xxxiv List of Abbreviations

SMA sensory monitoring agency

SNR Signal to Noise Ratio
SPM ship state parameters monitoring
SSM sea state monitoring
Tbps Terrabytes per second
TDR time domain reflectometry
THM tracking the head motion
TMS Tether Management System
TSHP system for tracking hand movements
TSP telesensor programming
UART Universal Asynchronous Receiver Transmitter
UAV Unmanned Aerial Vehicle
UCF Unconscious Cognitive Functions
UMTS Universal Mobile Telecommunications System
UPEC University Paris-Est Creteil
UR Underwater Robot
USART Universal Synchronous/Asynchronous Receiver Transmitter
USB Universal Serial Bus
USD United States Dollars
UUV Unmanned Underwater Vehicle
VB.NET Visual Basic .NET
VSM virtual ship model
WCM weather conditions model
WCS Windows Color System
WIFI Wireless Fidelity
WLAN Wireless Local Area Network
 1
A Modular Architecture for Developing
Robots for Industrial Applications

A. Faíña1 , F. Orjales2 , D. Souto2 , F. Bellas2

and R. J. Duro2
1 IT
University of Copenhagen, Denmark
2 Integrated
Group for Engineering Research, University of Coruna, Spain
Corresponding author: A. Faíña <[email protected]>

Abstract
This chapter is concerned with proposing ways to make feasible the use of
robots in many sectors characterized by dynamic and unstructured environ-
ments. In particular, we are interested in addressing the problem through a new
approach, based on modular robotics, to allow the fast deployment of robots
to solve specific tasks. A series of authors have previously proposed modular
architectures, albeit mostly in laboratory settings. For this reason, their designs
were usually more focused on what could be built instead of what was
necessary for industrial operations. The approach presented here addresses the
problem the other way around. In this line, we start by defining the industrial
settings the architecture is aimed at and then extract the main features that
would be required from a modular robotic architecture to operate successfully
in this context. Finally, a particular heterogeneous modular robotic architec-
ture is designed from these requirements and a laboratory implementa-
tion of it is built in order to test its capabilities and show its versatility
using a set of different configurations including manipulators, climbers and
walkers.

Keywords: Modular robots, industrial automation, multi-robot systems.

Advances in Intelligent Robotics and Collaborative Automation, 1–26.


c 2015 River Publishers. All rights reserved.
2 A Modular Architecture for Developing Robots for Industrial Applications

1.1 Introduction
There are several industrial sectors, such as shipyards or construction, where
the use of robots is still very low. These sectors are characterized by presenting
dynamic and unstructured work environments where the work is not carried out
in a chain production line, but rather, the workers have to move to the structures
that are being built and these structures change during the construction process.
Basically, these are the main reasons to explain the low level of automation in
these sectors. Despite this, there are some cases in which robot systems have
been considered in order to increase automation in these areas. However, they
were developed for very specific tasks, that is, as specialists. Some examples
are robots for operations such as grit-blasting [1, 2], welding [3], painting
[4, 5], installation of structures [6, 7] or inspection [8, 9]. Nevertheless, their
global impact on the sector is still low [10]. The main reason for this low
penetration is the high cost of the development of a robot for a specialized
task and the large number of different types of tasks that must be carried out
in these industries. In other words, it is not practical to have a large group of
expensive robots, each one of which will only be used for a particular task
and will be doing nothing the rest of the time.
In the last few years, in order to increase the level of automation in the
aforementioned environments, several approaches have been proposed based
on multi-component robotics systems as an alternative to the use of one robot
for each task [11–13]. These approaches seek to obtain simple robotic systems
capable of adapting, easily and quickly, to different environments and tasks
according to the requirements of the situation.
Multi-component robotic systems can be classified into three categories:
distributed, linked and modular robots [14]; however, in this work, only the
last category will be taken into account. Thus, we explore an approach based
on modular robotics, which basically seeks the re-utilization of pre-designed
robotic modules. We want to develop an architecture that with a small set of
modules can lead to many different types of robots for performing different
tasks.
In the last two decades, several proposals of modular architectures for
autonomous robots have been made [15, 16]. An early approach to modular
architectures resulted in what was called ‘modular mobile robotic systems.’
These robots can move around the environment, and they can connect to
one another to form complex structures for performing tasks that cannot be
carried out by a single unit. Examples are CEBOT [17] or SMC-Rover [18].
Another type of modular architecture is lattice robots. These robots can form
 1.1 Introduction 3

compact three-dimensional structures or lattices over which one module or a

set of them can move. Atron [19] and Tetrobot systems [20] are examples of
this architecture.
Adifferent approach to modularity is provided by the so-called chain-based
architecture, examples of which are modular robots such as Polybot [21], M-
TRAN [22] or Superbot [23]. This kind of architecture has shown its versatility
in several tasks such as carrying or handling payloads, climbing staircases or
ropes or locomotion in long tests or in sandy terrains [24–26]. In addition,
some of them were designed specifically for dynamic and unstructured envi-
ronments. This is the case of the Superbot system, which was developed for
unsupervised operation in real environments, resisting abrasion and physical
impacts, and including enhanced sensing and communications capabilities.
However, and despite the emphasis on real environments, they are mostly
laboratory concept testing approaches with an emphasis on autonomous robots
and self-reconfigurable systems rather than on. That is, these architectures
were not designed to work in industrial settings and, consequently, their
components and characteristics were not derived from an analysis of the needs
and particularities of these environments. In fact, they are mostly based on the
use of a single type of module to simplify their implementation. Additionally,
these homogeneous architectures lead to the need of using a large number of
modules to perform some very simple tasks.
On the other hand, we can find another expression of modular robotics,
which appears as a result of the addition of modularity to robot architectures.
An example is modular manipulators which have mostly been studied for
their use in industrial environments. These types of manipulators can be re-
coupled to achieve, for example, larger load capacities or to extend their
workspace. Most of them can obtain a representation of their morphology
or configuration and automatically obtain their direct and inverse kinematics
and dynamics. There are homogeneous architectures and there are also
architectures with modules specialized in different movements but mainly
with rotational joints. Nevertheless, they are usually aimed at static tasks
[27, 28] and are much less versatile than real complete modular architectures.
In this line, companies such as Schunk Intec Inc or Robotics Design Inc.
are commercializing products inspired by this last approach. Both companies
have developed modular robotic manipulators with mechanical connections,
but these manipulators still need an external control unit configured with the
arm’s topology.
Currently, new research lines have emerged proposing models that take
characteristics of the two areas commented above. For example, some research
4 A Modular Architecture for Developing Robots for Industrial Applications

groups have begun to propose complete versatile heterogeneous modular

systems that are designed with industrial applications in mind. An example
of this approach is the work of [29] and their heterogeneous architecture.
These authors propose a heterogeneous architecture, but in its development,
they concentrate on using spherical actuators with 3 degrees of freedom and
with a small number of attachment faces in each module. Similarly, other
authors have proposed the use of a modular methodology to build robots
flexibly and quickly with low costs [30]. This architecture is based on two
different rotational modules and several end-effectors such as grippers, suckers
and wheels or feet. It has shown its strong potential in a wall-climbing robot
application [31]. These approaches are quite interesting, but they still lack
some of the features that would be desirable from a real industrially usable
heterogeneous modular architecture. For instance, the actuator modules in the
first architecture are not independent; they need a power and communications
module in order to work. The second system only allows serial chain topolo-
gies, which reduces its versatility, or the robot is not able to recognize its own
configuration in both architectures.
In this chapter, we are going to address in a top-down manner the
main features a modular robotic system or architecture needs to display in
order to be adequate for operation in dynamic and unstructured industrial
environments. From these features, we will propose a particular architecture
and will implement a reduced scale prototype of it. To provide an idea of
its appropriateness and versatality, we will finally present some practical
applications using the prototype modules.
The rest of the chapter is structured as follows: Section 2 is devoted to
the definition of the main characteristics the proposed architecture should
have to operate in industrial environments and what design decisions will
be taken. Section 3 contains different solutions we have adopted through
the presentation of a prototype implementation. Section 4 shows different
configurations that the architecture can adopt. Finally, Sections 5 and 6
correspond to the introduction of this architecture in real environments and
the main conclusions of the chapter, respectively.

1.2 Main Characteristics for Industrial Operation and

Design Decisions
Different aspects need to be kept in mind to decide on a modular robotic
architecture for operation in a set of industrial environments. On the one
hand, it is necessary to determine the types of environments the architecture
 1.2 Main Characteristics for Industrial Operation and Design Decisions 5

is designed for and their principal characteristics, the missions the robots will
need to perform in these environments and the implications these have on
the motion and actuation capabilities of the robots. Obviously, there are also
a series of general characteristics that should be fulfilled when considering
industrial operation in general. Consequently, we will first start by identifying
here the main features and characteristics a modular architecture should
display in order to be able to handle a general dynamic and unstructured
industrial environment. This provides the requirements to be met by the
architecture so that we can address the problem of providing a set of solutions
to comply with these requirements. An initial list of required features would
be the following:
• Versatility: The system has to allow to easily build a large number of
different configurations in order to adapt to specific tasks;
• Fast deployment: The change of configuration or morphology has to
be performed easily and in a short time so that robot operation is not
disrupted;
• Fault tolerance: In case of the total failure of a module, the robot has to
be able to continue operating minimizing the effects of this loss;
• Robustness: The modules have to be robust to allow working in dirty
environments and resisting external forces;
• Reduced cost: The system has to be cheap in terms of manufacturing and
operating costs to achieve an economically feasible solution;
• Scalability: The system has to be able to operate with a large num-
ber of modules. In fact, limits on the number of modules should be
avoided.
To fulfil these requirements, a series of decisions were made. Firstly, the
new architecture will be based on a modular chain architecture made up of
heterogeneous modules. This type of architecture has been selected because
it is well known that it is the general architecture that maximizes versatility.
On the other hand, using homogeneous modules is the most common option
in modular systems [15, 16, 21–23], because it facilitates module reuse.
However, it also limits the range of possible configurations and makes the
control of the robot much more complex. In the types of tasks we are
considering here, there are several situations that would require a very simple
module (e.g., a linear displacement actuator), but which would be very
difficult (complex morphology), or even impossible in some cases, to obtain
using any of the homogeneous architectures presented. Thus, for the sake
of flexibility and versatility, we have chosen to use a set of heterogeneous
6 A Modular Architecture for Developing Robots for Industrial Applications

modules (specialized modules for each type of movement). This solution

makes it easier for the resulting robots to perform complex movements
as complex kinematic chains can be easily built by joining a small set of
different types of modules. Moreover, each module was designed to perform
a single basic movement, that is, only one degree of freedom is allowed.
This permits using simple mechanisms within the modules, which increases
the robustness of the system and reduces the operating and manufacturing
costs.
Having decided on the nature of the architecture, now the problem is to
decide what modules would be ideal in terms of having the smallest set of
modules that covers all possible tasks in a domain. In addition, it should be
taken into account that the number of different types of modules needs to
be low in order to accomplish the scalability and reduced production cost
requirements. To do this, we chose to follow a top-down design strategy.
To this end, we studied some typical unstructured industrial environments
(shipyards) and defined a set of general missions that needed automation.
These missions were then subdivided into tasks and these into operations or
sub-tasks that were necessary. From these we deduced the kinematic pairs and
finally a simple set of actuator and end-effector modules that would cover
the whole domain was obtained. This approach differentiates the architecture
presented here from other systems, which are usually designed with a bottom-
up strategy (the modules are designed as the first step and then the authors try
to figure out how they can be applied).
We have only considered five general types of modules in the architecture:
• Actuators: Modules with motors to generate the robot’s motions;
• Effectors: Modules to interact with the environment such as magnets,
suckers or grippers;
• Expansion: Modules that increase computational capabilities, memory
or autonomy through batteries;
• Sensors: Modules to measure and obtain data from the environment such
as cameras and infrared or ultrasonic sensors;
• Linkers: Modules used to join other modules mechanically.
The architecture incorporates these five types of modules, but in this work, we
have focused only on the actuator modules. They are the ones around which
the morphological aspects of the robots gravitate, and we only employ other
modules when strictly necessary to show application examples. Therefore,
each module includes a processing unit, one motor, a battery, capabilities
to communicate with other modules and the necessary sensors to control its
 1.2 Main Characteristics for Industrial Operation and Design Decisions 7

motions. This approach permits achieving a fast deployment of functional

robots and their versatility as compared to cases where they require external
control units.
The process followed to decide on the different actuator modules corre-
sponds with a top-down design process as presented in Figure 1.1. As a first
step, we have considered three basic kinds of general mission the modular
robot could accomplish. These are the surface, linear and static missions (top
layer). Surface missions are those related with tasks requiring covering any
kind of surface (like cleaning a tank). Linear missions are those implying
a linear displacement (like weld inspection) and Static missions are those
where the robotic unit has a fixed position (like an industrial manipulator).

Figure 1.1 Diagram of the selected missions, tasks and sub-tasks considered, and the required
actuators and effectors.
8 A Modular Architecture for Developing Robots for Industrial Applications

The next layer shows the set of possible particular tasks we have considered
as necessary according to the previous types of mission, such as grit-blasting,
tank cleaning etc. The sub-task layer represents the low-level operations the
modular system must carry out to accomplish the task of the previous layer.
The next layer represents the kinematic pairs that can be used to perform
all the sub-tasks of the last layer. As mentioned above, these pairs only
have one degree of freedom. In this case, we have only chosen two kinds
of kinematic pairs: prismatic and revolution joints. Nevertheless, each joint
was implemented in two different modules in order to specialize the modules
to different motion primitives. For the prismatic joint, we have defined a
telescopic module with a contraction/expansion motion and a slider module
with a linear motion over its structure. The revolution joint also leads to
two specialized modules: a rotational module where the rotational axis goes
through the two parts of the module, like in wheels or pulleys, and a hinge
module. Finally, in the last layer we can see five examples of different effector
modules.
Once the actuator modules have been defined, we have to specify the shape
or morphology and the connecting faces of each module. Also, and again to
increase the versatility of the architecture, each module has been endowed with
a large number of attachment faces. This also permits reducing the number
of mechanical adapters needed to build different structures. The distribution
of the attachment faces will be located on cubic nodes or connection bays
within each module. This solution allows creating complex configurations,
even closed chains, with modules that are perpendicular, again increasing the
versatility of the architecture.
These mechanical connections have to be easily operated in order to
allow for the speedy deployment of different configurations. To this end,
each attachment face has been provided with mechanisms for transmitting
energy and communications between modules in order to avoid external wires.
We have also included mechanisms (proprioceptors) that allow the robot to
know its morphology or configuration, that is, what module is attached to what
face. This last feature is important because it allows the robot to calculate its
direct and inverse kinematics and dynamics in order to control its motion in
response to high-level commands from an operator.
The robots developed have to be connected to an external power supply
with one cable to guarantee the energy needed by all the actuators, effectors and
sensors. Nevertheless, the energy is shared among the modules to avoid wires
form module to module. In addition, each module contains a small battery
to prevent the risk of failure by a sudden loss of energy. These batteries,
 1.3 Implementation of a Heterogeneous Modular Architecture Prototype 9

combined with the energy bus between the modules, allow the robot to place
itself in a secure state, maximizing the fault tolerance and the robustness of the
system.
Finally, for the sake of robustness, we decided that the communications
between modules should allow three different communication paths: a fast and
global channel of communications between all the modules that make up a
robot, a local channel of communications between two attached modules and a
global and wireless communication method. These three redundant channels
allow efficient and redundant communications, even between modules that
are not physically connected or when a module in the communications path
has failed.
Summarizing, the general structure of a heterogeneous modular robotic
architecture has been obtained from the set of requirements imposed by
operation in an industrial environment and the tasks the robots must perform
within it. It turns out that given the complexity of shipyard environments, on
which the design was based, the design decisions that were made have led to
an architecture that can be quite versatile and adequate for many other tasks
and environments. In the following section, we will provide a more in-depth
description of the components of the architecture and their characteristics as
they were implemented for tests.

1.3 Implementation of a Heterogeneous Modular

Architecture Prototype
In the previous section, the main features and components of the developed
architecture were presented. Here we are going to provide a description of
the different solutions of actuator modules we have adopted through the
presentation of a prototype implementation. Throughout this section, the
design and morphology of the modules will be explained as well as the
different systems needed for it to operate, such as the energy supply system,
communications, control system, etc.
Figure 1.2 displays some of the different types of modules that were
developed. On the left part, it presents some of the effectors, on the top a
linker, a slider on the right and a rotational module, a telescopic module and
a hinge in the center. The different modules (actuators, linkers and effectors)
have been fully designed and a prototype implementation has been built for
each one of them. They all comprise nodes built using fiber glass from milled
printed circuit boards (PCBs). These parts are soldered to achieve a solid but
lightweight structure. Each module is characterized by having one or more
10 A Modular Architecture for Developing Robots for Industrial Applications

Figure 1.2 Different types of modules developed in this project: three effectors on the left
part, a linker on the top, a slider on the right, and in the middle there is a rotational module, a
hinge module and a telescopic module.

nodes, which act as connections bays. The shape of the nodes varies depending
on the type of module (e.g., it is a cube for the nodes of the slider and telescopic
modules). All of the free sides of these nodes provide a connection mechanism
that allows connecting them to other modules. The size of the nodes without
the connection mechanism is 48x48x48 mm; it is 54x54x54 mm including the
connectors.

1.3.1 Actuator Modules

To develop the prototype of the architecture, four different types of actuator
modules have been built in accordance to the main features of the architecture
described in the previous section. The modules only present one degree
of freedom in order to increase robustness and they have different types
of joints so that it is easy to build most of the kinematic chains used by
real robotic systems in industry. To this end, two linear actuators (slider
and telescopic modules) and two rotational actuators (rotational and hinge
modules) have been developed. In the case of linear actuators, the slider
module has a central node capable of a linear displacement between the
end nodes. Any other module can be connected to this central node. The
telescopic module only has two nodes and the distance between them can be
modified.
 1.3 Implementation of a Heterogeneous Modular Architecture Prototype 11

On the other hand, the rotational modules have two nodes and allow
their relative rotation. These modules are differentiated by the position of
the rotation shaft. Whereas the rotational axis of the rotation module goes
through the center of both modules, in the hinge it is placed in the union of
both nodes and perpendicularly to the line connecting their centers. The main
characteristics of the actuator modules are described in Table 1.1.

1.3.1.1 Slider module

This module has two end nodes that are joined together using three carbon
fiber tubes and an additional node that slides along the tubes between the end
nodes. The distance between the end nodes is 249 mm and the stroke of the
slider node is 189 mm. One of the end nodes has a servo with a pulley, which
moves a drive belt. The node on the other end has the return pulley and the
slider node is fixed to the drive belt. The central node contains the electronics
of the module, with power and data wires connecting it to one of the end nodes.
There is also a mechanism that coils the wires to adapt them to the position of
the slider node.

1.3.1.2 Telescopic module

The telescopic module has two nodes and the distance between them can
increase or decrease. Each node has two carbon fiber tubes attached to it.
There is an ABS plastic part at the end of the tubes. These parts have two
holes with plain bearings to fit the tubes of the other node. One node contains
a servo with a drive pulley and the return pulley is in the ABS part of this
node. The drive belt that runs in these pulleys is connected to the ABS part of
the opposite node. The other node has the electronic board.

1.3.1.3 Rotational module

This module has two nodes that can rotate with respect to each other. A low
friction washer between the nodes and a shaft prevents misalignments. One

Table 1.1 Actuator Modules

Slider Telescopic Rotational Hinge
Type of movement Linear Linear Rotational Rotational
Stroke 189mm 98mm 360º (1 turn) 200º
Nº nodes 3 2 2 2
Nº connection faces 5–4-5 5–5 5–5 1–1
per node
Weight 360g 345g 250g 140g
12 A Modular Architecture for Developing Robots for Industrial Applications

node carries a servo with a gear that engages another gear coupled to the shaft.
The reduction ratio is 15:46. The servo is modified and its potentiometer is
outside attached to a shaft that is operating at a 1:2 ratio with respect to the
main shaft. This configuration permits rotations of the module of 360º.

1.3.1.4 Hinge module

This module does not have any connection bay in its structure, only one
connection mechanism in each main block. A shaft joins two main parts built
from milled PCBs. These parts rotate relative to each other. The reduction of
the servo to the shaft is 1:3. The potentiometer of the servo is joined to the
shaft to sense the real position of the module.

1.3.2 Connection Mechanism

Different types of physical couplings between modules can be found in the
literature, including magnetic couplings, mechanical couplings or even shape
memory wires. In this work, we have decided to use a mechanical connection
due to the high force requirements in some tasks and due to the power
consumption of other options, like in the case of magnetic couplings.
Several mechanical connectors have been developed for modular robots,
but most designers focus their efforts in the mechanical aspects, paying less
attention to power transmission and communications between modules. Here
we have designed a connection mechanism that is able to join two modules
mechanically and, at the same time, transmit power and communications.
Currently, the connector is manually operated, but its automation is under
development.
The connector design can be seen in Figure 1.2 and it has two main parts: a
printed circuit board and a resin structure. The resin structure has four pins and
four sockets to allow four connections in a multiple of 90 degrees like in [16]
and [27]. Inside the resin structure, there is a PCB that can rotate 15 degrees.
The PCB has to be forced to fit inside the resin structure, so the PCB remains
fixed. When two connectors are faced, the rotation of the PCB of one connector
blocks the pins of the other one, and vice versa. The space between the pins
of the two connectors is the same as the thickness of the two connector PCBs.
The PCB has four concentric copper tracks on the top side. A mill breaks
these tracks in order to provide a cantilever. A small quantity of solder is
deposited in the end of the cantilever track. When two connectors are attached,
this solder forces the cantilever tracks to bend, so a force is generated. This
force maintains the electrical contacts fixed even under vibrations.
 1.3 Implementation of a Heterogeneous Modular Architecture Prototype 13

Two of the tracks are wider than the other two because they are employed
to transmit power (GND and +24V). The other two tracks are employed to
transmit data: a CAN bus and local asynchronous communications. The local
asynchronous communications track in each connector is directly connected to
the microcontroller, while the other tracks are shared for all the connectors of
the module. To share these tracks in the node, we choose a surface mount
and insulating displacement connector placed at the bottom of the PCB.
This solution is used to serially connect the PCBs of the node together in
a long string and it allows two modules on the same robot to communicate
even in the case of a failure in a module in the path of the message.

1.3.3 Energy
A need for the modular system to require a wire or tether to obtain power
or perform communications would limit the resulting robots’ motions and
their independence. Therefore, one aim of this work is for the architecture
to allow for fully autonomous modular robots. This is achieved by means
of the installation of batteries in each module and, when the robot needs
more power, expansion modules with additional batteries can be attached
to it. However, in industrial environments it is often the case that the tools
the robots need to use do require cables and hoses to feed them (welding
equipment, sandblasting heads, etc.) and, for the sake of simplicity and length
of time the robot can operate, it makes a lot of sense to use external power
supplies. For this reason, the architecture also allows for tethered operation
when this is more convenient, making sure that the power line reaches just
one of the modules and then it is internally distributed among the rest of the
modules.
The modules developed in this work are powered at 24V, but each module
has its own dc converter to reduce the voltage to 5V to power the servomotors
and the different electronic systems embedded in each module.

1.3.4 Sensors
All of the modules contain specific sensors to measure the position of their
actuator. To this end, the linear modules have a quadrature encoder with
0.32 mm accuracy in their position. The rotational modules are servo con-
trolled, so it is not necessary to know the position of the module. But, in order
to improve the precision of the system, we have added a circuit that senses
the value of the potentiometer after applying a low pass filter.
14 A Modular Architecture for Developing Robots for Industrial Applications

Furthermore, all the modules have an accelerometer to provide their

spatial orientation. In addition, the local communications established in each
attachment face permit identifying the type and the face of the module that
is connected to it. This feature, combined with the accelerometer, allows
determining the morphology and attitude of the robot without any external
help.
All the above-mentioned sensors are particular to each individual module.
It means that they only get the data from the module as well. Nevertheless, to
perform some tasks (welding, inspection, measuring, etc.), it is necessary to
provide to the robot with specific sensors such as camera, ultrasound sensor or
whatever. These specific sensor modules are attached to the actuator module
that requires it. They are basically nodes (morphologically similar to the
rest of the nodes in most modules) with the particular sensor and the pro-
cessing capabilities to acquire and communicate the data from the particular
sensor.

1.3.5 Communications
One of the most difficult tasks in modular robotics is the design of the
communications systems (local and global). On the one hand, it has to ensure
the adequate coordination between modules, and on the other hand, it has to
be able to respond quickly to possible changes in the robot’s morphology.
That is, it has to adapt when a new module is attached, unattached or even
when one module fails. The robot’s general morphology has to be detected
through the aggregation of the values of the local sensing elements in each
module as well as the information they have on the modules they are linked to.
For this, we use an asynchronous local communications line for inter-module
identification (morphological proprioception).
On the other hand, a CAN bus is used for global communications. It allows
performing tasks requiring a critical temporal coordination between remote
modules. Also, a MiWi wireless communications system is implemented as
a redundant system that is used when we have isolated robotic units or when
the CAN bus is saturated.
Additionally, all the modules, except the rotational one, have a micro-USB
connection to allow communications to an external computer. This feature and
a boot loader allow us to employ a USB memory to load the program without
the use of a programmer for microcontrollers. Figure 1.3 shows the printed
circuit board (PCB) of the slider module containing all the communications
elements.
 1.3 Implementation of a Heterogeneous Modular Architecture Prototype 15

Figure 1.3 Control board for the slider module and its main components.

1.3.6 Control
The system control is responsible for controlling and coordinating all the local
tasks within each module, as well as the behaviour of the robot. To do this,
in this work, each module carries its own electronics board with its micro-
controller (PIC32MX575F512) and a DC/DC converter for power supply.
The micro-controller is responsible of the low-level tasks of the module:
controlling the actuator, managing the communications stacks and measuring
the values of its sensors. As each actuator module has its own characteristics
(number of connection faces, encoder type, etc.) and the available space inside
the modules is very limited, we have developed a specific PCB for each kind
of actuator module. As an example, Figure 1.3 shows the top and bottom side
of the control board for the slider module.
Besides the low-level tasks, this solution permits choosing the type of
control to be implemented: centralized or distributed. While in a distributed
control scheme, each of the modules contributes to the final behaviour through
the control of its own actions depending on its sensors or communications to
other modules. In a centralized control scheme, one of the modules would
be in charge of controlling the actions of all the other modules, with the
advantage of having redundant units in case of failure. Additionally, all
modules employ the CAN bus to coordinate their actions and to synchronize
16 A Modular Architecture for Developing Robots for Industrial Applications

their clocks. Obviously, this architecture allows for any intermediate type of
control scheme.

1.4 Some Configurations for Practical Applications

In this section, we will implement some example configurations using the
architecture to show how easy it is to build different types of robots as well as
how versatile the architecture is. For the sake of clarity and in order to show the
benefits of a heterogeneous architecture, we will show simple configurations
developed with only a few modules (videos of these configurations can be
found in vimeo.com/afaina/ad-hoc-morphologies).
All the experiments were carried out with the same setup. First, the modules
were manually assembled in the configuration to test and we connected one
cable for power supply and an USB cable to connect one module to a laptop.
After powering up the system, the module that communicates to the laptop
is selected as a master module. This master module uses the CAN bus to
find other connected modules. Then, it uses the asynchronous communica-
tions and the orientation of each module to discover the topology of the
robot.

1.4.1 Manipulators
One of the most important pillars of industrial automation are manipulators.
Traditional manipulators present a rigid architecture, which complicates their
use in different tasks, and they are very heavy and big to be transported in
dynamic and unstructured environments. Nevertheless, modular manipulators
can be very flexible as they can be entirely reconfigured to adapt to a specific
task and the modules can be transported easily across complex environments
and then they can be directly assembled on the workplace.
The configuration choice of the manipulator is highly application depen-
dent and it is mostly determined by the workspace shape and size, as well as
other factors such as the load to be lifted, the required speed, etc. For instance,
the different types of modules in the architecture can also be used to easily
implement spherical or polar manipulators. These type of manipulators present
a rotational joint at their base and a linear joint for the radial movements
as well as another rotational joint to control their height. Thus, a spherical
manipulator is constructed using just five modules as shown in the pictures of
Figure 1.4. This robot has a magnetic effector to adhere to the metal surface:
a rotational module, a hinge module and a prismatic module for motion and a
 1.4 Some Configurations for Practical Applications 17

Figure 1.4 Spherical manipulator moving a load from one place to another.

final magnetic effector to manipulate metal pieces. We can see how the robot
is able to take an iron part using the electromagnet placed at the end of the
manipulator and carry it to another place. The whole process takes around
10 seconds.
Another very common type of manipulator is the cartesian robot. They are
constructed using just linear joints and are characterized by a cubic workspace.
The ease with which it is possible to produce speed and position control
mechanisms for them, their ability to move large loads and their great stability
are their major advantages.
An example of a very simple and fully functional cartesian robot is
displayed on the left image of Figure 1.5. It is constructed using only two
linear modules and a telescopic module for the implementation of its motions,
two magnetic effectors to adhere to the metal surface and a smaller magnet
that is used as a final effector. The two large magnets used to adhere the robot
18 A Modular Architecture for Developing Robots for Industrial Applications

Figure 1.5 Cartesian manipulators for Static missions.

to the metal surface provide better stability than the previous spherical robot
and reduce the vibrations on the small magnetic end-effector. In addition, we
could implement a gantry style manipulator, as we can observe on the right
image of Figure 1.5. This gantry manipulator has great stability as it uses
four magnets to adhere to the surface and provides a very stable structure
to achieve a high accuracy positioning of its end-effector. Furthermore, this
implementation can lift and move heavier loads as it has two pairs of modules
working in parallel.

1.4.2 Climber and Walker Robots

The most appropriate configurations to carry small loads or sensors and to
move the robots themselves to the workplace are the so-called climber or
walker robot configurations. Modular robots should be able to get to hard to
reach places and, more importantly, their architecture should allow for their
reconfiguration into appropriate morphologies to move through different types
of terrains, different sized tunnels or over obstacles. This reconfigurability
allows reaching and working in areas where it would be impossible for other
kinds of robots to operate. Consequently, being able to obtain simple modular
configurations that allow for these walking or climbing operations is important,
and in this section we will describe three configurations using our architecture
that allow for this.
 1.4 Some Configurations for Practical Applications 19

One of the most prototypical configuration in modular robots is a serial

chain of hinge modules. It is called a worm configuration and can be employed
for inspection tasks inside pipes using a camera or to pass through narrow
passages. Here, we show on Figure 1.6 that we can achieve a working worm
type robot using two hinge modules of our architecture. The whole sequence
takes around 8 seconds, but the robot’s speed could be increased if we use a
worm configuration with more modules.
Another example of how using this architecture a functional robot climber
can be constructed with just with a few modules is the linear wall climber. This
robot consists in a combination of a slider module for motion and two magnet
effectors to stick to the metal surface. This simple robot, which is displayed
on Figure 1.7. (left), can be used on tasks like measuring ship rib thickness or
inspecting a linear weld.
Obviously, the linear climber is unable to avoid obstacles or to turn. Thus,
a possibility to achieve configurations with greater capabilities is to use a few
more modules. A wall climber robot is shown in Figure 1.7 (right). It can be
constructed through the combination of two slider modules, each one of them
with two magnetic effectors to adhere to the metal surface, a linear module and

Figure 1.6 A snake Robot that can inspect inside a pipe.

Figure 1.7 Climber and Walker Robots for linear and surface missions.
20 A Modular Architecture for Developing Robots for Industrial Applications

Figure 1.8 A Biped Robot able to overpass obstacles.

a hinge module between them. This configuration allows the robot to move
and to turn, making it useful for surface inspection tasks performed with an
ultrasonic sensor or other final effectors.
Approximations that are more complex can be created with better locomo-
tion capabilities using other sets of modules. For example, a well-known way
to move through an environment is by walking. This way of moving also allows
stepping over small obstacles or irregularities. A very simple implementation
of a walking robot is shown in Figure 1.8. This configuration is made up of
two hinge modules, each one of them with a magnetic effector, joined together
by a rotational module. This biped robot is capable of walking over irregular
surfaces, stepping over small obstacles and even of moving from a horizontal
to a slanted surface.

1.5 Towards Industrial Applications

In this work, we have analyzed the main features and characteristics that
a modular architecture should display in order to be able to handle a
general dynamic and unstructured industrial environment: versatility, fast
deployment, fault tolerance, robustness, reduced cost and scalability. Cur-
rently, modular commercial systems have achieved a good fault tolerance,
robustness and reduced cost, but they still lack versatility to operate in
dynamic industrial environments and their deployment needs at least some
hours.
 1.6 Conclusions 21

Here, we have developed a new modular architecture taking into account

these dynamic environments. An initial analysis has shown that some impor-
tant features for an architecture of this type is that it should be an heterogeneous
modular architecture with a high number of standardized connection faces, dif-
ferent channels of communication, common power buses, and an autonomous
and independent control for each module or the robot’s ability to discover
its morphology. In order to test the architecture in useful tasks, we have
implemented some modular prototypes. The results show that we can deploy
complex robots for specific tasks in a few minutes and they can be easily
controlled through a GUI in a laptop. Furthermore, we can deploy different
configurations for similar tasks where we can increase the stability and
accuracy of the robot’s end-effector using parallel robots.
An industrial implementation of this architecture is still in a develop-
ment stage, but it will allow working reliably in dynamic and unstructured
environments. It will have the same features of our architecture but with
an industrial-oriented implementation. The main changes will affect to the
robustness of the modules and the connectors. First, modules will be able to
support loads and momentums generated by the most typical configurations. In
addition, it will be ruggedized to work in real environments, which can present
dust or humidity. Regarding the connectors, they will be able to support high
loads and, at the same time, they will allow the fast deployment of the robot
configurations. We can find one connector with these characteristics in [32]
but, additionally, they will have to distribute the energy and communications
buses. As the robots have to work in environments with a high presence of
ferromagnetic material, such as shipyards, we cannot use the magnetometer
values to calculate the relative orientation of the module. Therefore, we will
include a sensor to measure the relative orientation between the modules.
Finally, one important issue to address is the security of the operators who
work near the robots. Most industrial robots are designed to work in close
environments with a shield for the worker’s protection. Thereby, our modules
will have to be compliant for security reasons. This solution is currently used
by some companies that sell compliant robots able to work in the presence of
humans [33].

1.6 Conclusions
A new heterogeneous modular robotic architecture has been presented which
permits building robots in a fast and easy way. The design of the architecture
is based on the main features that we consider, in a top-down fashion, that a
22 A Modular Architecture for Developing Robots for Industrial Applications

modular robotic system must have in order to work in industrial environments.

A prototype implementation of the architecture was created through the con-
struction of a basic set of modules that allows for the construction of different
types of robots. The modules provide for autonomous processing and control,
one degree of freedom actuation and a set of communications capabilities so
that, through their cooperation, different functional robot structures can be
achieved. To demonstrate the versatility of the architecture, a set of robots
was built and tested for simple operations, such as manipulation, climbing or
walking. Obviously, this prototype implementation is not designed to work
in real industrial environments. Nevertheless, the high level of flexibility
achieved with very few modules shows that this approach is very promising.
We are now addressing the implementation of the architecture in more rugged
modules that allow testing in realistic environments.

References
[1] C. Fernandez-Andres, A. Iborra, B. Alvarez, J. Pastor, P. Sanchez,
J. Fernandez-Merono and N. Ortega, ‘Ship shape in Europe: cooperative
robots in the ship repair industry’, Robotics & Automation Magazine,
IEEE, vol. 12, no. 3, pp. 65–77, 2005.
[2] D. Souto, A. Faiña, A. Deibe, F. Lopez-Peña and R. J. Duro, ‘A robot
for the unsupervised grit-blasting of ship hulls’, International Journal of
Advanced Robotic Systems, vol. 9, no. 82, 2012.
[3] G. de Santos, M. Armada and M. Jimenez, ‘Ship building with rower’,
IEEE Robotics & Automation Magazine, vol. 7, no. 4, pp. 35–43, 2000.
[4] B. Naticchia, A. Giretti and A. Carbonari, ‘Set up of a robotized system
for interior wall painting’, in Proceedings of the 23rd International
Symposium on Automation and Robotics in Construction (ISARC), pp.
3–5, 2006.
[5] Y. Kim, M. Jung, Y. Cho, J. Lee and U. Jung, ‘Conceptual design and
feasibility analyses of a robotic system for automated exterior wall
painting’, International Journal of Advanced Robotic Systems, vol. 4,
no. 4, pp. 417–430, 2007.
[6] S. Yu, S. Lee, C. Han, K. Lee and S. Lee, ‘Development of the curtain
wall installation robot: Performance and efficiency tests at a construction
site’, Autonomous Robots, vol. 22, no. 3, pp. 281–291, 2007.
[7] P. Gonzalez de Santos, J. Estremera, E. Garcia and M. Armada, ‘Power
assist devices for installing plaster panels in construction’, Automation
in Construction, vol. 17, no. 4, pp. 459–466, 2008.
 References 23

[8] C. Balaguer, A. Gimenez and CM Abderrahim. ‘Roma robots for inspec-

tion of steel based infrastructures’, Industrial Robot: An International
Journal, vol. 29, no. 3, pp. 246–251, 2002.
[9] J. Shang, T. Sattar, S. Chen and B. Bridge. ‘Design of a climbing
robot for inspecting aircraft wings and fuselage’, Industrial Robot: An
International Journal, vol. 34, no. 6, pp. 495–502, 2007.
[10] T. Yoshida, ‘A short history of construction robots research & devel-
opment in a Japanese company’, in Proceedings of the International
Symposium on Automation and Robotics in Construction, 2006,
pp. 188–193.
[11] C. A. C. Parker, H. Zhang and C. R. Kube, ‘Blind bulldozing. Multiple
robot nest construction’, In IEEE/RSJ International Conference on
Intelligent Robots and Systems, (IROS 2003), vol. 2, pp. 2010–2015,
2003.
[12] N. Correll and A. Martinoli, ‘Multirobot inspection of industrial
machinery’, Robotics & Automation Magazine, IEEE, vol. 16, no. 1,
pp. 103–112, 2009.
[13] A. Breitenmoser, F. Tâche, G. Caprari, R. Siegwart and R. Moser,
‘Magnebike: toward multi climbing robots for power plant inspection’,
In Proceedings of the 9th International Conference on Autonomous
Agents and Multiagent Systems: Industry track, pp. 1713–1720. Inter-
national Foundation for Autonomous Agents and Multiagent Systems,
2010.
[14] R. J. Duro, M. Graña and J. de Lope, ‘On the potential contributions of
hybrid intelligent approaches to multicomponent robotic system devel-
opment’, Information Sciences, vol. 180 no. 14, pp. 2635–2648, 2010.
Special Section on Hybrid Intelligent Algorithms and Applications.
[15] M. Yim, W. M. Shen, B. Salemi, D. Rus, M. Moll, H. Lipson et al.,
‘Modular self-reconfigurable robot systems’, IEEE Robotics and
Automation Magazine, vol. 14, no. 1, pp. 43–52, 2007.
[16] K. Stoy, D. Brandt and D. J. Christensen, ‘Self-reconfigurable robots:
An Introduction’, MIT Press, 2010.
[17] T. Fukuda, S. Nakagawa, Y. Kawauchi and M. Buss, ‘Self-organizing
robots based on cell structures-cebot’, In IEEE International Workshop
on Intelligent Robots, pp. 145–150, 1988.
[18] A. Kawakami, A. Torii, K. Motomura and S. Hirose, ‘Smc rover:
planetary rover with transformable wheels’, In SICE 2002. Proceed-
ings of the 41st SICE Annual Conference, vol. 1, pp. 157–162 vol.1,
2002.
24 A Modular Architecture for Developing Robots for Industrial Applications

[19] M. W. Jorgensen, E. H. Ostergaard and H. H. Lund, ‘Modular atron:

Modules for a self-reconfigurable robot’, In IEEE/RSJ International
Conference on Intelligent Robots and Systems, pp. 2068–2073, 2004.
[20] G. J. Hamlin and A. C. Sanderson, ‘Tetrobot modular robotics: proto-
type and experiments’. In Intelligent Robots and Systems ’96, IROS
96, Proceedings of the 1996 IEEE/RSJ International Conference on,
vol. 2, pp. 390–395, 1996.
[21] M. Yim, D. Duff and K. Roufas, ‘Polybot: a modular reconfigurable
robot’, in IEEE International Conference on Robotics and Automation,
vol. 1. IEEE; pp. 514–520, 2000.
[22] S. Murata, E. Yoshida, A. Kamimura, H. Kurokawa, K. Tomita
and S. Kokaji, ‘M-tran: Self-reconfigurable modular robotic system’,
IEEE/ASME transactions on mechatronics, vol. 7, no. 4, pp. 431–441,
2002.
[23] B. Salemi, M. Moll and W. Shen, ‘Superbot: A deployable, multifunc-
tional, and modular self-reconfigurable robotic system’, in IEEE/RSJ
International Conference on Intelligent Robots and Systems (IROS),
pp. 3636–3641, 2006.
[24] D. Brandt, D. Christensen and H. Lund, ‘Atron robots: Versatility
from self-reconfigurable modules’, in International Conference on
Mechatronics and Automation, 2007. (ICMA), pp. 26–32, 2007.
[25] N. Ranasinghe, J. Everist and W.-M. Shen, ‘Modular robot climbers’, in
Workshop on Self-Reconfigurable Robots, Systems & Applications in
IEEE/RSJ International Conference on Intelligent Robots and Systems
(IROS), 2007.
[26] F. Hou, N. Ranasinghe, B. Salemi and W. Shen, ‘Wheeled locomo-
tion for payload carrying with modular robot’, IEEE/RSJ International
Conference on Intelligent Robots and Systems (IROS), pp. 1331–1337,
2008.
[27] T. Matsumaru, ‘Design and control of the modular robot system:
Tomms’, in IEEE International Conference on Robotics and Automation
(ICRA), pp. 2125–2131, 1995.
[28] C. J. Paredis, H. B. Brown and P. K. Khosla, ‘A rapidly deployable
manipulator system’, in IEEE International Conference on Robotics and
Automation, pp. 1434–1439, 1996.
[29] J. Baca, M. Ferre and R.Aracil, ‘Aheterogeneous modular robotic design
for fast response to a diversity of tasks’, Robotics and Autonomous
Systems, vol. 60, no. 4, pp. 522–531, 2012.
 References 25

[30] Y. Guan, L. Jiang, X. Zhang, H. Zhang and X. Zhou, ‘Development

of novel robots with modular methodology’, IEEE/RSJ International
Conference on Intelligent Robots and Systems (IROS), pp. 2385–2390,
2009.
[31] Y. Guan, H. Zhu, W. Wu, X. Zhou, Li Jiang, C. Cai et al., ‘A Modular
Biped Wall-Climbing Robot With High Mobility and Manipulating
Function’, IEEE/ASME Transactions on Mechatronics, vol. 18, no. 6,
2013.
[32] M. Nilsson, ‘Connectors for self-reconfiguring robots’, IEEE/ASME
Transactions on Mechatronics, vol. 7, no. 4, pp. 473–474, 2002.
[33] E. Guizzo and E. Ackerman, ‘The rise of the robot worker’, IEEE
Spectrum, vol. 49, no. 10, pp. 34–41, 2012.
 2
The Dynamic Characteristics of a
Manipulator with Parallel Kinematic
Structure Based on Experimental Data

S. Osadchy1 , V. Zozulya2 and A. Timoshenko3

1,2 Facultyof Automation and Energy, Kirovograd National Technical
University, Ukraine
3 Faculty of Automation and Energy, Kirovograd Flight Academy

National Aviation University, Ukraine

Corresponding author: S. Osadchy <[email protected]>

Abstract
The chapter presents two identification techniques which the authors found
most useful in examining the dynamic characteristics of a manipulator with
a parallel kinematic structure as an object of control. These techniques
emphasize a frequency domain approach. If all input/output signals of an
object can be measured, then the first one of such techniques may be used for
identification. In the case when all disturbances can’t be measured, the second
identification technique may be used.

Keywords: Manipulator with parallel kinematics, structural identification,

control system.

2.1 Introduction
Mechanisms with parallel kinematics [1, 2] compose the basis for the
construction of single-stage and multi-stage manipulators. A single-stage
manipulator consists of an immobile basis, a mobile platform and six guide
rods. Each rod can be represented as two semi rods Aij and an active kinematics
pair Bij (Figure 2.1).

Advances in Intelligent Robotics and Collaborative Automation, 27–48.


c 2015 River Publishers. All rights reserved.
28 The Dynamic Characteristics of a Manipulator with Parallel Kinematic

Figure 2.1 Kinematic diagram of single-section mechanism.

We will consider two systems of co-ordinates: inertial O0 X0Y0 Z0 with the

origin in the center of the base O0 and mobile O1 X1Y1 Z1 , with the origin
O1 in the platform center of mass. From Figure 2.1 it is evident, that such
mechanisms consist of thirteen mobile links and eighteen kinematics pairs.
That is why, in accordance with [2], the number of its possible motions
equals six.
Let us propose the following definitions ( j∈[1:6]): l1j – length of rod
number j; Mx ,My ,Mz – projections of the net resistance moment vector on the
axes of the co-ordinate system O0 X0 Y0 Z0 .
Obviously, while lengths l 1, j are changing, then the co-ordinates of the
platform’s center of mass and the projections of the resistance moment vector
are changing too.
From the point of view of automatic control theory, the mechanism with
parallel kinematics belongs to the array of mobile control objects with two
 2.2 Purpose and Task of Research 29

multidimensional entrances (control signals and disturbances) and one output

vector (platform co-ordinates).

2.2 Purpose and Task of Research

The main purpose of this research is to construct a mathematical model
which characterizes the interrelation between control signals, disturbances and
co-ordinates of the platform center of mass on the base of experimental data.
If one assembles the length of rod changes in the vector of control signals
u1 , the projections of force resistance moment changes in the disturbance
vector ψ and the coordinates of the platform center of mass changes in the
output vector x
⎡ ⎤ ⎡ ⎤ ⎡ ⎤
l1,1 Mx Xc
⎢ ⎥
u1 = ⎣ ... ⎦ , ψ = ⎣ My ⎦ , x = ⎣ Yc ⎦ , (2.1)
l1,6 Mz Zc

then the block diagram of the mechanism with parallel kinematics can
be represented as shown on the Figure 2.2 where W u is an operator which
characterizes the influence of the control signals vector u on the output vector x
and W ψ is an operator which describes the influence of the disturbance vector
ψ on the output vector x. In this case, in order to find the mathematical model,
it is necessary to define these operators. If we want to find such operators

Figure 2.2 Block diagram of the mechanism with parallel kinematics.

30 The Dynamic Characteristics of a Manipulator with Parallel Kinematic

based on experimental data, then two variants of the research task can be
enunciated.
The first variant will be applied if the components of vectors u1 , x and ψ
can be measured fully (the complete data). The second variant will be applied
in the case when only the components of vectors u1 and x can be measured
(the incomplete data).
So the research on the dynamics of the mechanism with parallel kinematics
can be formulated as follows: to find transfer function matrices Wu , Wψ and
also to estimate the influence of vectors u1 and ψ on vector x on the base of
known complete or incomplete experimental data.
The solution of such a problem has been found as a result of three stages of:
• The development of algorithms for the structural identification of a
multivariable Dynamic object with the help of complete or incomplete
data;
• Collecting and processing experimental data about vectors u1 , x and ψ;
• The verification of the results of the structural identification.

2.3 Algorithm for the Structural Identification of the

Multivariable Dynamic Object with the Help of the
Complete Data
Let’s suppose the identification object dynamics is characterized by a transfer
function matrix Wob (Figure 2.2), which may have unstable poles. Suppose
that as a result of the processing of regular components of vectors u and x, the
 
Laplace transformations U p and X p are defined

 u1 
U p = L {u} = L ψ
, X p = L {x} . (2.2)

Thus, the Laplace transformation of output vector X p has unstable poles

of vector U p and unstable poles of matrix Wob . Therefore, it is possible to

remove all unstable poles from X p [3] which differ from the unstable poles of

U p and to define a diagonal polynomial matrix W2 such as
 
Y p = W2 · X p . (2.3)

In this case, the interdependence between vectors Yp and Up is expressed
with the help of equation
 2.3 Algorithm for the Structural Identification 31

 
Y p = F1p · U p , (2.4)
where F1p is a transfer function matrix, all the poles of which are located
in the left half-plane (LHP) of complex variables. It is equal to

F1p = W2 · Wob . (2.5)

Consequently, the identification problem consists in determining a physi-
cally implemented matrix F1p that minimizes a quality functional
j∞
1
J= · tr(ε · ε∗ · A) · ds, (2.6)
2·π·j −j∞

where ε – identification error, which is equal to

 
ε = F1p · U p − Y p , (2.7)

A– is a positively defined polynomial weight matrix.

To solve this problem, the ratio (2.7) must be submitted in a vector-matrix
form   
 U
ε = F1p −En · p . (2.8)
Yp
The Hermitian conjugated vector ε∗ from Equation (2.6) is equal to
 
  F1p ∗
ε∗ = U p∗ Y p∗ · . (2.9)
−En

After introducing the expressions (2.8) and (2.9) into the Equation (2.6),
the quality functional can be shown as follows:

1

j∞ 
J = 2·π·j · tr F1p −En ×
−j∞

   (2.10)
   
Up · U p∗ Up · Y p∗ F1p ∗
×     · ·A · ds.
Yp · U p∗ Yp · Y p∗ −En

Thus, the problem of structural identification is reduced to the minimization

of the functional (2.10) on the class of a steady variation matrix of transfer
functions F1p . Such minimization has been carried out as a result of the
32 The Dynamic Characteristics of a Manipulator with Parallel Kinematic

application of the Wiener-Kolmogorov procedure. In accordance with such

procedure [5], the first variation of the quality functional (2.10) has been
defined as
1
δJ = · tr{A0∗ · [A0 · F1p · D − (H0 + H+ + H− )] · L · D∗
2·π·j
× δF1p∗ + δF1p · D · L · [D∗ · F1p∗ · A0∗ − (H0 + H+ + H− )∗ ]
× A0 } · ds, (2.11)
where A0 is a result of the factorization [4] of the matrix A the determinante
pf which has zeros with the negative real parts

A = A0∗ · A0 ; (2.12)
D is a fraction-rational matrix with particularities in the left half-plane (LHP)
which is defined on the basis of the algorithms in articles [3, 4] from the
following equation
 
D · L · D∗ =U p · U p∗ , (2.13)
where L is a singular matrix, each element of which is equal to one; bottom
index * designates the Hermitian conjugate operation; H0 +H+ +H− is a
fraction-rational matrix which is equal to
 
H0 + H+ + H− = A0 · Y p · U p∗ ·D∗−1 · L+ ; (2.14)
L+ is the pseudo inverse to matrix L [5]; matrix H0 is the result of the division;
H+ is a proper fractional rational matrix with poles that are analytic only in the
right half-plane (RHP); H− is a proper fractional rational matrix with poles that
are analytic in LHP. In accordance with the chosen minimization procedure,
a steady and physically realized variation F1p which delivers a minimum to
the functional (2.10) is equal to
F1p = A−1 −1
0 · (H0 + H+ ) · D . (2.15)
If one takes into account matrices W2 , F1p from Equations (2.3) and (2.15),
then an unknown object transfer function matrix Wob can be identified with
the help of the following expression
Wob = W2−1 · F1p . (2.16)
The separation [4] of the transfer function matrix (2.16) makes it possible to
find the unstable part of the object transfer function matrix with the help of
equation
 2.4 Algorithm for the Structural Identification 33

Wob2 = W− , (2.17)
where W− is a fraction-rational matrix with particularities in the RHP.
An algorithm for the structural identification of the multivariable dynamic
object with an unstable part on the base of the vectors u and x emplees the
implementation of the following operations:
• Search the matrix W2 as a result of the left-hand removal of the unstable
poles from Xp , which differ from the poles of Up ;
• Factorization of the weight matrix A from (2.12);
• Identification of the analytical complex variable matrix D
Equation (2.13);
• Calculation of H0 + H+ as a result of the separation (2.14);
• Calculation of F1p on the basis of the Equation (2.15);
• Identifying Wob2 by the separation of the product (2.16).
In this way, we have substantiated the algorithm for the structural identification
of the multivariable dynamic object with the help of the complete experimental
data.

2.4 Algorithm for the Structural Identification of the

Multivariable Dynamic Object with the Help of
Incomplete Data
Let’s suppose that the identification object dynamics is characterized by
a system of ordinary differential equations with constant coefficients. The
Fourier transformation of such system, subject to the zero initial conditions,
can be shown as follows:

P · x = M · u1 + ψ, (2.18)

where P and M are polynomial matrices of the appropriate dimensions; ψ

is the Fourier image of a centered multivariable stationary random process
with the unknown spectral densities matrix Sψψ . Let us admit also that vectors
u and x are the centered multivariable stationary random processes with the
matrices of the spectral and cross-spectral densities Sxx , Suu , Sxu , Sux known
as a result of the experimental data processing. It is considered that the random
process ψ can be formed by a filter with the transfer function matrix Ψ and is
equal to
34 The Dynamic Characteristics of a Manipulator with Parallel Kinematic

ψ = Ψ · Δ, (2.19)
where Δ is the vector of the single δ-correlated “white” noises.
If one takes into account expression (2.19), then Equation (2.18) can be
rewritten as follows:

x = P −1 · M · u1 + P −1 · Ψ · Δ (2.20)

and a transfer function matrix which must be identified can be defined as the
expression  
φ = φ11 φ12 = P −1 · M P −1 Ψ . (2.21)
So, the Equation (2.20) can be simplified to the form

x = φ · y, (2.22)

where y is an extended vector of the external influences

 
y = u  Δ . (2.23)

Thus, the identification problem can be formulated as follows. Using the

records of the vectors x and y, choose the sectional matrix string φ (2.21)
that provides minimum to the following quality functional
j∞
1 
J= tr{Sεε R}ds, (2.24)
j
−j∞

where J is equal to the sum of the identification errors variances as the

elements of the identification errors vector ε

ε = x − φ · y. (2.25)

Sεε is a transposed matrix of the identification errors spectral densities
    
Sεε = Sxx − Syx φ∗ − φSxy + φSyy φ∗ ; (2.26)

  

Syx = Sux SΔx ; (2.27)


Suu Om×n

Syy =  , (2.28)
On×m SΔΔ
 2.4 Algorithm for the Structural Identification 35
 , S  , S  are the transposed spectral density matrices of the vectors x,
Sxx uu yy
 , S  , S  is the transposed cross-spectral density matrices between
u, y; Sxy yx ux
vectors x and y, y and x, u and x; SΔx is a transposed cross-spectral density
matrix which is found on the basis of the Wiener’s factorization [3] of the
additional connection equation
−1
SxΔ SΔx = Sxx − Sxu Suu Sux , (2.29)

R is a positively defined polynomial weight matrix.

An algorithm for the set problem decision, which is grounded in [8] and
allows defining the sought after matrix φ which minimizes the functional
(2.24), has the following form

φ = R0 (K0 + K+ )D−1 , (2.30)

in which matricees R0 and D are results of the Wiener’s factorization [3] of

 so that
matrices R and Syy

R0∗ R0 = R; DD∗ = Syy ; (2.31)

K0 +K+ is a transfer function matrix with the stable poles, which is defined as
a result of the following equation right part separation [7]

K0 + K+ + K− = R0 Syx D∗−1 . (2.32)

An algorithm for the structural identification of a multivariable dynamic object

with the help of the stochastic stationary components of vectors u1 and x
implies the following operations:
• Search for the spectral and cross-spectral densities matrices Sxx , Suu ,
Syy , Sux , Sxu on the base of the experimental data processing;
• Factorization of the weight matrix R from (2.31);
• Factorization of the additional connection Equation (2.29);
• Factorization of the transposed spectral densities matrix (2.28);
• The Wiener’s separation of the matrix (2.32);
• Calculation of matrix φ based on Equation (2.30);
• Identification of matrices φ11 and φ12 with the help of Equation (2.21).
In this way, we substantiate the algorithm for the structural identification
of the multivariable object on the base of the incomplete experimental
data.
36 The Dynamic Characteristics of a Manipulator with Parallel Kinematic

2.5 The Dynamics of the Mechanism with a Parallel

Structure Obtained by Means of the Complete Data
Identification
To identify the models of a dynamics, we used tracks of changes in the compo-
nents of vectors u, ψ and x, obtained as a result of the behavior of a modeling
platform using a virtual model. The case was considered when the motion
platform center of mass O1 remained in the plane of the manipulator symmetry
O0 X0 Y0 . Thus, it is evident that in this case, instead of six rods only three
(Figure 2.1) may be considered and the dimension of vector u (2.1) is equal 3.
As a result of the computational experiment, all the above vectors’
components were obtained and all graphs of their changes (Figures 2.3–2.5)
were built.
For solving of the identification problem, the control and perturbations
vectors were combined into a single vector u of the input signals
T
u= l1,1 l1,2 l1,3 Mx My Mz . (2.33)
In this case, the identification problem is to estimate the order and
the parameters of a differential equations system which characterizes the
mechanism motion.

Figure 2.3 Graphs of changes in the length of the rods.

 2.5 The Dynamics of the Mechanism with a Parallel Structure Obtained 37

Figure 2.4 Graphs of changes in the projections of the resistance moments.

Figure 2.5 Graphs of chances in the coordinates of the platform’s center of mass.
38 The Dynamic Characteristics of a Manipulator with Parallel Kinematic

The state space dynamic model of the mechanism is defined with the help
of the System Identification Toolbox of the Matlab environment. Considering
the structure of vector u, defined by (2.33), allows to obtain the equation of
the hexapod’s state as follows:

ẏ(t) = Ay(t) + Bu u(t) + Bψ ψ(t),

x(t) = Cy(t) + Du u(t) + Dψ ψ(t), (2.34)

where the matrices Bu , Bψ , Du and Dψ are easily determined.
The analysis of the obtained model of dynamics shows that the moving
object is fully controllable and observable.
As a result, the Laplace transformation of the left and right parts of the
Equations (2.34) obtained the following relations

(sEn − A)y(s) = Bu u(s) + Bψ ψ(s)
(2.35)
x(s) = Cy(s) + Du u(s) + Dψ ψ(s),

where y(s), x(s), u(s), ψ(s) – the Laplace image of the vector, En – the identity
matrix, s – the independent complex variable.
After solving the system of Equations (2.35) with respect to the vector
of the initial coordinates of the mechanism x, the following matrices of the
transfer functions Wu and Wψ (Figure 2.2) were obtained

Wu = C (sEn − A)−1 Bu + Du , (2.36)

Wψ = C (sEn − A)−1 Bψ + Dψ . (2.37)

Substituting the appropriate numerical matrices C, Bu , Bψ , Du , Dψ , in
expressions (2.36), (2.37) allowed determining that

 −0.37(s + 1.606)(s2 + 21.03s + 790.2) −0.599(s − 57.51)(s2 − 2.215s + 8.456)

(s + 1.55)(s2 + 3.463s + 117.8) (s + 1.55)(s2 + 3.463s + 117.8)
Wu = 0.29(s + 2.232)(s2 + 21.65s + 529.7) 2.23(s − 14.01)(s2 − 0.67s + 87.82)
(s + 1.55)(s2 + 3.463s + 117.8) (s + 1.55)(s2 + 3.463s + 117.8)

0.92796(s + 0.4741)(s2 − 8.066s + 404.6)


(s + 1.55)(s2 + 3.463s + 117.8)
0.36433(s + 75.92)(s2 − 2.029s + 91.39) , (2.38)
(s + 1.55)(s2 + 3.463s + 117.8)
 2.5 The Dynamics of the Mechanism with a Parallel Structure Obtained 39

 −0.006(s − 2.782)(s2 + 10.9s + 450.5) 0.0023(s + 43.94)(s2 − 3.357s + 8.447)

(s + 1.55)(s2 + 3.463s + 117.8) (s + 1.55)(s2 + 3.463s + 117.8)
Wψ = 0.046892(s − 10.1)(s2 + 3.372s + 133.1) 0.059(s + 0.162)(s2 − 2.875s + 124.3)
(s + 1.55)(s2 + 3.463s + 117.8) (s + 1.55)(s2 + 3.463s + 117.8)

0.0049(s + 57.79)(s + 8.592)(s − 2.702)


(s + 1.55)(s2 + 3.463s + 117.8)
−0.014(s + 22.92)(s2 + 1.572s + 174.3) . (2.39)
(s + 1.55)(s2 + 3.463s + 117.8)

The analysis of the matrix structure of (2.38) and (2.39) and the Bode
diagrams (Figure 2.6) shows that this mechanism can be classified as a multi-
resistant mechanical filter with the input signals and the disturbances energy
bands lying in the filter spectral band pass. The eigen frequency of such a filter
is close to 11s−1 and depends on the moments of inertia and the mass of the
moving elements of the mechanism (Figure 2.1).
The ordinary differential equations dynamics model of the system can
be obtained if you present the transfer function matrices Wu and Wψ as a
product of the polynomial matrices P, M and Mψ with the minimum order
polynomials:
Wu = P −1 M, (2.40)
Wψ = P −1 Mψ . (2.41)
To find the polynomial matrices P and M with the minimum order polynomials,
we propose the following algorithm:
• By moving the poles to the right [3], the transfer function matrix Wu
should be introduced as follows:
−1
Wu = NR DR (2.42)

and the diagonal polynomial matrix DR should be found;

• On the basis of the polynomial matrices NR and DR found by the CMFR
algorithm, substantiated in [7], the unknown matrices P and M should
be identified
−1
P −1 M = NR DR ; (2.43)
• From Equation (2.41) and the known matrices P and Wψ the polynomial
matrix Mψ should be found

Mψ = P Wψ . (2.44)
40 The Dynamic Characteristics of a Manipulator with Parallel Kinematic

Figure 2.6 Bode diagrams of the mechanism with a parallel structure.

 2.6 The Dynamics of the Mechanism with a Parallel Structure 41

Figure 2.7 Block diagram of the mechanism with a parallel kinematics.

The application of the algorithms (2.42) and (2.43) to the original data
represented by expressions (2.38), (2.39) made it possible to obtain the
polynomial matrices P, M, Mψ . Then, application of the inverse Laplace
transform under the zero initial conditions allowed determining the following
system of ordinary differential equations

P0 ẍ + P1 ẋ + P2 x = M0 ü + M1 u̇ + M2 u + Mψ0 ψ̈ + Mψ1 ψ̇ + Mψ2 ψ, (2.45)

where Pi , Mi , ψ i – the numeric matrices.

Representing the dynamic model (2.45) allowed to reconstruct a
block diagram of a parallel kinematics mechanism in the standard form
(Figure 2.7), where P−1 is an inverse of matrix P.

2.6 The Dynamics of the Mechanism with a Parallel

Structure Obtained by Means of the Incomplete
Data Identification
The incomplete experimental data arises when not all entrance signals of
vector u (Figure 2.2) can be measured and recorded. Such a situation appears
during the dynamics identification of a manipulator with controlled diode
motor-operated drive (Figure 2.8). In this case, only signals of the platforms
center of mass set position which form vector u1 and signals of the platform’s
center of mass current position which form vector x are accessible for
measuring. Thus, the task of the identification is to define matrices φ11 , φ12
from Equation (2.21) by records of vectors u1 and x.
42 The Dynamic Characteristics of a Manipulator with Parallel Kinematic

Figure 2.8 Manipulator with a controlled diode motor-operated drive.

Solving this task is achieved as a result of algorithm (2.30–2.32) applied

to the estimations of the spectral and cross-spectral density matrices Suu , Sxx ,
Sux and SΔx . For the illustration of this algorithm application, we used the
records of the «input-output» vectors u, x with the following structure
⎡ ⎤
u1
u = ⎣ u2 ⎦ ; x = [Zc ] , (2.46)
u3
where u1 – the required value of the manipulators’ platform center of mass
O1 projection on the axis O0 X0 (Figure 2.1); u2 – the required value of the
manipulators platform center of mass O1 projection on the axis O0 Y0 ; u3 – the
required value of the manipulators’ platform center of mass O1 projection on
the axis O0 Z0 (Figure 2.1).
As a result of the experiment, all the above vectors’ (2.46) components
were obtained and all the graphs of their changes (Figures 2.9, 2.10) were built.
 2.6 The Dynamics of the Mechanism with a Parallel Structure 43

Figure 2.9 Grapas of the vector u componente changes.

Figure 2.10 Grapas of vector x component changes.

In accordance with this algorithm (2.26–2.32) on the first stage of the

calculations, estimations of matrices Suu , Sxx , Sux were found.
Approximation of such estimations made it possible to construct the fol-
lowing spectral densities matrices with the help of the logarithmic frequency
descriptions method [4]
44 The Dynamic Characteristics of a Manipulator with Parallel Kinematic

⎡ 12(s+0.095)(−s+0.15)
⎤
7.87|s+0.12|2 0 |s2 +0.29s+0.034|2
⎢ |s +0.29s+0.034|2
2 ⎥
 ⎢ 45.5|s+0.075|2
0 ⎥
Suu =⎢ 0 |s2 +0.64s+0.16|2
⎥;
⎣ 12(−s+0.095)(s+0.15) ⎦
2.88|s+0.095|2
|s2 +0.29s+0.034|2 0 |s2 +0.29s+0.034|2

 4.59 |(s + 0.08) (s + 1.3)|2

Sxx = ;
|(s + 0.8) (s2 + 0.29s + 0.034)|2

 4.89(−s+0.1)(s+0.2)(−s+2.1)
Sux = (s+0.8)|s2 +0.29s+0.034|2
0

1.2 (s + 3) |s + 0.095|2
. (2.47)
(s + 0.8)|s2 + 0.29s + 0.034|2
The introduction of the found matrices (2.47) into the Equation (2.29) and
its factorization made it possible to find the cross spectral density SΔx

0.89 (s + 1.6) (s + 0.031)

SΔx = . (2.48)
(s + 0.8) (s2 + 0.29s + 0.034)
 from
The factorization [4] of the transposed spectral densities matrix Syy
expression (2.31) allowed finding the following matrix D

⎡ ⎤
8.87(s+0.1) −0.54
s2 +0.29s+0.034
0 s2 +0.29s+0.034 0
⎢ 6.75(s+0.075) 0 ⎥
⎢ 0 0 ⎥
D=⎢ 1.34(s+0.11)
s2 +0.64s+0.16
1.04(s−0.057) ⎥. (2.49)
⎣ 0 s2 +0.29s+0.034
0 ⎦
s2 +0.29s+0.034
0 0 0 1

Taking into account the dimension of the output co-ordinates vector x

(2.46), we have accepted that matrix R is equal to the identity matrix. At
that rate matrix R0 also equals 1. Substitution of the results (2.47–2.49) in
expression (2.32) and its separation allowed defining that
⎡ ⎤
−0.44(s−2.57)(s+0.12)
(s+0.8)(s2 +0.29s+0.034)
⎢ ⎥
⎢ 0 ⎥
K0 + K+ = ⎢
⎢ 1.95(s+1.16)(s−0.04) ⎥ .
⎥ (2.50)
⎣ (s+0.8)(s2 +0.29s+0.034) ⎦
0.89(s+1.6)(s+0.03)
(s+0.8)(s2 +0.29s+0.034)
 2.7 Verification of the Structural Identification Results 45

Substitution of matrices (2.49), (2.50) in expression (2.30) and taking into

account the vectors u and x made it possible to solve the problem and to find
such matrices φ11 and φ12 as
⎡ −0.33(s+0.61)(s−0.034)
⎤
(s+0.03)(s+0.8)
⎢ ⎥
φ11 = ⎣ 0 ⎦ ; (2.51)
1.87(s+1.12)(s−0.008)
(s+0.03)(s+0.8)

0.89 (s + 1.6) (s + 0.031)

φ12 = (2.52)
(s + 0.8) (s2 + 0.29s + 0.034)
Taking into account the flow diagram on Figure 2.2 and the physical sense
of matrices φ11 and φ12 made it possible to formulate the equation

Wu = φ11 ; Wψ = φ12 . (2.53)

For the definition of the incomplete data identification error, the

Equation (2.26) is used and the error spectral density is found in the form
which is shown below:

 0.023
Sεε = . (2.54)
|s + 0.037|2
The identification error mathematical mean is equal to zero and its relative
variances is equal to


j∞
Sεε ds
−j∞
Eε = = 0.0157. (2.55)

j∞
Sxx ds
−j∞

Obviously it is clear that the main part of the error ε oscillations power
density is concentrated in the area of the infrasonic frequencies. The presence
of such an error is explained by the limited duration of the experiment.

2.7 Verification of the Structural Identification Results

Verification of the identification results was implemented with the help of
the modeling tool SIMULINK from Matlab. The principle of the verifica-
tion of the identification results exactness was based on the comparison
46 The Dynamic Characteristics of a Manipulator with Parallel Kinematic

of vector x records to the results of the platform center of mass position

simulation.
On Figure 2.11, the simulation model block u is designed to create the
set of the input signals. The block m generates a set of projections of the net

Figure 2.11 The scheme simulation model of the mechanism with parallel kinematics.

Figure 2.12 Graphs of the change of coordinates X of the center of mass of the platform.
 2.8 Conclusions 47

Figure 2.13 Graphs of thee changes of Y coordinates of the center of mass of the platform.

resistance moment vector on the axis of the co-ordinate system O0 X0Y0 Z0 .

The output of the block xy formed the vector x of records. The blocks Wu
and Wpsi are designed for storing the matrices of the transfer functions
Wu and Wψ .
According to the simulation results, the grapas (Figures 2.12 and 2.13) are
built. Analysis of these graphs show that they are close enough.

2.8 Conclusions
The conducted research on the mechanism with a parallel structure dynamics
made it possible to obtain the following scientific and practical results:
• Substantiate two new algorithms for the structural identification of the
multivariable moving object dynamic models. The first one of them
allows to define the structure and parameters of a transfer function matrix
of the object with unstable poles using the regular vectors "input-output".
The second one allows identifying not only the model of a mobile object
but also the model of the non-observed stationary stochastic disturbance;
• Three types of models which characterize the dynamics of the manipula-
tor with a parallel kinematics are identified. This allows to use different
modern multidimensional optimal control systems synthesis methods for
designing the optimal mechatronic system;
48 The Dynamic Characteristics of a Manipulator with Parallel Kinematic

• It is shown that the mechanism with a parallel kinematics as an object of

control is a multi-resistant mechanical filter.

References
[1] J. P. Merlet, ‘Parallel robots. Solid mechanics and its application’,
V.74 – Kluwer Academic Publishers, 2000, pp. 394
[2] S. V. Volkomorov, J. P. Hagan, A. P. Karpenko, ‘Modeling and optimiza-
tion of some parallel mechanisms’, Information Technology. Application
2010, No. 5. pp.1–32 (in Russian).
[3] M. C. Davis, ‘Factoring the spectral matrix’, IEEE Trans. Automat.
Cointr. – 1963.- AC-8, N 4, pp. 296–305.
[4] V. N. Azarskov, L. N. Blokhin, L. S. Zhitetsky, ‘The methodology of
constructing optimal systems stochastic stabilization’, Monograph, Ed.
Blokhin L. N. - K.: NAU Book Publishers, 2006, pp. 440 - Bibliography:
pp. 416–428 (in Russian).
[5] F. R. Gantmakher, ‘Theory matrits.-4th ed’, Nauka, 1988. p. 552 (in
Russian).
[6] V. Kucera ‘Discrete line control: the polynomial equation approach’.
Praha: Akademia, 1979, pp. 206
[7] F. A. Aliev, V. A. Bordyug, V. B. Larin, ‘Time and frequency methods
for the synthesis of optimal regulators’, Baku: Institute of Physics of the
Academy of Sciences, 1988, pp. 46 (in Russian).
 3
An Autonomous Scale Ship Model for
Parametric Rolling Towing Tank Testing

M. Míguez González, A. Deibe, F. Orjales, B. Priego and F. López Peña

Integrated Group for Engineering Research, University of A Coruña, Spain

Corresponding author: M. Míguez González <[email protected]>

Abstract
This chapter presents the work carried out for developing a self-propelled
scale ship model for model testing, with the main characteristic of not having
any material link to a towing device to carry out the tests. This model has been
fully instrumented in order to acquire all the significant raw data, process them
onboard and communicate with an inshore station.

Keywords: Ship scale model, autonomous systems, data acquisition, para-

metric resonance, towing tank tests.

3.1 Introduction
Ship scale model testing has traditionally been the only way to accurately
determine ship resistance, propulsion, maneuvering and seakeeping charac-
teristics. These tests are usually carried out in complex facilities, where a large
carriage, to which the model is attached, moves it following a desired path
along a water tank.
Ship model testing could be broadly divided into four main types of
tests. Resistance tests, either in waves or still water, intended to obtain the
resistance of the ship without taking the effects of its propulsion system into
consideration. Propulsion tests aimed at analyzing the performance of the ship
propeller when it is in operation together with the ship itself. Maneuvering
tests the objective of which is to analyze the capability of the ship for carrying

Advances in Intelligent Robotics and Collaborative Automation, 49–72.


c 2015 River Publishers. All rights reserved.
50 An Autonomous Scale Ship Model for Parametric Rolling Towing Tank Testing

out a set of defined maneuvers. And finally, seakeeping tests aimed at studying
the behavior of the ship while sailing in waves [1].
The former two tests are carried out in towing tanks, which are slender
water channels where the model is attached to the carriage that tows it along
the center of the tank. The latter two, on the other hand, are usually performed
in the so-called ocean basins where the scale model can be either attached to
a carriage or radio controlled with no mechanical connection to it.
However, there exist certain kinds of phenomena in the field of seakeeping
that can be studied in towing tanks (which are cheaper and have more
availability than ocean basins) and that are characterized by showing very
large amplitude nonlinear motions. The interactions between the carriage and
the model due to these motions, which are usually limited to a maximum
in most towing tanks, and the lack of space under the carriage, reduce the
applicability of slender channels for these types of tests.
One of these phenomena is that of ship parametric roll resonance, also
known as parametric rolling. This is a well-known dynamical issue affecting
ships, especially containerships, fishing vessels and cruise ships, and it can
generate very large amplitude roll motions in a very sudden way, reaching
the largest amplitudes in just a few rolling cycles. Parametric roll is due to
the periodic alternation of wave crests and troughs along the ship, which
produce the changes in ship transverse stability that lead to the aforementioned
roll motions.
Resonance is most likely to happen when the ship sails in longitudinal seas
and when a certain set of conditions are present, which include a wave encoun-
ter frequency ranging twice the ship’s natural roll frequency, a wavelength
almost equal to the ship length and a wave amplitude larger than a given
threshold [2].
Traditionally, parametric roll tests in towing tanks have been carried out
by using a carriage towed model, where the model is free to move in just
some of the 6 degrees of freedom (typically heave, roll and pitch, the ones
most heavily influencing the phenomenon) [3, 4]. However, this arrangement
limits the possibility of analyzing the influence of the restrained degrees of
freedom [5], which may also be of interest while analyzing parametric roll
resonance, or it may interfere on its development.
The main objective of the present work is to overcome the described
difficulties for carrying out scale tests in slender towing tanks where large
amplitude motions are involved, while preserving the model capabilities for
being used in propulsion (free-running), maneuvering and seakeeping tests.
 3.2 System Architecture 51

This has been done by using a self-propelled and self-controlled scale

model, able to freely move in the six degrees of freedom and to measure, store
and process all the necessary data without direct need of the towing carriage.
In addition, the model could be used for any other tests, both in towing tanks or
ocean basins, with the advantage of being independent of the aforementioned
carriage.
The type and amount of data to be collected is defined after taking into
consideration the typology of tests to be carried out. Taking into account that
all three, propulsion, maneuvering and seakeeping tests should be considered,
all the data related with the motion of the ship, together with those monitoring
the propulsion system and the heading control system, have to be collected.
Moreover, a processing, storage and communication with shore system was
also implemented and installed onboard.
Finally, the results obtained by using the model in a test campaign aimed at
predicting the appearance of parametric roll, where a detection and prevention
algorithm has been implemented in the onboard control system, are presented.

3.2 System Architecture

The first implementation of the proposed system has been on a scale ship
model that has been machined from high-density (250 kg/m3 ) polyurethane
blocks to a scale of 1/15th, painted and internally covered by epoxy
reinforced fiberglass. Mobile lead weights have been installed on sup-
porting elements, allowing longitudinal, transverse and vertical displace-
ments for a correct mass arrangement. Moreover, two small weights have
been fitted into a transverse slider for fast and fine-tuning of both the
transverse position of the center of gravity and the longitudinal moment
of inertia.
The propulsion system consists of a 650W brushless, outrunner, three-
phase motor, an electronic speed control (ESC) and a two-stage planetary
gearbox, which move a four-bladed propeller. The ESC electronically gen-
erates a three-phase low voltage electric power source to the brushless
motor. With this ESC, the speed and direction of the brushless motor can
be controlled. It can even be forced to work as a dynamic brake if need be.
The rudder is linked to a radio control servo with a direct push-pull linkage,
so that the rudder deflection angle can be directly controlled with the servo
command.
Both the brushless motor and the rudder may be controlled either by
using a radio link or by the onboard model control system, which is the
52 An Autonomous Scale Ship Model for Parametric Rolling Towing Tank Testing

Table 3.1 Measured Parameters

Type of Test ITTC Guideline Measured Parameters
√
Propulsion 7.5-02-03-01.1 Model speed
√
Resistance/External tow force
√
Propeller thrust and torque
√
Propeller rate of revolution
√
Model trim and sinkage
√
Maneuverability 7.5-02-06-01 Time
√
(additional to Position
√
Propulsion) Heading
√
Yaw rate
√
Roll angle
√
Rudder angle
Force and moment on steering x
devices
√
Seakeeping 7.5-02-07-02.1 Model motions (6 d.o.f.)
√
(additional to Model motion rates (6 d.o.f.)
√
Maneuverability) Model accelerations (6 d.o.f.)
Impact pressures x
Water on deck x
Drift angle x

default behavior. The user can choose between both control methods any time;
there’s always a human at the external transmitter to ensure maximum safety
during tests.

3.2.1 Data Acquisition

In order to ensure that the model can be used for propulsion, maneuvering and
seakeeping tests and is in accordance with the ITTC (International Towing
Tank Conference) requirements, the following parameters have been measured
(Table 3.1):
In order to obtain data for all the representative parameters, the following
sensors have been installed onboard:
• Inertial Measurement Unit (IMU): having nine MEMS embedded sen-
sors; including 3 axis accelerometers, 3 axis gyros and 3 axis magne-
tometers. The IMU has an internal processor that provides information
on accelerations in the OX, OY and OZ axis, angular rates around these
three axes and quaternion based orientation vectors (roll, pitch, yaw),
both in RAW format and filtered by using Kalman techniques. This sensor
 3.2 System Architecture 53

has been placed near the ship’s center of gravity with the objective of
improving its performance;
• Thrust sensor: a thrust gauge has been installed to measure the thrust
generated by the propeller at the thrust bearing;
• Revolution and torque sensor: in order to measure the propeller revolu-
tions and the torque generated by the engine, a torque and rpm sensor
has been installed between both elements;
• Sonars: intended to measure the distance to the towing tank walls and
feed an automatic heading control system;
• Not directly devoted to tested magnitudes, there are also a temperature
sensor, battery voltage sensors and current sensors.
Data acquisition is achieved through an onboard mounted PC, placed forward
on the bottom of the model. The software in charge of the data acquisition
and processing and ship control is written in Microsoft .Net, and installed in
this PC. This software is described in the following section. There is a Wi-Fi
antenna at the ship’s bow, connected to the onboard PC that enables a Wi-Fi
link to an external, inshore workstation. This workstation is used to monitor
the ship operation.
An overview of the model where its main components are shown is
included in Figure 3.1.

3.2.2 Software Systems

In this section, the software designed and implemented to control the ship is
described. Regarding the operational part of the developed software, there are
three main tasks that have to be performed in real time in order to monitor

Figure 3.1 Ship scale model overview.

54 An Autonomous Scale Ship Model for Parametric Rolling Towing Tank Testing

and control the ship: acquisition, computation and actuation. In every time
step, once the system is working, all the sensor measurements are collected.
The indispensable sensors that need to be connected to the system are: the
Inertial Measurement Unit (IMU) which provides the acceleration, magnetic,
angular rate and attitude measurements, the sonars and thrust gauge, in this
case connected to a data acquisition board, and the revolution and torque
sensor (Figure 3.2).
Once the sensor data are acquired, the system computes the proper signals
to modify the rudder servo and the motor ESC commands. These signals can be
set manually (using the software interface from the Wi-Fi linked workstation,
or the external RC transmitter) or automatically. Applying a controller over
the rudder servo, using the information from the sonar signals, it is possible
to keep the model centered and in course along the towing tank. Another
controller algorithm is used to control the ship speed.
The software is based on VB. NET and to interact with the system from an
external workstation, a simple graphical user interface has been implemented
(Figure 3.3).
From an external workstation, the user can start and finish the test, activate
the sensors to measure, monitor the data sensors in real time, control the rudder
servo and the motor manually using a slider or stop the motor and finish the
test. All the acquired data measurements are saved in a file for a future analysis
of the test.

Figure 3.2 Connectivity between Mini-PC and sensors onboard.

 3.2 System Architecture 55

Figure 3.3 Graphical user interface to monitor/control the model from an external
workstation.

3.2.3 Speed Control

The speed control of the model is done by setting an rpm command, which
keeps engine revolutions constant by using a dedicated controller programmed
in the governing software. Alternatively, a servo command may be used for
setting a constant power input for the engine. In calm waters and for a given
configuration of the ship model, there is a relationship between ship speed
and propeller revolutions. By performing some preliminary tests at different
speeds, this relation can be adjusted and used thereafter for testing within a
simple, open loop controller. However, in case of testing with waves, these
waves introduce an additional and strong drag component on the ship forward
movement, and there is no practical way of establishing a similar sort of
56 An Autonomous Scale Ship Model for Parametric Rolling Towing Tank Testing

relationship. For these cases, the towing carriage is used as a reference and
the speed is maintained by keeping the ship model in a steady relative position
to the carriage.
The speed control strategy to cope with this composed speed was initially
tested as shown in Figure 3.4. It was initially done by means of a double PID
controller; the upper section of the controller tries to match the ship speed
with a set point selected by the user, cv . This portion of the controller uses the
derivative of the ship position along the tank, x, as an estimation of the ship
speed, evx . The position x is measured through the Laser Range Finder sensor
placed at the beginning of the towing tank, facing the ship poop, to capture the
ship position along the tank, and send this information through a dedicated RF
modem pair, to the Mini-PC onboard. The bottom section, on the other hand,
uses the integral of the ship acceleration in its local x-axis from the onboard
IMU, va , as an estimation of the ship speed, eva . Each branch has its own PID
controller, and the sum of both outputs is used to command the motor.
Both speed estimations come from different sensors, in different coor-
dinate systems, with different noise perturbations and, over all, they have
different natures. The estimation based on the derivative of the position along
the tank has little or zero drift over time, and its mean value matches the real
speed on the tank x axis, and changes slowly. On the other hand, the estimation
based on the acceleration along the ship’s local x-axis is computed by the
onboard IMU, from its MEMS sensors, and is prone to severe noises, drift over
time and changes quickly. Furthermore, the former estimation catches the slow

Figure 3.4 Double PID Speed control.

 3.2 System Architecture 57

behavior of the ship speed, and the latter its quick changes. This is the reason
to use different PID controllers with both estimations. The resulting controller
follows the user-selected speed setpoint, with the upper branch eliminating
any steady-state speed error, while minimizing quick speed changes with the
lower branch.
Later on, a different speed control approach was introduced in order to
improve its performance. Being the Laser Range Finder output an absolute
measure of ship position, the speed obtained from its derivative is significantly
more robust than the one estimated from the IMU in the first control scheme,
and has no drift over time or with temperature. The new speed control
algorithm is based in a complementary filter [6]. This filter estimates the ship
speed from two different speed estimations, with different and complementary
frequency components, as shown in Figure 3.5.
This two signal complementary filtering is based upon the use and
availability of two independent noisy measurements of the ship speed, v(s):
the one from the derivative of the range finder position estimation (v(s)+n1 (s))
and the one from the integration of IMU acceleration (v(s)+n2 (s)). Each of
these signals has their own spectral characteristics, here modeled by their
different noise levels, n1 (s) and n2 (s). If both signals have complementary
spectral characteristics, transfer functions may be chosen in such a way as to
minimize speed estimation error. The general requirement is that one of the
transfer functions complement the other. Thus, for both measurements of the
speed signal [7]:
H1 (s) + H2 (s) = 1.
This will allow the signal component to pass through the system undistorted
since the output of the system will always sum to one. In this case, n1 is
predominantly high-frequency noise and n1 is low-frequency noise; these two

Figure 3.5 Speed control. complementary filter.

58 An Autonomous Scale Ship Model for Parametric Rolling Towing Tank Testing

noise sources have complementary spectral characteristics. Choosing H 1 (s) to

be a low-pass filter, and H 2 (s) a high-pass filter, both with the same, suitably
cut frequency, the output v will not suffer from any delay in dynamic response
due to low-pass filtering, and will be free of both high- and low-frequency
noise.

3.2.4 Track-Keeping Control

Regarding heading control, IMU and sonar data are used for keeping the model
centered and in course along the towing tank. In case these values are not
accurate enough, heading control may be switched to a manual mode and an
external RC transmitter could be used for course keeping. At first, the signals
of the sonars to maintain the model centered on the tank and a Kalman filter
taken data from the IMU were used to keep the course, the magnetometers’
signals being of primary importance in this Kalman filter.
During testing, this arrangement showed to be not very effective because
the steel rails of the carriage, placed all along at both sides of the tank, induced
a shift in the magnetometer signals when the ship model was not perfectly
centered at the tank. In addition, the magnetometers were also sensible to
the electrical power lines coming across the tank. For these reasons only the
sonar signals were used to keep both course and position, with the help of
the relative position to the carriage, which have been used to keep the speed
constant anyway.

3.2.5 Other Components

The power for all the elements is provided by two 12 V D.C batteries,
placed abaft in the ship, providing room in their locations for longitudinal
and transverse mass adjustment. These batteries have enough capacity for a
whole day of operation.
The main propulsion system block, consisting of the already mentioned
brushless motor (1), electronic speed control (ESC, not in view), two-stage
planetary gearbox (2), rotational speed and torque sensor (3), elastic cou-
plings (4), and an output shaft with Kardan couplings (5), is represented in
Figure 3.6. The disposition of these elements in a single block allows a modular
implementation of the whole system and simplifies the operations needed to
install and uninstall it from a given ship model.
Finally, two adjustable weights have been placed in both sides of the ship
model, and another one has been placed forward in the centerline. In both
 3.3 Testing 59

Figure 3.6 Propulsion system block.

cases, enough room has been left as to allow transversal and vertical mass
adjustment. Moreover, two sliders with 0.5 kg weights have been installed for
fine tuning of the mass distribution.

3.3 Testing
As mentioned above, the proposed model is mainly intended to be used in
seakeeping tests in towing tanks, where the carriage may interfere in the motion
of the ship. The application of the proposed model to one of these tests, aimed
at predicting and preventing the appearance of parametric roll resonance, will
be presented in this section.
As it has been already described, parametric roll resonance could generate
large roll motions and lead to fatal consequences. The need for a detection
system, and even for a guidance system, has been recursively stated by the
maritime sector [8]. However, the main goal is to obtain systems that can, in
a first stage, detect the appearance of parametric roll resonance, but that, in a
second stage, can prevent it from developing.
As mentioned, there are some specific conditions that have to be present for
parametric roll to appear, regarding both ship and waves characteristics. Wave
encounter frequency should be around twice the ship natural roll frequency,
their amplitude should be over a certain threshold that depends on ship
characteristics, and their wavelength should be near the ship´s length. Ship roll
damping should be small enough not to dissipate all the energy that is generated
by the parametric excitation. And finally, ship restoring arm variations due to
wave passing along the hull, should be large enough as to counteract the effect
of roll damping.
60 An Autonomous Scale Ship Model for Parametric Rolling Towing Tank Testing

If parametric roll resonance wants to be prevented, it is obvious that there

are three main approaches that could be used, and that consist in avoiding the
presence of at least one of the aforementioned conditions.
The first alternative consists in acting on the ship damping components,
or using stabilizing forces that oppose the heeling arms generated during the
resonance process. The second alternative, aimed at reducing the amplitude of
restoring arm variations, necessarily implies introducing modifications in hull
forms. Finally, the third alternative is focused on avoiding wave encounter
frequency being twice the ship natural roll frequency. If we consider a ship
sailing at a given loading condition (and so with a constant natural roll
frequency) in a specific seaway (frequency and amplitude), the only way of
acting on the frequency ratio is by modifying encounter frequency. So, this
alternative would be based on ship speed and heading variations that take the
vessel out of the risk area defined by ωωne = 2. This last approach is the one
adopted in this work.

3.3.1 Prediction System

Regarding the prediction system, Artificial Neural Networks (ANN) have been
used as a roll motion forecaster, exploiting their capabilities as nonlinear
time series estimators, in order to subsequently detect any parametric roll
event through the analysis of the forecasted time series [9]. In particular,
multilayer perceptron neural networks (MPNNs) have been the structures
selected to predict ship roll motion in different situations. The base neural
network architecture used in this work is a multilayer perceptron network
with two hidden layers, 30 neurons per layer, and one output layer. This initial
structure is modified by increasing the number of neurons per layer, or the
number of layers, in order to compare the prediction performance of different
architectures and according to the complexity of the cases under analysis.
The network is fed with time series of roll motion, which are 20 seconds
long and sampled at a frequency Fs = 2 Hz; hence the input vector x0
has 40 components. The network has only one output, which is the one
step-ahead prediction. By substituting part of the input vector with the network
output values, and recursively executing the system, longer predictions can
be obtained from the initial 20 seconds. However, as the number of iterations
grows, the prediction performance deteriorates accordingly.
The length of the roll time series has been selected taking into account
two factors. On the one hand, the natural roll periods of the vessels chosen for
the testing. In the loading conditions considered for these tests, this value is
 3.3 Testing 61

7.48 seconds. On the other hand, parametric roll fully develops in a short
period of time, usually no more than four rolling cycles [10].
The training of the artificial neural network forecaster has been achieved
by using the roll time series obtained in the towing tank tests that will be
described in the following section.
These algorithms have been implemented within the ship onboard control
system, and their performance analyzed in some of the aforementioned towing
tank tests.

3.3.2 Prevention System

Once parametric resonance has been detected, the control system should act
modifying the model speed or heading in order to prevent the development
of the phenomenon. In order to do so, the concept of stability diagrams has
been applied. These diagrams display the areas in which, for a given loading
condition and as a function of wave height and encounter frequency – natural
roll frequency ratio and for different forward speeds, parametric roll takes
place. From the analysis of these regions, the risk state of the ship at every
moment could be determined and its speed and heading modified accordingly
to take the model out of the risk area.
In order to set up these stability diagrams, a mathematical model of the
ship behavior has been developed, validated, and implemented within the ship
control system in order to compute the stability areas for the different loading
conditions tested.
This model is a one degree of freedom nonlinear model of the ship roll
motion:
(Ixx + A44 ) · φ̈ + B44, T · φ̇ + C44 (φ, t) = 0,
where Ixx and A44 are respectively the mass and added mass moments of
inertia in roll, B44, T represents the nonlinear damping term and C44 (φ, t) is
the time varying nonlinear restoring coefficient. In this case, the moment and
the added moment of inertia have been obtained by measuring the natural roll
frequency from a zero speed roll decay test, carried out with the developed
ship scale model.
While trying to model parametric roll resonance, it has been seen that
both heave and pitch motions have a clear influence in the appearance of the
phenomenon, together with the effects of the wave moving along the hull.
So, including the influence of these factors is of paramount importance for a
good performance of the mathematical model. In this work, the influence of
these factors has been taken into account while computing the restoring term,
62 An Autonomous Scale Ship Model for Parametric Rolling Towing Tank Testing

by adopting the quasi-static “look up table” approach, described by [11] and

required by the ABS Guidelines [12] for modelling the variation of the ship
restoring capabilities in longitudinal waves.
Moreover, and regarding the roll damping, it has been shown that it is
essential for a good simulation of large amplitude roll motion, to consider this
parameter as highly nonlinear. In order to account for this fact, roll damping
has been decomposed into two terms, one linear component, which is supposed
to be dominant at small roll angles, and a quadratic one, which is necessary
to model the effects of damping at large roll angles. This approach has also
been applied by many authors for modelling parametric roll, that is, [13],
with accurate results. In order to determine the two damping coefficients
(B44, a , B44, b ) in the most accurate way, roll decay tests for different forward
speeds have been carried out in still water for the loading condition under
analysis. The procedure followed for determining the damping coefficients
from these tests is that described in [14].
Once the model was correctly set up and validated, it was executed for
different combinations of wave height and frequency, the maximum roll
amplitude for these time series was computed and the stability diagrams
developed.

3.3.3 Towing Tank Tests and Results

The proposed system has been used to perform different tests, some of which
have been published elsewhere [15]. The main objective of these tests was to
analyze, predict and prevent the phenomenon of parametric roll resonance. It
is in this sort of tests, characterized by large amplitude oscillations in both roll
and pitch motions, where the proposed system performs best as it can take
information on board without disturbing the free motion of the ship model.
To illustrate the influence of the towing device on the measures obtained
in this kind of tests, Figure 3.7 is presented. On it, the pitch and roll motions
of a conventional carriage-towed model (Figure 3.8) in a similar parametric
rolling test, are included.
As it can be observed, the ship pitch motion presents a series of peaks
(the most relevant in seconds 140, 180, 220 and 300), which are due to the
interference of the towing device. These interferences not only influence the
model pitch motion, but could also affect the development of parametric roll
and so, the reliability of the test.
The tests campaign has been carried out in the towing tank of the Escuela
Técnica Superior de Ingenieros Navales of the Technical University of Madrid.
 3.3 Testing 63

Figure 3.7 Roll and pitch motions in parametric roll resonance. Conventional carriage-towed
model.

Figure 3.8 Conventional carriage-towed model during testing.

64 An Autonomous Scale Ship Model for Parametric Rolling Towing Tank Testing

This tank is 100 meters long, 3.8 meters wide and 2.2 meters deep. It is
equipped with a screen type wave generator, directed by a wave generation
software, capable of generating longitudinal regular and irregular waves
according to a broad set of parameters and spectra. The basin is also equipped
with a towing carriage able to develop a speed of up to 4.5 m/s. As it has already
been mentioned, in this test campaign, trials at different forward speeds and
also at zero speed have been carried out.
Regarding the zero speed runs, in order to keep the model in position and
to try to avoid as much as possible any interferences of the restraining devices
in the ship motions, two fixing ropes with two springs have been tightened to
the sides of the basin and to a rotary element fixed to the model bow. Moreover,
another restraining rope has been fitted between the stern of the model and the
towing carriage, stowed just behind of it. However, this last rope has been kept
loose and partially immersed, being enough for keeping the model head to seas
without producing a major influence on its motion. In the forward speed test
runs, the model has been left sailing completely free, with the exception of a
security rope that would be used just in case the control of the ship was lost.
In order to set the adequate speed for each test run, a previous calibration
for different wave conditions has been carried out to establish the needed
engine output power (engine servo command) for reaching the desired speed
as a function of wave height and frequency. The exact speed value developed
in each of the test runs has been measured by following the model with the
towing carriage, which has provided the instantaneous speed along the run.
The calibration curve has been updated as the different runs were completed,
providing more information for subsequent tests. However and considering
that the model is free to move in the six degrees of freedom, the instantaneous
speed of the model may be affected by surge motion, especially at the
conditions with highest waves.
A total number of 105 test runs have been carried out in head regular waves.
Different combinations of wave height (ranging from 0.255 m to 1.245 m
model scale) and ratio between encounter frequency and natural roll frequency
(from 1.80 to 2.30) have been considered for three different values of forward
speed (Froude numbers 0.1, 0.15 and 0.2) and zero speed, and two different
values of metacentric height (0.370 m and 0.436 m). From the whole set of
test runs, 55 correspond to the 0.370 m GM case, while 50 correspond to a
GM of 0.436 m.
The results obtained from the different test runs have been applied for
determining ship roll damping coefficients and validating the performance of
the mathematical model described in the previous subsection, for validating
 3.3 Testing 65

Figure 3.9 Proposed model during testing.

the correctness of the so computed stability diagrams, and for training and
testing the ANN detection system.
In addition to this, the results of pitch and roll motion obtained with the
proposed model are presented in Figure 3.10, for the sake of comparison with
the results obtained with the conventional model (Figure 3.7).As it can be seen,
the pitch time series doesn’t present the peaks observed in the conventional

Figure 3.10 Roll and pitch motions in parametric roll resonance. Proposed model.
66 An Autonomous Scale Ship Model for Parametric Rolling Towing Tank Testing

model measurements, as no interference between model and carriage occurs

in this case.

3.3.3.1 Mathematical model validation

As a first step, the results obtained from the towing tank tests have been used for
validating the performance of the roll motion nonlinear mathematical model
to predict the amplitude of the resulting roll motion.
In most cases, the results obtained from the model are quite accurate, and
correctly simulate the roll motion of the ship both in resonant and non-resonant
conditions.
Examples for both situations could be seen in Figures 3.11 and 3.12:

Figure 3.11 Comparison between experimental and numerical data. Resonant case.

Figure 3.12 Comparison between experimental and numerical data. Non-Resonant case.
 3.3 Testing 67

3.3.3.2 Validation of stability diagrams

Once the model has been validated, it has been recursively executed for
a set of combinations of different wave frequencies and heights, covering
the frequency ratio range from 1.70 to 2.40 and wave heights from 0.20
to 1.20 m, at each operational condition under consideration (four speeds),
in order to set up the stability diagrams. Once computed, the results have
been compared to those obtained in the towing tank tests. To illustrate this
comparison, the results obtained from these experiments (Figure 3.13) have
been superimposed on the corresponding mathematical plots (Figure 3.14);
light dots represent non-resonant conditions, while resonance is shown by the
dark dots.
As can be seen in Figure 3.14, the shape of the unstable region matches
the results obtained in the towing tank tests, that together with the accurate
results obtained while computing roll amplitude, show a good performance
of the mathematical model.

Figure 3.13 Experimental stability diagrams. Fn = 0.1. (Left), Fn = 0.15 (Right),

GM = 0.370 m.

Figure 3.14 Comparison between experimental and numerical stability diagrams.

Fn = 0.1. (Left), Fn = 0.15 (Right), GM = 0.370 m.
68 An Autonomous Scale Ship Model for Parametric Rolling Towing Tank Testing

3.3.3.3 Prediction system tests

To forecast the onset and development of the parametric roll phenomena,
some standard perceptron ANNs have been used. Several ANN architectures
were tested and the overall best results were obtained with 3 layers of
30 neurons each.
The training cases for the ANNs have been obtained from the experiments
that have been carried out with different values of wave frequency and
amplitude at a Froude number of 0.2, consisting of 14 time series averaging
a full scale length of 420 seconds. IMU output data was sampled by the on-
board computer at a rate of 50 data sets per second. As for this particular ship
model, the roll period of oscillation is of several seconds, the data used for
training the ANNs was under-sampled to 2 Hz, which was more than enough
to capture the events and permit reducing the size of the data set used. This
resulted in a total number of 11169 training cases.
Encounter frequency – natural roll frequency ratio ranged from 1.8 to
2.3, implying that there would be cases where parametric roll was not
present. Due to this fact, the performance of the system in a condition where
only small roll amplitudes appear due to small transversal excitations or
when roll motions decrease (cases that were not present in the mathematical
model tests as no other excitation was present apart from head waves) could
be evaluated.
The tests have been performed on both regular and irregular waves in
cases ranging from small to heavy parametric roll. In regular waves, the RMS
error when predicting 10 seconds ahead has been of the order of 10−3 in cases
presenting large roll amplitudes and it reduces to 10−4 in cases with small
amplitudes.
Two examples of these predictions are shown in Figures 3.15 and 3.16,
which include one case with fully developed parametric roll, and another
one without any resonant motions. On these figures, the ANN forecast is
represented by the dotted line, while the real data is represented by full lines.
Figure 3.15 is a typical example of a case presenting large amplitudes, while
in Figure 3.16 no resonance motion takes place. As can be observed in both
cases, the results are very good.
Further details of the characteristics and performance of the forecasting
ANN system have been presented by the authors in [16]. There, the forecasting
system has been implemented on a ship model instrumented with accelerom-
eters and tested by using standard towing tank methods. The data used for the
Figure 3.7 plot has been obtained during this testing campaign.
 3.4 Conclusions and Future Work 69

Figure 3.15 Forecast results. 30 neuron, 3 layer MP. 10 seconds prediction. Resonant case.

Figure 3.16 Forecast results. 30 neuron, 3 layer mp. 10 Seconds prediction.

Non Resonant case.

3.4 Conclusions and Future Work

The development and implementation of an autonomous scale ship model
for towing tank testing has been presented as well as some of the results
obtained with it during a real towing tank test campaign. The system is aimed
to be installed on board of self-propelled models, acting as an autopilot that
controls speed and track, the latter by maintaining course and keeping the
model centered in the tank. It also has an IMU with a 3-axis accelerometer,
a gyroscope and a magnetometer and, in addition, it measures the torque,
rotational speed and propulsive force at the propeller. A model ship so
70 An Autonomous Scale Ship Model for Parametric Rolling Towing Tank Testing

instrumented is able to move without any restriction along any of its six degrees
of motion; consequently, the system produces optimal measurements even in
tests cases presenting motions of large amplitude.
At is present development stage, the system only needs to use the towing
carriage as a reference for speed and position. A more advanced version that
could eliminate the use of this carriage is under development. This towing
carriage, together with its rails, propulsion and instrumentation, is a very
costly piece of hardware. The final version of the system could be constructed
at a fraction of this cost and it will be a true towless towing tank, as it
will allow performing any standard towing tank test without the need of an
actual tow.

References
[1] T. I. Fossen, ‘Handbook of Marine Craft Hydrodynamics and Motion
Control’, John Wiley & Sons, 2011.
[2] W. N. France, M. Levadou, T. W. Treakle, J. R. Paulling, R. K. Michel
and C. Moore, ‘An investigation of head-sea parametric rolling and its
influence on container lashing systems’, Marine Technology, vol. 40(1),
pp. 1–19, 2003.
[3] A. Francescutto, ‘An experimental investigation of parametric rolling
in head waves’, Journal of Offshore Mechanics and Arctic Engineering,
vol. 123, pp. 65–69, 2001.
[4] I. Drummen, ‘Experimental and Numerical Investigation of Nonlinear
Wave-Induced Load Effects in Containerships considering Hydroelas-
ticity’, PhD Thesis, Norwegian University of Science and Technology,
Norway, 2008.
[5] International Towing Tank Conference (ITTC), ‘Testing and Extrapo-
lation Methods. Loads and Responses, Stability. Model Tests on Intact
Stability’, ITTC 7.5–02-07–04.1. 2008.
[6] R. Brown, P. Hwang, ‘Introduction to Random Signals and Applied
Kalman Filtering, Second Edition’, John Wiley and Sons Inc., 1992.
[7] E. R. Bachmann, ‘Inertial and Magnetic Tracking of Limb Segment
Orientation for Inserting Humans into Synthetic Environments’, PhD
Thesis, Naval Posgraduate School, USA, 2000.
[8] K. Døhlie, ‘Parametric Roll - a problem solved?’, DNV Container Ship
Update, 1, 2006.
[9] R. M. Golden, ‘Mathematical methods for neural network analysis and
design’, The MIT Press, 1996.
 References 71

[10] International Maritime Organization (IMO), ‘Revised Guidance to the

Master for Avoiding Dangerous Situations in Adverse Weather and
Sea Conditions (Vol. IMO MSC.1/Circ. 1228)’, IMO Maritime Safety
Committee, 82nd session, 2007.
[11] G. Bulian, ‘Nonlinear parametric rolling in regular waves - a gen-
eral procedure for the analytical approximation of the GZ curve and
its use in time domain simulations’, Ocean Engineering, 32 (3–4),
pp. 309–330, 2005.
[12] American Bureau of Shipping (ABS), ‘Guide for the Assessment of
Parametric Roll Resonance in the Design of Container Carriers’, 2004.
[13] M. A. S. Neves, C. A. Rodríguez, ‘On unstable ship motions resulting
from strong non-linear coupling’, Ocean Engineering, 33 (14, 15),
1853–1883, 2006.
[14] Y. Himeno, ‘Prediction of Ship Roll Damping.AState of theArt’, Depart-
ment of Naval Architecture and Marine Engineering, The University of
Michigan College of Engineering, USA, 1981.
[15] M. Míguez González, F. López Peña, V. Díaz Casás, L. Pérez Rojas,
‘Experimental Parametric Roll Resonance Characterization of a Stern
Trawler in Head Seas’, Proceedings of the 11th International Conference
on the Stability of Ships and Ocean Vehicles, Athens, 2012.
[16] F. López Peña, M. Míguez González, V. Díaz Casás, R. J. Duro, D. Pena
Agras, ‘An ANN Based System for Forecasting Ship Roll Motion’, Pro-
ceedings of the 2013 IEEE International Conference on Computational
Intelligence and Virtual Environments for Measurement Systems and
Applications, Milano, Italy, 2013.
 4
Autonomous Knowledge Discovery Based
on Artificial Curiosity-Driven Learning
by Interaction

K. Madani, D. M. Ramik and C. Sabourin

Images, Signals & Intelligent Systems Lab. (LISSI / EA 3956) University

PARIS-EST Créteil (UPEC) –Sénart-FB Institute of Technolog, Lieusaint,
France
Corresponding author: K. Madani <[email protected]>

Abstract
In this work, we investigate the development of a real-time intelligent system
allowing a robot to discover its surrounding world and to learn autonomously
new knowledge about it by semantically interacting with humans. The learning
is performed by observation and by interaction with a human. We describe the
system in a general manner, and then we apply it to autonomous learning of
objects and their colors. We provide experimental results both using simulated
environments and implementing the approach on a humanoid robot in a real-
world environment including every-day objects. We show that our approach
allows a humanoid robot to learn without negative input and from a small
number of samples.

Keywords: Visual saliency, autonomous learning, intelligent system, artifi-

cial curiosity, automated interpretation, semantic robot-human interaction.

4.1 Introduction
In recent years, there has been a substantial progress in robotic systems able
to robustly recognize objects in the real world using a large database of
pre-collected knowledge (see [1] for a notable example). There has been,
however, comparatively less advance in the autonomous acquisition of such

Advances in Intelligent Robotics and Collaborative Automation, 73–94.


c 2015 River Publishers. All rights reserved.
74 Autonomous Knowledge Discovery Based on Artificial Curiosity-Driven

knowledge: if contemporary robots are often fully automatic, they are rarely
fully autonomous in their knowledge acquisition. If the aforementioned
substantial progress is commonsensical regarding the last-decades’ significant
developments in methodological and algorithmic approaches relating visual
information processing, pattern recognition and artificial intelligence, the
languishing in the machine’s autonomous knowledge acquisition is also
obvious regarding the complexity of the additional necessary skills to achieve
such “not algorithmic” but “cognitive” task.
Emergence of cognitive phenomena in machines have been and remain
an active part of research efforts since the rise of Artificial Intelligence (AI)
in the middle of the last century, but the fact that human-like machine-cognition
is still beyond the reach of contemporary science only proves how difficult
the problem is. In fact, nowadays there are many systems, such as sensors,
computers or robotic bodies, that outperform human capacities; nonetheless,
none of the existing robots can be called truly intelligent. In other words, robots
sharing everyday life with humans are still far away. Somewhat, it is due to
the fact that we are still far from fully understanding the human cognitive
system. Partly, it is so because it is not easy to emulate human cognitive skills
and complex mechanisms relating those skills. Nevertheless, the concepts of
bio-inspired or human-like machine-cognition remain the foremost sources of
inspiration for achieving intelligent systems (intelligent machines, intelligent
robots, etc. . . ). This is the way we have taken (e.g. through inspiration
from biological and human knowledge acquisition mechanisms) to design the
investigated human-like machine-cognition based system able to acquire high-
level semantic knowledge from visual information (e.g. from observation). It
is important to emphasize that the term “cognitive system” means here that
characteristics of such a system tend to those of human cognitive systems. This
means that a cognitive system, which is supposed to be able to comprehend
the surrounding world on its own, but whose comprehension would be
non-human, would afterward be incompetent of communicating about it with
its human counterparts. In fact, human-inspired knowledge representation and
human-like communication (namely semantic) about the acquired knowledge
become key points expected from such a system. To achieve the aforemen-
tioned capabilities, such a cognitive system should thus be able to develop
its own high-level representation of facts from low-level visual information
(such as images). Accordingly to the expected autonomy, the processing from
the “sensory level” (namely visual level) to the “semantic level” should be
performed solely by the robot, without human supervision. However, this does
not mean excluding interaction with humans, which is, on the contrary, vital for
 4.1 Introduction 75

any cognitive system, be it human or machine. Thus, the investigated system

has to share its perceptual high-level knowledge of the world with the human
by interacting with him. The human on his turn shares with the cognitive robot
his knowledge about the world using natural speech (utterances) completing
observations made by the robot.
In fact, if a humanoid robot is required to learn to share the living space with
its human counterparts and to reason about it in “human terms”, it has to face at
least two important challenges. One, coming from the world itself, is the vast
number of objects and situations the robot may encounter in the real world.
The other one comes from humans’ richness concerning various ways they
use to address those objects or situations using natural language. Moreover, the
way we perceive the world and speak about it is strongly culturally dependent.
This is shown in [2] regarding usage of color terms by different people around
the world, or in [3] regarding cultural differences in description of spatial
relations. A robot supposed to defeat those challenges cannot rely solely on a
priori knowledge that has been given to it by a human expert. On the contrary,
it should be able to learn on-line, within the environment in which it evolves
and by interaction with the people it encounters in that environment (see
[4] for a survey on human-robot interaction and learning and [5] for an
overview of the problem of anchoring). This learning should be completely
autonomous, but still able to benefit from interaction with humans in order
to acquire their way of describing the world. This will inherently require that
the robot has the ability of learning without an explicit negative evidence
or “negative training set” and from a relatively small number of samples.
This important capacity is observed in children learning the language [6].
This problem has been addressed to different degrees in various works. For
example, in [7] a computational model of word-meaning, acquisition by
interaction is presented. In [8], the authors present a computational model for
the acquisition of a lexicon describing simple objects. In [9], a humanoid robot
is taught to associate simple shapes to human lexicon. In [10], a humanoid
robot is taught through a dialog with untrained users with the aim to learn
different objects and to grasp them properly. More advanced works on robots’
autonomous learning and dialog are given by [11, 12].
In this chapter, we describe an intelligent system, allowing robots (as for
example humanoid robots) to learn and to interpret the world in which they
evolve using appropriate terms from human language, while not making use
of a priori knowledge. This is done by word-meaning anchoring based on
learning by observation and by interaction with its human counterpart. Our
model is closely inspired by human infants’early-ages learning behaviour (e.g.
76 Autonomous Knowledge Discovery Based on Artificial Curiosity-Driven

see [13, 14]). The goal of this system is to allow a humanoid robot to anchor
the heard terms to its sensory-motor experience and to flexibly shape this
anchoring according to its growing knowledge about the world. The described
system can play a key role in linking existing object extraction and learning
techniques (e.g. SIFT matching or salient object extraction techniques) on one
side, and ontologies on the other side. The former ones are closely related to
perceptual reality, but are unaware of the meaning of objects they are treated,
while the latter ones are able to represent complex semantic knowledge about
the world, but, they are unaware of the perceptual reality of concepts, which
they are handling.
The rest of this chapter is structured as follows. Section 4.2 describes
the architecture of the proposed approach. In this section, we detail our
approach by outlining its architecture and principles, we explain how beliefs
about the world are generated and evaluated by the robot and we describe
the role of human-robot interaction in the learning process. Validation of
the presented system on colors learning and interpretation, using simulation
facilities, is reported in Section 4.3. Section 4.4 focuses on the implementation
and validation of the proposed approach on a real robot in a real-world
environment. Finally, Section 4.5 discusses the achieved results and outlines
future work.

4.2 Proposed System and Role of Curiosity

Curiosity is a key skill for human cognition and thus it appears as an appealing
concept in conceiving artificial systems that gather knowledge, especially
when they are supposed to gather knowledge autonomously. Accordingly to
Berlyne’s Theory of human curiosity [15], two kinds of curiosities stimulate
the human’s cognitive mechanism. The first one is the so-called “perceptual
curiosity”, which leads to increased perception of stimuli. It is a lower-level
function, more related to perception of new, surprising or unusual sensory
input. It relates reflexive or repetitive perceptual experiences. The other
one is called “epistemic curiosity”, which is more related to the “desire
for knowledge that motivates individuals to learn new ideas, to eliminate
information-gaps, and to solve intellectual problems.
According to [16] and [17], the general concept of the presented archi-
tecture could include one unconscious visual level which may contain a
number of Unconscious Cognitive Functions (UCF) and one conscious visual
level which may contain a number of Conscious Cognitive Functions (CCF).
Conformably with the aforementioned concept of two kinds of curiosity,
 4.2 Proposed System and Role of Curiosity 77

an example of knowledge extraction from visual perception, involving both

kinds of curiosity, is shown on Figure 4.1. The perceptual curiosity moti-
vates or stimulates what we call the low-level knowledge acquisition and
concerns “reflexive” (unconscious) processing level. It seeks “surprising”
or “attention-drawing” information in given visual data. The task of the
perceptual curiosity is realized by perceptual saliency detection mechanisms.
This gives the basis for operation of high-level knowledge acquisition,
which is stimulated by epistemic curiosity. Being previously defined as
the process that motivates to “learn new ideas, eliminate information-
gaps, and solve intellectual problems”: as those relating the interpretation
of visual information or the belief’s generation concerning the observed
objects.
The problem of learning brings an inherent problem of distinguishing
the pertinent sensory information and the impertinent one. The solution to
this task is not obvious even if we achieve joint attention in the robot.
This is illustrated on Figure 4.2. If a human points to one object (e.g. an
apple) among many others, and describes it as “red”, the robot still has to
distinguish which of the detected colors and shades of the object the human
is referring to.
To achieve correct anchoring in spite of such an uncertainty, we adopt the
following strategy. The robot extracts features from important objects found
in the scene along with the words the tutor used to describe the objects. Then,
the robot generates its beliefs about which word could describe which feature.
The beliefs are used as organisms in a genetic algorithm. Here, the appropriate

Figure 4.1 General Bloc-diagram of the proposed curiosity driven architecture (left) and
principle of curiosity-based stimulation-satisfaction mechanism for knowledge acquisition
(right).
78 Autonomous Knowledge Discovery Based on Artificial Curiosity-Driven

Figure 4.2 A Human would describe this Apple as “Red” in spite of the fact, that this is not
the only visible color.

fitness function is of major importance. To calculate the fitness, we train a

classifier based on each belief and using it, we try to interpret the objects the
robot has already seen. We compare the utterances pronounced by the human
tutor in the presence of each such an object with the utterances the robot would
use to describe it based on the current belief. The closer the robot’s description
is to the one given by the human, the higher the fitness is. Once the evolution
has been finished, the belief with the highest fitness is adopted by the robot
and is used to interpret occurrences of new (unseen) objects. On Figure 4.3,
important parts of the system proposed in this paper are depicted.

4.2.1 Interpretation from Observation

Let us suppose a robot equipped with a sensor observing the sur-
rounding world. The world is represented as a set of features I =
{i1 , i2 , · · · , ik } ,which can be acquired by this sensor [18]. Each time the
robot makes an observation o, a human tutor gives it a set of utterances Um
describing the important (e.g. salient) objects found. Let us denote the set of
all utterances ever given about the world as U. The observation o is defined as
an ordered pair o = {Il , Um } , where Il ⊆ I, expressed by Equation (4.1),
stands for the set of features obtained from observation and Um ⊆ U is a set of
utterances (describing o) given in the context of that observation. ip denotes
the pertinent information for a given u (i.e. features that can be described
 4.2 Proposed System and Role of Curiosity 79

Figure 4.3 A Human would describe this Toy-frog as green in spite of the fact, that this is
not the only visible color.

semantically as u in the language used for communication between the human

and the robot), ii the impertinent information ii (i.e. features that are not
described by the given u, but might be described by another ui ∈ U ) and
sensor noise ε.The goal for the robot is to distinguish the pertinent information
present in the observation from the impertinent one and to correctly map the
utterances to appropriate perceived stimuli (features). In other words, the robot
is required to establish a word-meaning relationship between the uttered words
and its own perception of the world. The robot is further allowed to interact
with the human in order to clarify or verify its interpretations.
 
Il = ip (u) + ii (u) + ε. (4.1)
Um Um

Let us define an interpretation X (u) = {u , Ij }of an utterance u an ordered

pair where Ij ⊆ I is a set of features from I. So, the belief B is defined
according to Equation (4.2) as an ordered set of X(u) interpreting utterances
u from U.
80 Autonomous Knowledge Discovery Based on Artificial Curiosity-Driven

B = {X (u1 ), · · · , X (un ) } . (4.2)

According to the criterion expressed by (4.3), one can calculate the belief
B, which interprets in the most coherent way the observations made so far:
in other words, by looking for such a belief, which minimizes across all the
observations oq ∈ O the difference between the utterances UHq made by the
human, and those utterances UBq , made by the system by using the belief B.
Thus, B is a mapping from the set U to I : all members of U map to one or
more members of I and no two members of U map to the same member of I.
⎛ ⎞
|O|

arg min ⎝ |UHq − UBq |⎠ . (4.3)
B
q=1

Figure 4.4 gives, through example, an alternative scheme of the defined notions
and their relationship. It depicts a scenario in which two observations o1 and
o2 are made corresponding to two description U1 and U2 of those observations,
respectively.
On first observation, features i1 and i2 were obtained along with utterances
u1 and u2 , respectively. Likewise for the second observation, features i3 , i4
and i5 were obtained along with utterance u3 . In this example, it is easily
visible that the entire set of features I = {i1 , · · · , i5 } contains two sub-sets
I1 and I2 . Similarly the ensemble of whole utterances {u1 , u2 , u3 } give the

Figure 4.4 Bloc-diagram of relations between observations, features, beliefs and utterances
in sense of terms defined in the text.
 4.2 Proposed System and Role of Curiosity 81

set UH and their sub-sets U1 and U2 refer to the corresponding observations

(e.g. q ∈ {1, 2}). In this view, an interpretation X (u1 ) is a relation of u1 with
a set of features from I (namely I1 ). Then, a belief B is a mapping (relation)
from the set U to I. All members of U map to one or more members of I and
no two members of U are associated to the same member of I.

4.2.2 Search for the Most Coherent Interpretation

The system has to look for a belief B, which would make the robot describe a
particular scene with utterances as close and as coherent as possible to those
made by a human on the same scene. For this purpose, instead of performing
the exhaustive search over all possible beliefs, we propose to search for a
suboptimal belief by means of a genetic algorithm. For doing that, we assume
that each organism within it has its genome constituted by a belief, which,
results into genomes of equal size |U | containing interpretations X(u) of all
utterances from U. The task of coherent belief generation is to generate beliefs
which are coherent with the observed reality.
In our genetic algorithm, the genomes’ generation is a belief generation
process generating genomes (e.g. beliefs) as follows. For each interpretation
X(u) the process explores the whole the set O. For each observation oq ∈ O,
if u ∈ UHq then features iq ∈ Ij (with Ij ⊆ I) are extracted. As
described in (1), the extracted set contains pertinent as well as impertinent
features. The coherent belief generation is done by deciding, which features
iq ∈ Ij may possibly be the pertinent ones. The decision is driven by
two principles. The first one is the principle of “proximity”, stating that
any feature i is more likely to be selected as pertinent in the context of,
if its distance to other already selected features is comparatively small.
The second principle is the “coherence” with all the observations in O. This
means that any observation oq ∈ O, corresponding to u ∈ UHq , has to have at
least one feature i assigned into Ij of the current X (u) = {u, Ij } [19]. Thus,
it is both the similarity of features and the combination of certain utterances,
describing observations from O (characterized by certain features), that guide
the belief generation process. These beliefs may be seen as “informed guesses”
on the interpretation of the world as perceived by the robot.
To evaluate a given organism, a classifier is trained, whose classes
are the utterances from U and the training data for each class u ∈
U are those corresponding to X (u) = {u, Ij } , i.e. the features
associated with the given u in the genome. This classifier is used through
the whole set O of observations, classifying utterances u ∈ U describing
82 Autonomous Knowledge Discovery Based on Artificial Curiosity-Driven

each oq ∈ Oaccording to its extracted features. Such a classification results

in the set of utterances UBq (meaning that a belief B is tested regarding
the qth observation). The fitness function evaluating the fitness of each
above-mentioned organism is defined as “disparity” between UBq and UHq
(defined in the previous subsection) which is computed according to the
Equation (4.4), where v is given by Equation (4.5) representing the number
of utterances that are not present in both sets UBq and UHq , which means
that they are either missed or are superfluous utterances interpreting the given
features.
1
D (ν) = (4.4)
1+ν
    
   
ν =  UHq UBq  −  UHq UBq  . (4.5)
At the end of the above-described genetic evolution process, the globally best
fitting organism is chosen as the belief that best explains the observations O
made (by the robot) so far about the surrounding world.

4.2.3 Human-Robot Interaction

Human beings learn both by observation and by interaction with the world
and with other human beings. The former is captured in our system in
the “best interpretation search” outlined in previous subsections. The latter
type of learning requires that the robot be able to communicate with its
environment and is facilitated by learning by observation, which may serve
as its bootstrap. In our approach, the learning by interaction is carried out in
two kinds of interactions: human-to-robot and robot-to-human. The human-to-
robot interaction is activated anytime the robot interprets the world wrongly.
When the human receives a wrong response (from the robot), he provides the
robot a new observation by uttering the desired interpretation. The robot takes
this new corrective knowledge about the world into account and searches for
a new interpretation according to this new observation. The robot-to-human
interaction may be activated when the robot attempts to interpret a particular
feature. If the classifier trained with the current belief classifies the given
feature with a very low confidence, then this may be a sign that this feature is
a borderline example. In this case, it may be beneficial to clarify its true nature.
Thus, led by epistemic curiosity, the robot asks its human counterpart to make
an utterance about the uncertain observation. If the robot does not interpret
according to the utterance given by the human (the robot’s interpretation was
 4.3 Validation Results by Simulation 83

wrong), this observation is recorded as new knowledge and a search for the
new interpretation is started.
Using these two ways of interactive learning, the robot’s interpretation of
the world evolves both in amount, covering increasingly more phenomena
as they are encountered, and in quality, shaping the meaning of words
(utterances) to conform with the perceived world.

4.3 Validation Results by Simulation

In the simulated environment, images of real-world objects were presented
to the system alongside with textual tags describing colors present on each
object. The images were taken from the Columbia Object Image Library
database (COIL: it contains 1000 color images of different views of 100
objects). Five fluent English speakers were asked to describe each object
in terms of colors. We restricted the choice of colors to “Black”, “Gray”,
“White”, “Red”, “Green”, “Blue” and “Yellow”, based on the color opponent
process theory [20]. The tagging of the entire set of images was highly coherent
across the subjects. In each run of the experiment, we have randomly chosen
a tagged set. The utterances were given in the form of text extracted from the
descriptions. The object was accepted as correctly interpreted if the system’s
and the human’s interpretations were equal.

Figure 4.5 Upper: the WCS color table. lower: the WCS color table interpreted by robot
taught to distinguish warm (marked by red), cool (blue) and neutral (white) colors.
84 Autonomous Knowledge Discovery Based on Artificial Curiosity-Driven

The rate of correctly described objects from the test set was approximately
91% after the robot had fully learned. Figure 4.5 gives the result of interpre-
tation by the system of the colors of the WCS table regarding “Warm” and
“Cool” colors.
Figure 4.6 shows the learning rate versus the increasing number of
exposures of each color. It is pertinent to emphasize the weak number of
learned examples (required examples) leading to a correct recognition rate

Figure 4.6 Evolution of number of correctly described objects with increasing number of
exposures of each color to the simulated robot.

Figure 4.7 Examples of obtained visual colors’ interpretations (lower images) and corre-
sponding original images (upper images) for several testing objects from COIL database.
 4.4 Implementation on Real Robot and Validation Results 85

of 91%. Finally, Figure 4.7 gives an example of objects’ colors interpretation

by the system.

4.4 Implementation on Real Robot and Validation Results

The validation of the proposed system has been performed on the basis
of both simulation of the designed system and by an implementation on
a real humanoid robot1 . As real robot we have considered the NAO robot
(a small humanoid robot from Aldebaran Robotics) which provides a number
of facilities such as onboard camera (vision), communication devices and
onboard speech generator. The fact that the above-mentioned facilities were
already available offers a huge save of time, even if those faculties remain
quite basic in that kind of robot.
Although the usage of the presented system is not specifically bound to
humanoid robots, it is pertinent to state two main reasons why a humanoid
robot has been used for the system’s validation. The first reason for this is
that from the definition of the term “humanoid”, a humanoid robot aspires
to make its perception close to the human one, entailing a more human-like
experience of the world. This is an important aspect to be considered in the
context of sharing knowledge between a human and a robot. Some aspects
of this problem are discussed in [21]. The second reason is that humanoid
robots are specifically designed in order to interact with humans in a “natural”
way by using a loudspeaker and microphone set. Thus, required facilities
for bi-directional communication with humans through speech synthesis and
speech recognition are already available on such kinds of robots. This is of
major importance when speaking is a central item for natural human-robot
interaction.

4.4.1 Implementation
The core of the implementation’s architecture is split into five main units:
Communication Unit (CU), Navigation Unit (NU), Low-level Knowledge
Acquisition Unit (LKAU), High-level Knowledge Acquisition Unit (HLAU)
and Behavior Control Unit (BCU). Figure 4.8 illustrates the bloc-diagram
of the implementation’s architecture. The aforementioned units control NAO
robot (symbolized by its sensors, its actuators and its interfaces in Figure 4.8)

1
A video capturing different parts of the experiment may be found online on:
http://youtu.be/W5FD6zXihOo
86 Autonomous Knowledge Discovery Based on Artificial Curiosity-Driven

Figure 4.8 Block diagram the implementation’s architecture.

through its already available hardware and software facilities. In other words,
the above-mentioned architecture controls the whole robot’s behavior.
The purpose of NU is to allow the robot to position itself in space with
respect to objects around it and to use this knowledge to navigate within
the surrounding environment. Capacities needed in this context are obstacle
avoidance and determination of distance to objects. Its sub-unit handling
spatial orientation receives its inputs from the camera and from the LKAU.
To get to the bottom of the obstacle avoidance problem, we have adopted a
technique based on ground color modeling. Inspired by the work presented
in [22], color model of the ground helps the robot to distinguish free-space
from obstacles. The assumption is made that obstacles repose on ground
(i.e. overhanging and floating objects are not taken into account). With this
assumption, the distance of obstacles can be inferred from monocular camera
data. In [23], some aspects of distance estimation from a static monocular
camera have been mentioned, proffering the robot the capacity to infer
distances and sizes of surrounding objects.
The LKAU ensures gathering of visual knowledge, such as detection
of salient objects and their learning (by the sub-unit in charge of salient
object detection) and sub-recognition (see [18, 24]). Those activities are
 4.4 Implementation on Real Robot and Validation Results 87

carried out mostly in an “unconscious” manner, that is, they are run as
an automatism in “background” while collecting salient objects and learn-
ing them. The learned knowledge is stored in Long-term Memory for
further use.
The HKAU is the center where the intellectual behavior of the robot is
constructed. Receiving its features from the LKAU (visual features) and from
the CU (linguistic features), this unit processes the belief generation, the most
coherent beliefs emergence and constructs the high-level semantic represen-
tation of acquired visual knowledge. Unlike the LKAU, this unit represents
conscious and intentional cognitive activity. In some way, it operates as a
baby who learns from observation and from verbal interaction with adults
about what he observes developing in this way his own representation and his
own opinion about the observed world [25].
The CU is in charge of robot communications. It includes an output
communication channel and an input communication channel. The output
channel is composed of a Text-To-Speech engine which generates human
voice through loudspeakers. It receives the text from the BCU. The input
channel takes its input from a microphone and through an Automated Speech
Recognition engine (available in NAO) the syntax and semantic analysis
(designed and incorporated in BCU) it provides the BCU labeled chain of
strings representing the heard speech. As it has been mentioned, the syntax
analysis is not available on NAO. Thus it has been incorporated in BCU. To
perform syntax analysis, the TreeTagger tool is used. Developed by the ICL at
University of Stuttgart, the TreeTagger tool is a tool for annotating text with
part-of-speech and lemma information. Figure 4.9 shows, through a simple
example of an English phrase, the operational principle of syntactic analysis
performed by this tool. “Part-of-speech” row gives tokens explanation and the
“Lemma” row shows lemmas output, which is the neutral form of each word in
the phrase. This information along with known grammatical rules for creation
of English phrases may further serve to determine the nature of the phrase as

Figure 4.9 Example of English phrase and the corresponding syntactic analysis output
generated by treetagger.
88 Autonomous Knowledge Discovery Based on Artificial Curiosity-Driven

declarative (for example: “This is a Box”), interrogative (for example: “What

is the name of this object?”) or imperative (for example: “Go to the office”).
It can be also used to extract the subject, the verb and other parts of speech,
which are further processed in order to make emerge the appropriate action
by the robot. Figure 4.10 gives the flow diagram of communication between
the robot and a human as it has been implemented in this work.
The BCU plays the role of a coordinator of robot’s behavior. It handles data
flows and issues command signals for other units, controlling the behavior of
the robot and its suitable reactions to external events (including its interaction
with humans). BCU received its inputs from all other units and returns
its outputs to each concerned unit including robot’s devices (e.g. sensors,
actuators and interfaces) [25].

4.4.2 Validation Results

A total of 25 every-day objects was collected for experimental purposes of
(Figure 4.11). They have been randomly divided into two sets for training and
for testing. The learning set objects were placed around the robot and then a
human tutor pointed to each of them calling it by its name. Using its 640x480
monocular color camera, the robot discovered and learned the objects from its
surrounding environment containing objects from the above-mentioned set.

Figure 4.10 Flow diagram of communication between a robot and a human which is used in
this work.
 4.4 Implementation on Real Robot and Validation Results 89

Figure 4.11 Everyday objects used in the experiments in this work.

The first validation involving the robot has aimed at verifying the leaning,
color interpretation, interaction with human and description abilities of the
proposed (e.g. investigated) system. To do this, the robot has been asked to
learn a subset of the 25 objects: in terms of associating the name of each
detected object to that object. At the same time, a second learning process has
been performed involving the interaction with the tutor who has successively
pointed the above-learned objects describing (e.g. telling) to the robot the
color of each object. Here below, an example of the Human-Robot interactive
learning is reported:
• Human: [pointing a red aid-kit] “This is a first-aid-kit!”
• Robot: “I will remember that this is a first-aid-kit.”
• Human: “It is red and white.”
• Robot: “OK, the first-aid-kit is red and the white.”
After learning the names and colors of the observed objects, the robot is
asked to describe a number of objects including also some of the already
learned objects but in a different posture (for example the yellow chocolate
90 Autonomous Knowledge Discovery Based on Artificial Curiosity-Driven

box presented in reverse posture, etc.) and a number of still unseen objects
(as for example a red apple or a white teddy-bear). The robot has successfully
described, in coherent linguistics, the presented seen and unseen objects. Here
below is an example of Human-Robot interaction during the recognition phase:
• Human: [pointing the unseen white teddy-bear]: “Describe this!”
• Robot: “It is white!”
• Human: [pointing the already seen, but reversed, yellow chocolate box]:
“Describe this!”
• Robot: “It is yellow!”
• Human: [pointing the unseen apple]: “Describe this!”
• Robot: “It is red!”
Figure 4.12 shows two photographs of the above-reported experimental vali-
dation, where the robot completes its knowledge by interacting with a human
and learning from him. Figure 4.13 shows another two photographs where
the robot describes different objects it could detect within its surrounding
environment.
In another experiment, taking into account the previously learned objects,
among which the robot has detected and has learned a black book (the only
available book in the learning set), the robot has been asked to search for the
book in a new (e.g. unseen) environment where a number of different objects,
including the above-mentioned black book, have been placed with two new
books (unseen) as well as other objects looking similar to a book. The robot’s
behavior and Human-Robot interaction during this experiment is given here
below:
• Human: “Go find the book!”
• Robot: “OK, I am searching for the book!”

Figure 4.12 Tutor pointing an aid-kit detected by robot describing its name and color to the
robot (left-side picture). Pointing, in the same way, another visible objects detected by robot,
tutor describes them to the robot (Right-Side picture).
 4.4 Implementation on Real Robot and Validation Results 91

Figure 4.13 Tutor pointing a yellow chocolate box which has been seen, interpreted and
learned (by the robot) in terms of colors then asking the robot to describe the chosen object
(left-side picture). Tutor pointing an unseen white teddy-bear asking the robot to describe the
chosen object (right-side picture).

• Robot: [after visually exploring the new surrounding environment]: “I

can see three books!”
• Human: “Get the red one!”
• Robot: [after turning toward the “red and grey” book and after walking
and reaching the aforementioned book’s location] “I am near the red
book!”
Figure 4.14 shows two photographs of the robot during its searching for the
required “red” book. Besides the correct recognition of the desired object
(asked by the tutor), what is pertinent and interesting to note is that the robot
also found two other unseen books. What is also very pertinent to emphasize,
and very interesting, is that even if there is no “red” book in that environment,
the robot has correctly interpreted the fact that the red book required by the
human was the “red and grey” book: the only book that may coherently

Figure 4.14 Images from a video sequence showing the robot searching for the book
(left-side picture) and robot’s camera view and visualization of color interpretation of the
searched object (right-side picture).
92 Autonomous Knowledge Discovery Based on Artificial Curiosity-Driven

be considered as “red” by the human. A video showing the experimental

validation may be found on http://youtu.be/W5FD6zXihOo. More details of
the presented work with complementary results can be found in [19, 25].

4.5 Conclusions
This chapter has presented, discussed and validated a cognitive system for
high-level knowledge acquisition from visual perception based on the notion
of artificial curiosity. Driving as well the lower as the higher levels of the
presented cognitive system, the emergent artificial curiosity allows such a
system to learn in an autonomous manner new knowledge about the unknown
surrounding world and to complete (enrich or correct) its knowledge by inter-
acting with a human. Experimental results, performed as well on a simulation
platform as using the NAO robot, show the pertinence of the investigated
concepts as well as the effectiveness of the designed system. Although
it is difficult to make a precise comparison due to different experimental
protocols, the results we obtained show that our system is able to learn
faster and from significantly fewer examples than most of more-or-less similar
implementations.
Based on the results obtained, it is thus justified to say that a robot
endowed with such artificial curiosity-based intelligence will necessarily
include autonomous cognitive capabilities. With respect to this, the fur-
ther perspectives regarding the autonomous cognitive robot presented in
this chapter will focus on integration of the investigated concepts in other
kinds of robots, such as mobile robots. There, it will play the role of an
underlying system for machine cognition and knowledge acquisition. This
knowledge will be subsequently available as the basis for tasks proper for
machine intelligence such as reasoning, decision making and an overall
autonomy.

References
[1] D. Meger, P. E. Forssén, K. Lai, S. Helmer, S. McCann, T. Southey,
M. Baumann, J. J. Little and D. G. Lowe, ‘Curious George: An attentive
semantic robot’, Robot. Auton. Syst., vol. 56, no. 6, pp. 503–511, 2008.
[2] P. Kay, B. Berlin and W. Merrifield, ‘Biocultural Implications of Systems
of Color Naming’, Journal of Linguistic Anthropology, vol. 1, no. 1,
pp. 12–25, 1991.
 References 93

[3] M. Bowerman, ‘How Do Children Avoid Constructing an Overly

General Grammar in the Absence of Feedback about What is Not a
Sentence?’, Papers and Reports on Child Language Development, 1983.
[4] M. A. Goodrich and A. C. Schultz, ‘Human-robot interaction: a survey’,
Found. Trends Hum.-Comput. Interact., vol. 1, no. 3, pp. 203–275, 2007.
[5] S. Coradeschi and A. Saffiotti, ‘An introduction to the anchoring
problem’, Robotics & Autonomous Sys., vol. 43, pp. 85–96, 2003.
[6] T. Regier, ‘A Model of the Human Capacity for Categorizing Spatial
Relations’, Cognitive Linguistics, vol. 6, no. 1, pp. 63–88, 1995.
[7] J. de Greeff, F. Delaunay and T. Belpaeme, ‘Human-robot interaction
in concept acquisition: a computational model’, Proc. of Int. Conf. on
Development and Learning, vol. 0, pp. 1–6, 2009.
[8] P. Wellens, M. Loetzsch and L. Steels, ‘Flexible word meaning in
embodied agents’, Connection Science, vol. 20, no. 2–3, pp. 173–191,
2008.
[9] J. Saunders, C. L. Nehaniv and C. Lyon, ‘Robot learning of lexical
semantics from sensorimotor interaction and the unrestricted speech of
human tutors’, Proc. of 2nd International Symposium on New Frontiers
in Human-Robot Interaction, Leicester, pp. 95–102, 2010.
[10] Lütkebohle, J. Peltason, L. Schillingmann, B. Wrede, S. Wachsmuth,
C. Elbrechter and R. Haschke, ‘The curious robot - structuring interactive
robot learning’, Proc. of the 2009 IEEE international conference on
Robotics and Automation, Kobe, pp. 2154–2160, 2009.
[11] T. Araki, T. Nakamura, T. Nagai, K. Funakoshi, M. Nakano and
N. Iwahashi, ‘Autonomous acquisition of multimodal information for
online object concept formation by robots’, Proc. of IEEE/ IROS,
pp. 1540–1547, 2011.
[12] D. Skocaj, M. Kristan, A. Vrecko, M. Mahnic, M. Janicek, G.-J.
M. Kruijff, M. Hanheide, N. Hawes, T. Keller, M. Zillich and K. Zhou,
‘Asystem for interactive learning in dialogue with a tutor’, Proc.of IEEE/
IROS, pp. 3387–3394, 2011.
[13] C. Yu, ‘The emergence of links between lexical acquisition and object
categorization: a computational study’, Connection Science, vol. 17,
3–4, pp. 381–397, 2005.
[14] S. R. Waxman and S. A. Gelman, ‘Early word-learning entails reference,
not merely associations’, Trends in cognitive science, 2009.
[15] D. E. Berlyne, ‘A theory of human curiosity’, British Journal of
Psychology, vol. 45, no. 3, August, pp. 180–191, 1954.
94 Autonomous Knowledge Discovery Based on Artificial Curiosity-Driven

[16] K. Madani, C. Sabourin, ‘Multi-level cognitive machine-learning based

concept for human-like artificial walking: Application to autonomous
stroll of humanoid robots’, Neurocomputing, S.I. on Linking of phe-
nomenological data and cognition, pp. 1213–1228, 2011.
[17] K. Madani, D. Ramik, C. Sabourin, ‘Multi-level cognitive machine-
learning based concept for Artificial Awareness: application to humanoid
robot’s awareness using visual saliency’, J. of Applied Computa-
tional Intelligence and Soft Computing,. DOI: 10.1155/2012/354785,
2012. (available on: http://dx.doi.org/10.1155/2012/354785).
[18] D. M. Ramik, C. Sabourin, K. Madani, ‘A Machine Learning based
Intelligent Vision System for Autonomous Object Detection and Recog-
nition’, J. ofApplied Intelligence, Springer, Vol. 40, Issue 2, pp. 358–374,
2014.
[19] D-M. Ramik, C. Sabourin. K. Madani, ‘From Visual Patterns to Semantic
Description: a Cognitive Approach Using Artificial Curiosity as the
Foundation’, Pattern Rgognition Letters, Elsevier, vol. 34, no. 14,
pp. 1577–1588, 2013.
[20] M. Schindler and J. W. v. Goethe, ‘Goethe’s theory of colour applied by
Maria Schindler’, New Knowledge Books, East Grinstead, Eng., 1964.
[21] V. Klingspor, J. Demiris, M. Kaiser, ‘Human-Robot-Communication and
Machine Learning’, Applied Artificial Intelligence, pp. 719–746, 1997.
[22] J. Hofmann, M. Jngel, M. Ltzsch, ‘A vision based system for goal-
directed obstacle avoidance used in the rc’03 obstacle avoidance chal-
lenge’, Lecture Notes in Artificial Intelligence, Proc. of 8th International
Workshop on RoboCup, pp. 418–425, 2004.
[23] D. M. Ramik, C. Sabourin, K. Madani, ‘On human inspired semantic
slam’s feasibility’, Proc. of the 6th International Workshop on Artifi-
cial Neural Networks and Intelligent Information Processing (ANNIIP
2010), ICINCO 2010, INSTICC Press, Funchal, pp. 99–108, 2010.
[24] R. Moreno, D. M. Ramik, M. Graña, K. Madani, ‘Image Segmentation
on the Spherical Coordinate Representation of the RGB Color Space’,
IET Image Processing, vol. 6, no. 9, pp. 1275–1283, 2012.
[25] D. M. Ramik, C. Sabourin, K. Madani, ‘Autonomous Knowledge Acqui-
sition based on Artificial Curiosity: Application to Mobile Robots in
Indoor Environment’, J. of Robotics and Autonomous Systems, Elsevier,
Vol. 61, no. 12, pp. 1680–1695, 2013.
 5
Information Technology for Interactive
Robot Task Training Through
Demonstration of Movement1

F. Kulakov and S. Chernakova

Laboratory of Informational Technologies for Control and Robots,

St. Petersburg Institute for Informatics and Automation of the Russian
Academy of Sciences, Russia
Corresponding author: F. Kulakov <[email protected]>

Abstract
Remote robot control (telecontrol) includes the solution of the following
routine problems: surveillance of the remote working area, remote operation
of the robot situated in the remote working area, as well as pre-training of the
robot. The current paper describes a new technique for robot control using
intelligent multimodal human-machine interfaces (HMI). The first part of the
paper explains the robot control algorithms, including testing of the results
of learning and of movement reproduction by the robot. The application of
the new training technology is very promising for space robots as well as for
modern assembly plants, including the use of micro-and nano-robots.

Keywords: Robot, telecontrol, task training by demonstration, human-

machine interfaces.

5.1 Introduction
The concept of telesensor programming (TSP) and relevant task-oriented
robot control techniques for use in space robotics was first proposed by
G. Herzinger [1].

1
The paper is published with financial support from the Russian Foundation for Basic Research,
projects 14-08-01225-a, 15-07-04415-a, 15-01-02021-a.

Advances in Intelligent Robotics and Collaborative Automation, 95–122.


c 2015 River Publishers. All rights reserved.
96 Information Technology for Interactive Robot Task Training

Within the framework of the ROTEX program, implemented in April 1993

for the SPACE-LAB space station, a simulation environment for multisen-
sory semiautonomous robot systems, with powerful man-machine interfaces
(laser range finders, 3D-stereo graphics and force/torque effort reflection),
was developed. This allowed the space robot manipulator to be remotely
programmed (teleprogrammed) from Earth.
The solution for the problem of remote control under non-deterministic
delays in the communications channel is based on the use of TSP with training
by demonstration for the sensitized robot.
Tasks such as assembly, joining of connectors and catching flying objects
were practiced. Actually, it was the first time that a human remotely trained
a robot through direct movement demonstration using a graphic model with
robot sensor simulation.
The effectiveness of interactive control (demonstration training) is high-
lighted in all cases of the application of pre-training technology to space and
medical robots, as the most natural way to transfer the operator’s experience
(SKILL TRANSFER) in order to ensure autonomous robot manipulator
operation in a complex non-deterministic environment [2–5].
However, in these studies it was only possible to conduct training with
the immediate recording of the movement trajectory positioning data and the
possibility of motion correction as per the signals from the robot’s sensors.
These studies did not solve the problem of complex robot motion repre-
sentation as a certain data structure that is easily adjustable by humans, or
“independently” modified by the autonomous robot, depending on changes in
the remote environment.
The current paper describes a new information technology-based approach
for interactive training by demonstration of the human operator’s natural hand
movements based on motion representation in the form of a frame-structured
model (FSM).
Here, a frame means a description of the shape of motion with indications
of its metric characteristics, methods and sequence of execution of the
separate parts of the movement. Training by demonstration means intelligent
robot manipulator programming aimed at training the robot for autonomous
work with the objects (among the objects) without point-to-point trajectory
recording. That is, by providing only separate fragments of movement in the
training stage, and sequentially executing them, depending on the task.
In order to train a robot manipulator to move among objects it was
suggested to use a remotely operated camera, fixed to the so-called “sensitized
glove”. This allows not only the registration of position and orientation of the
 5.2 Conception and Principles of Motion Modeling 97

hand in space, but also the position of the object (experimental models) models’
characteristic points relative to the camera on the hand.

5.2 Conception and Principles of Motion Modeling

5.2.1 Generalized Model of Motion
A variety of robot motion types in the external environment (EE), including
manipulation of items (objects and tools) as well as the complexity and
variability of EE configurations, are typical for aerospace, medical, industrial,
technological and assembly operations.
Let us consider the problem of training the robot manipulator to perform
motion relative to EE objects in two cases: examination motion and manipula-
tive motion. The main issue in forming the motion patterns, set, in this case, by
the motions of the operator’s head and arm, is to have a method for recording
and reproducing the three-dimensional trajectories of the robot manipulator
grip relative to EE objects.
The problem of the alignment of the topology and the semantics of objects,
well known in geographic information systems (GIS), is basically close to the
problem of motion modeling and route planning in robotics.
In the case of navigational routing tasks using intelligent GIS, the authors
basically consider motion along a plane (on the surface of the sphere) or
several planes (echelon gratings). Moreover, in most cases, the moving
object is taken as a mathematical point, not having its own orientation in
space.
The motion path configuration in space often does not matter, so routing
is carried out over the shortest distance. Thus, while following the curvature
of the relief, the motion tracks its shape.
For object shape modeling and motion formation, we propose using a
common structured description language, which considers that the object
shape model is defined and described by a frame of its elements, and the motion
trajectory model is described by a frame of descriptions of the elementary
motions. It is important to note that the elementary motions (fragments) can
be given appropriate names and be considered to be the language operators,
providing the possibility of describing robot actions in a rather compact
manner.
For interactive motion demonstration robot training, we propose using a
combination of the EE (MEE) objects’ shape models and the motion shape
models (MFM). In this case, the generalized frame-structured model (FSM)
98 Information Technology for Interactive Robot Task Training

is defined as a method for storing information not only about the shape of the
EE objects, but also about the shape of the motion trajectory.
The description language used in FSM is a multilevel hierarchical sys-
tem of frames, similar to M. Minski frames [6], containing a description
of the shape elements, metric characteristics and methods and procedures
for working with these objects. MFM, as one of the FSM components,
stores the structure of the shape of motion trajectories demonstrated by the
human operator during the process of training the robot to perform specified
movements [7, 8].
The generalized FSM of the remotely operated robot IE includes:
• Object models, models of the objects’ topology (location) in a particular
IE (MIE);
• Models of different typical motions and topology models (interrelations,
locations) of these movements in a particular IE (MIE).
It is also proposed to store, in the MIE, the coordinates and images of
objects from positions convenient both for remote-camera observation (which
enables the most accurate measurement of the coordinates of the characteristic
features of object images) and for grabbing objects with the robot gripper
(Figure 5.1) [9].
Training of motion can be regarded as a transfer of knowledge of motor,
sensory, and behavioral skills from a human operator to the robot control
system (RCS), which in this case should be a multimodal man-machine
interface (MMI), developed to the greatest possible extent (intelligent) to
provide adequate and effective perception of human actions. Consequently,

Figure 5.1 Images of the Space Station for two positions: “Convenient for observation” and
“Convenient for grabbing” objects with the virtual manipulator.
 5.2 Conception and Principles of Motion Modeling 99

it is assumed that a generalized model of description of the robot knowledge

on the EE based on the FSM will be created, including the robot itself and its
possible (necessary) actions within it.
The preliminary results of the research on algorithms and technologies
for the robot manipulator task training by demonstration, using the motion
description in the form of MFM, are presented below.

5.2.2 Algorithm for Robot Task Training by Demonstration

In order to task-train the robot by demonstration, a special device, the so-
called “sensitized glove,” is put on the hand of the trainer. It is equipped with
a television camera and check points (markers) [10].
This allows the execution of two functions simultaneously (Figure 5.2):
• Using the television camera on the glove, record the image and determine
the coordinates of the objects’ characteristic points, over which the hand
of the human operator moves;
• Using the sensors of the intelligent MMI system, determine the spatial
position and orientation of the hand in the work location by means of
3–4 check points (markers) on the glove.
Considering the processes for task-training a robot to perform elementary
operations and reproducing these operations, an important feature is revealed.
This feature consists in the fact that algorithms for training and repro-
duction present fragments, which are used in different operations without

Figure 5.2 “Sensitized Glove” with a camera and the process of training the robot by means
of demonstration.
100 Information Technology for Interactive Robot Task Training

modifications or with very minor changes, and may also be repeated several
times in a single operation.
From among the various movements of the robot manipulator, most of
them can be represented as a sequence of a limited number of elementary
motions (motion fragments), for example:
• Transfer motion of the gripper along an arbitrary complex trajectory
g = g(l) from the current position to a certain final position;
• Correction motion, using the sequence of characteristic points (CP) of
the EE objects’ images, as input information;
• Surveillance movement in the process by which the following are
sequentially created: matrices of the gripper position Tb , Tb1 , Tb2 , joint
coordinate vectors gb , gb1 , gb2 , and geometric trajectory g = g(l);
• Movement to a convenient position for surveillance;
• Movement to a convenient position for grabbing;
• Movement for “tracking” the object (approaching the object);
• Movement to grab the object.
In traditional training systems using one or the other method, a sequence of
points of the motion trajectory of the robot gripper is obtained. It can be
represented as a function of some parameter l, which can be considered as the
preliminary result of training the robot to perform the fragment of the gripper
movement from one position to the other:
g = g (l) , lb ≤ l ≤ le
gb = g (lb ) , ge = g (le ) ,
where: lb – parameter of the trajectory in the initial position, l – parameter
of the trajectory in the current position, le – parameter of the trajectory in the
final position.
In this case, the training algorithm for performing motions ensures the
formation of geometric trajectory g(l) and includes the following:
• Formation of a sequence of triplets of the two-dimensional vectors
ximb (1) , ximb (2) , ximb (3) ; ximI (1) , ximI (2) , ximI (3) ; . . . ; xime (1) , xime (2) ,
xime (3) , conforming to the image positions of the 3 CP on the object
during training;
• Formation of the sequence Tb , TI , TII , . . . , Te of the matrices of the glove
position;
• Solution of systems of equations (5.1):
(i) (i) (i)
xim = (xim1 , xim2 ) = k (i) TX (i) , (5.1)
 5.2 Conception and Principles of Motion Modeling 101

where: k(i) is a variable scale, defined as: k(i) = f /d(i) -f, where d(i) is the
distance from the point to the TV camera showing the plane; f is the focal
distance
 of thelens, Tis a (2x4) matrix made up of the first two rows of matrix:
 α Xn 
T = , characterizing the rotation and displacement of the system of
0 1 
coordinates (CS), in conjunction with the camera on the glove, relative to the
object CS, where a is the direction cosine matrix of the reference CS rotation
angle, Xn the displacement vector of the beginning of the CS and X(i) – the
two-dimensional vectors of the position of the image of the characteristic
points of the object in the image plane.
This data is sufficient to construct a sequence of matrices of the gripper
positions Tb , T1 , TII , . . . , Te during movement. The orientation blocks in these
matrices are matrices αb , α1 , αII , . . . , αe . The block of the gripper pole
position corresponds to the initial position of the gripper. According to this
sequence, the geometric, and, in line with it, the temporal motion trajectory
of the gripper can be built.
When teaching this action, the operator must move his hand with the glove
on it in the manner in which the gripper should move during the process of
the surveillance motion, whereas the position of the operator’s wrist can be
arbitrary and convenient for the operator.
Furthermore, for each case of teaching a new motion, it is necessary to
memorize a new volume of motion information in the form of several sets of
coordinates mentioned above.
When teaching the motions, e.g. IE surveillance, it is necessary to store a
considerable amount of information in the memory of the robot control system
(RCS), including:
• Values of matrix T, which characterize the position and orientation
of the glove in the coordinate system of the operator’s workstation,
corresponding to initial Tb , final Te and several intermediate T1 , TII , . . .
gripper positions, which it must take when performing movements;
• Several images of the object from the glove-mounted TV camera, cor-
responding to the various gripper positions, to control the accuracy of
training;
• Characteristic identification signs, characteristic points (CP) of the dif-
ferent images of the object, at different glove positions during the training
process;
• Coordinates of the CP of the images of the object in the base coordinate
system;
102 Information Technology for Interactive Robot Task Training

• Parameters of gripper opening and the permissible compressive force

applied to the subject.
To reduce the amount of information and to present the motion trajectory in a
language is close to the natural one, it is suggested to use a frame-structured
description in the motion shape model (MFM), the basic principles of which
are described in the previous papers by the authors [11, 12].

5.2.3 Algorithm for Motion Reproduction after Task Training by

Demonstration
The specific feature of the robot’s motion reproduction in a real IE is that
fragments of elementary movements, stored during task training can follow
a different sequence depending on the external conditions when reproduced.
Due to the aforementioned features, it appears to be reasonable to teach the
robot to do different fragments of motion in various combinations of the given
fragments.
The number of applied elementary motions (fragments) increases along
with the number of reproduced operations. However, this increase will be
much smaller than the increase in the number of operations for which the
robot is used. It is important to note that proper names can be assigned to
the given elementary motions and they can be considered to be operators of
the language with the help of which the robot’s actions can be described in a
sufficiently compact manner.
On the basis of the frame-structured description of the MFM, obtained
during task training, the so-called “tuning of the MFM” for a specific task
is performed before starting the reproduction of motion by the robot in a
particular IE situation.
Practically, this is done by masking or selection of only those descrip-
tions of motion in the MFD that satisfy the conditions of the task and the
external conditions of the situation in the IE according to their purpose
and shapes (semantic and topological features). The selected movements are
automatically converted into a sequence of elementary movements g = g(l).
In the case of the reproduction of the elementary motion along the trained
trajectory g = g(l) in systems without sensor offsetting it is necessary to
construct a parameter change function l(t) in the area lb ≤ l ≤ le . Typically,
the initial and final velocities l(t)are known, and they are most commonly
equal to zero:

l b = le = 0.
 5.2 Conception and Principles of Motion Modeling 103

In the simplest case of the formation of l(t), three intervals can be singled
out in it: the “acceleration” interval from the initial velocity (l b ) to some
permissible speed (l d ), the interval of motion at a predetermined speed and
the deceleration interval from the predetermined velocity to zero (l e ).
During acceleration and deceleration a constant acceleration (l d ) must
take place. Its value should be such that the value of the velocity g  and
acceleration g vectors can be physically implementable under the existing
restrictions of the control vector (U) of the robot manipulator’s motors.
The values of these limitations can be determined based on the consider-
ation of the dynamic model (R) of the robot manipulator, which connects the
control vector (U) to the motion dynamics vectors (g, g , g ):
 
U = R g, g  , g  .
In the case of the motion reproduction transfer of function l = l(t), it is defined
by the following ratio:
|l |
• During the acceleration interval (0 < t = t1 ), where t1 = sign (ld ) ld ;
| d|
• During the interval of motion at a constant velocity - (t1 ≤ t ≤ t2 ), where
|l |t2 sign(l )·t2
t2 = t1 + |lel−l b|
− dl 1 : l (t) = ld (t − t1 ) + lb 2 ld 1 ;
| d|

| d| | d|
• During the deceleration interval (t2 ≤ t ≤ t3 ), where
|l |t2
t3 = 2t1 + |lel−l b|
− dl 1 :
| d|

| d|
sign(ld )t21 sign(ld )ld (t − t2 )
l (t − t1 ) = ld (t − t1 ) + lb + − .
2 2
The reproduction of movement over time by the robot is carried out as per the
implementation of the obtained function l(t) in the motion trajectory g(l):
g = g (l (t)) .
To determine the drives’ control vector U = R (g, g , g ) the substitution of
values g, g , g by the values of function g = g (l (t)) is carried out. This results
in the formation of the control function of the motors of the robot manipulator
over time.
It should be noted that a man performs natural motions with constant accel-
eration, in contrast to the robot manipulator, whose motions are characterized
by a constant rate (speed). Therefore, the robot has to perform its motions as
per its own dynamics, which differ from the dynamic properties of the human
operator.
104 Information Technology for Interactive Robot Task Training

5.2.4 Verification of Results for the Task of Training the

Telecontrolled (Remote Controlled) Robot
Remotely operated robots must be sufficiently autonomous and trainable to be
able to efficiently perform operations in remote environments distant from the
human operator. Naturally, task training for space robots must be performed
in advance right here on Earth, and medical robots shall be trained out of the
operation theaters.
At the same time, a possibility for the remote correction of training
outcomes must be provided, for possible additional training by the human
operator, located at a considerable distance from the robot in space or from a
remotely controlled medical robot.
For greater reproduction reliability, it is necessary to implement an auto-
mated process control over motion reproduction by the remotely controlled
robot using copies of the MFM and MEE from the RCS. Remote control
over the robot movements by a human operator must be carried out using
prediction, taking into consideration the potential interference and time delays
in the communications channel.
Actual remote control of the space robot or the remotely operated medical
robot must be carried out as follows:
• With some time advance (prediction), simultaneously with working
motion execution by the robot, control over the current robot motion
is performed on the simulator, which includes the MEE, MFM and the
intelligent MMI system;
• Data from the RCS, arriving with delay, is reflected on the MEE and
MFM and is compared to the predicted movement in order to provide the
possibility of emergency correction;
• The trajectory of motion relative to the real location of the MEEs is
automatically adjusted by the RCS as per sensor signals;
• By human command (or automatically by the RCS) correction of param-
eters and operational replacement of the motion fragments are carried
out in accordance to the pre-trained alternative variants of the working
motion.
After the execution by the robot of the regular working movement, actual
motion trajectories in the form of a description in the language of MFM,
compiled after an automatic motion analysis in the RCS, are transferred from
the RCS to the human operator in the modeling environment. This information,
together with the results of the real EE scanning by the robot during the robot’s
 5.2 Conception and Principles of Motion Modeling 105

execution of the working movement, is to be used for correction of the MFM

and MEE.
In the absence of significant corrections in the process of executing the
working movement, the training is considered to be correct, understanding
between the human operator and the RCS is considered to be adequate, and
the results of the robot task training can be used in the future.

5.2.5 Major Advantages of Task Training by Demonstration

The proposed algorithm for task training by demonstration of the motion
has a number of advantages over conventional methods of programming
trajectories or motion copying, when the operator, for example, moves in
space the manipulator’s gripper along the desired trajectory with continuous
recording of the current coordinates in the memory of the robot. Let us list the
main ones.
Using the professional skills and intuitive experience of the human being.
The human being, using his professional skills and intuitive experiences,
demonstrates motions by hand, which are automatically analyzed by the MMI
(for configuration acceptance and safety) and are conveyed to the robot in the
form of a generalized MFM description. Conventional means of supervisory
control, in which remote control or a joystick, are used to set the generalized
command, are further developed in this case.
Simplicity and operational efficiency of training. Training is performed
by simple movements of the human hand without any programming of the
complex spatial displacements. It is more natural and easier for the human
being to control the position of his hand during the execution of movement,
than doing the same using the buttons, mouse or joystick. Experiments
have shown that practically everyone can learn to control a robot through
hand motion and it can be done in just a few hours. Time and cost of
personnel training, for the control and training of robots are significantly
reduced.
Relaxation in the requirements for motion accuracy. Instead of the exact
copying and recording of arrays of coordinates of the motion trajectory during
robot manipulator training, the operator gives only assignment (name) and
shape of the spatial motion trajectory, including the manipulation of items,
tools and EE objects. The free movement is set by the human being and is
reproduced by the robot at a certainly safe distance from the objects; therefore,
minute accuracy of such movements is not required. In the case where the robot
gripper approaches the object, the motion is automatically adjusted according
106 Information Technology for Interactive Robot Task Training

to information from the sensors, including the force-torque sensing. There is

no need to copy the exact motions by the remotely operated robots, which
are commonly used in partially nondeterministic EE, when there is no precise
information about the location of the robot and obstacles.
Reliability of control over the autonomous robot. One of the advantages
lies in the fact that there is no need for the operator to be close to the working
area or to be present in the working area of the remotely operated robot, for
example, inside the space station, or on the outer surface of the orbital station
for operational intervention in the robot’s actions, avoiding therefore delays
and interferences in the communications channels. Based on the descriptions
of the MFM and MEE, the intelligent RCS can automatically adjust the shape,
and even the sequence of the trained motions of the robot.
Ease of control, correction and transfer of motion experience. The visual
appearance of motion presentation in a MFM, its proximity to natural language
of the frame structured movement description, allow reliable checking, in-flow
change of composition, sequence and shape of the complex working move-
ments directly according to the motion description text using the graphical
model of robot manipulator, as well as a human model (“avatar”) [13].

5.3 Algorithms and Models for Teaching Movements

5.3.1 Task Training by Demonstration of Movement among
the Objects of the Environment
Robot task training and remote control is performed using the modeling
environment, which contains an EE model (MEE), a model of the shape of
motion (MFM) and an intelligent system for the multimodal interface (IMI),
creating the so-called human “presence effect” in a remote EE using the
three-dimensional virtual models and images of real items. Using the IMI,
the operator can observe 3-D images on either side, like in holographs, can
touch or move the virtual objects, feeling with his hand the tactile and force
impact through simulations of object weight or its weight in zero-gravity
environment [14, 16].
Instead of real EE models the virtual MEE image can be used, as well
as a computer hand model, controlled by the hand position tracking system
(HPTS), included in the IMI [17].
The process of training the robot to move among OE objects implies that
the operator’s hand, dressed in a sensitized glove, executes the motion in OE
space that must be subsequently performed by the manipulator gripper. In
 5.3 Algorithms and Models for Teaching Movements 107

order to do this, it is necessary to perform the following operations (in on-line

mode) in the training stage:
• Demonstrate a fragment of the operator’s hand motion among objects of
the OE model or of the virtual hand model in the graphical model of the
OE (MOE);
• Register through the IMM system and store in memory the fragment of
motion containing the timestamps and a corresponding vector (X) of 6
dimensions (x = x (l), y = y (l), z = z (l)) and orientation ( ϕx = ϕx (l),
ϕy = ϕy (l), ϕz = ϕz (l)) of the operator’s hand;
• Recognize the motion shape through the IMI system and record the results
in the form of a frame-based description in MFM;
• Record the images of objects, obtained through the TV camera on the
glove in the process of moving the hand and carry out recognition,
identification and measurement of the coordinates of the characteristic
points of the objects’ images and enter this data into the MOE;
• Add to the MFM the information about the location of MOE objects
relative to the glove at the moment of execution of the fragment of
movement.
The position of the sensitized glove relative to objects of the EO model is
determined during the process of training by demonstration by solving the
so-called “navigation task”. This research offers a unique solution of the
navigation task for the given case [18].
While training the robot by means of demonstration, the objects (models)
of the OE come in view of the TV camera fixed on the sensitized glove. There
can be objects of manipulation or foreign objects (obstacles).
For the industrial and aerospace application of robots, the objects generally
have regular geometric shapes, angles and edges, which may be used as
characteristic features and characteristic points.
Characteristic points (CP) can be small-sized (point) details of objects,
which can be easily distinguished on the image, as well as special marks
or pointed sources of light. These points are the easiest way to determine
the position and orientation of the camera on the sensitized glove relative to
the OE objects, that is, to solve the so-called “navigation problem” during the
process of robot task training.
Let us consider a case, where the position vectors of the object’s CP X(i) ,
(i = 1, 2, 3 – No. of CP) in a coordinate system associated to the object
(CS) are known beforehand. Images of 3 CPs of the object (Xim (1) , Xim (2) ,
Xim (3) ) on the TV camera’s image surface are projections of the real points
108 Information Technology for Interactive Robot Task Training

(CP1 ... CP3) on this plane in a variable scale k(i) =f /d(i) -f, inversely
proportional to the distance d(I) from the point to the imaging plane of the
lens, where f is the focal length of the lens.
Let us assume that the CS, associated with the camera lens, and, therefore,
with the glove, is located as shown in Figure 5.3, i.e. axes x1 and x2 of the
CS are located in the image plane, x3 is perpendicular to them and is directed
away from the lens towards the object. In Figure 5.3: x1 , x2 , x3 are the axes of
the coordinate system associated with the object; x(1) , x(2) , x(3) are the vectors
defining the position of characteristic points in the coordinate system of the
camera lens; xim (2) 1 , xim (2) 2 are 2 projections of the vector from the center
of the CCD matrix to the image of point 2 (this can also be shown for points
1 and 3).
Then, distance d(i) is equal to the projection of the i-th CP on the third axis
of the CS associated with the camera: d(i) = xim3 (i) , and the location of the
object X(i) in the image plane will be represented by two-dimensional vectors:
(i) (i) (i)
xim = (xim1 , xim2 ) = k (i) TX (i) , (5.2)

Figure 5.3 Formation of images of 3 characteristic points of the object.

 5.3 Algorithms and Models for Teaching Movements 109

where:
 T is a (2×4) matrix made up of the first two rows of matrix
 α Xn 
T =  characterizing the rotation and displacement of the CS asso-
0 1 
ciated to a camera on the glove, relative to the CS of the object, where a is a
direction cosine matrix relative to the turning angle of the SC and Xn = (Xn1 ,
Xn2 , Xn3 ) is the displacement vector of the SC’s origin,
(i) (i) (i)
X (i) = (x1 , x2 , x3 , 1).

Then: d(i) = xim3 (i) = T3 . x(i) , where T3 is the third row of matrix T.
It is obvious that matrix T completely determines the spatial position of
the glove in the CS associated to the object, and its elements can be found as
the result of solving the abovementioned navigation problem of determining
the spatial position of the glove during training.
During the CP image processing, vectors xim (i) , i = 1, 2, 3 are determined,
so the left side of Equations (5.2) is known, and these equations represent a
system of six equations concerning 12 unknown elements of matrix T, which
are the three components of vector Xn and nine elements of matrix a.
Since the elements of matrix a are linked by six more equations of
orthogonality and orthonormality, there are a total of 12 equations, that is,
as many as the unknowns. These are obviously sufficient to determine the
desired matrix T.
During the “training” motion of the operator’s hand at a given frequency,
a procedure involving an operation for the selection of the object’s CP image
and an operation for calculating the values of two-dimensional vectors xim (i) ,
i = 1, 2, 3 and their position in the image plane must be performed.
As a result of these actions, a sequence of values of the vector triplets
from the starting one Ximb (i = 1,2,3) to the finishing one xime (i = 1,2,3) :
(ximb (1) , Ximb (2) , Ximb (3) ); (ximI (1) , XimI (2) , XimI (3) ); (ximII (1) , XimII (2) ,
XimII (3) ); . . . (xime (1) , Xime (2) , Xime (3) ) is accumulated in the IMI database,
corresponding to the sequence of the glove’s positions during the movement of
the operator’s hand, which will be later reproduced by the robot. Each element
of this sequence carries enough information to solve the navigation task,
that is, to obtain the sequence Tb , TI , TII , . . . , Te of the matrix values, which
is the result of training.
After training, the robot reproduces a gripper motion based on sequence
Tb , TI , TII , . . . , Te using a motion correction algorithm based on the signals
from the camera, located in the gripper, by solving the so-called “correction
task” of the gripper relative to real OE objects.
110 Information Technology for Interactive Robot Task Training

5.3.2 Basic Algorithms for Robot Task Training by Demonstration

The most typical example of the robot’s interaction with OE objects is the
manipulation of arbitrarily oriented objects. In practice, the task of grabbing
objects has several cases. The simplest case is when there is one known object.
The robot must be trained to perform an operation of grabbing this object
irrespective of any minor changes in its position and orientation.
A more complicated case is when the position and orientation of a
known object are not known beforehand. The most typical case is when
there are several known objects with a priori unknown positions and ori-
entations. And an even more complex task is when among the known
objects there are unknown objects and obstacles that may hinder the grabbing
procedure.

5.3.3 Training Algorithm for the Environmental Survey Motion

During the training to perform the environmental survey motion, the operator’s
hand executes one or more types of search movements: rotation of the hand at
two angles, zigzag motion, etc. Information about the typical search motions
is recorded in the MFM. Survey motion may consist of several fragments of
different movements. The sequence and shape of these motions, dependent
on the task, are determined during the training phase and stored in the MFM.
After the execution of separate motion fragments, a break can be taken for
further analysis of the OE objects’ images.
Any OE object recognition algorithm suitable for a particular purpose can
be used, including a structural recognition algorithm [19].
It is necessary to note that object image analysis must include the
following:
• Recognition of the target subject through a set of characteristic features
(XT1, XT2, . . . XTk) that are sufficient to identify it using the description
stored in the MOE;
• Selection of a set of reference points (XT1, XT2, . . . XTn) that are
sufficient for navigating the robot gripper in the space of the real OE,
from among a set of points (XT1, XT2, . . . XTk), usually no more than
n = 4–6, depending on their location.
If the number of CPs observed on the object is insufficient (k < n), it is
necessary to perform the following search motions:
• Change the position or orientation of the camera on the glove so that
k = n or change the CP filter parameters in the recognition algorithm;
 5.3 Algorithms and Models for Teaching Movements 111

• Change the observation conditions or camera parameters for reliable

detection of CP, such as lighting of the operator workstation or focus of
the camera lens:
• Add artificial (contrasting, color) marks on the graphical model or
on the object model to recommend the use of these marks on real
objects;
• If k ≥ n, it is possible to skip to the calculation of the spatial position
and orientation of the glove relative to the object in accordance to
the algorithm of the “navigation task” (see above).
Once the specified object is detected and identified and the position and
orientation of the hand (glove) relative to this object is determined, the training
to execute the first survey motion is deemed finished.
The purpose of the next motion the robot is trained to execute involves a
gripper motion to the so-called “convenient for observation” position. In this
position, the maximum identification reliability and measurement accuracy of
the gripper position relative to the object are achieved.
The variants of the shift from the starting point of object detection to
the position which is “convenient for observation” must be shown by the
movement of the operator’s wrist using his intuitive experience.
There is also an option of task training by demonstration, for survey
movement, performed through natural head movements. In this case, the
camera with a reference device is fixed on the operator’s head.
The training process ends automatically, for example, upon a signal
from the IMI system after reaching the specified recognition reliability and
measurement accuracy parameters for the position of a hand or a head relative
to the OE object. A training halt signal can be given by voice (using the speech
recognition system of the IMI) or by pressing the button on the glove. In this
case, the object coordinates defined by its image are recorded in the MFM as
a vector of coordinates (X0 ).

5.3.4 Training Algorithm for Grabbing a Single Object

In this case, the grabbing process consists of three movements:
• Gripper motion from the initial position to the “convenient for grabbing”
the object position;
• Object gripping motion, for example, a simple translational motion along
the gripper axis;
• Object grabbing motion, such as a simple closing of the gripper, exerting
a given compression force.
112 Information Technology for Interactive Robot Task Training

Let’s consider the task training to perform only the first action, where the oper-
ator freely moves his hand, with sensitized glove on it, to the object (model)
from the initial position at the most convenient for grabbing side and sets it
at a short distance from the object with the desired orientation. Information
about the motion path and the hand position relative to the object, at least
at the end point of the motion, is memorized in the MFM through the IMI
system, which is necessary for adjusting the robot’s gripper position relative
to the object during the motion reproduction. It’s also desirable that at least
1 or 2 CPs of the object’s image get into the camera’s view in the “convenient
for grabbing” gripper position, so that the grabbing process can be supervised.
The training of the grabbing motion is performed along the easiest path, in
order to reduce the guidance inaccuracy. If the gripper is equipped with object
detection sensors, then the movement ends upon receiving signals from these
sensors.
During the training to grab objects, it is necessary to memorize the
transverse dimensions of the object at the spot of grabbing and the gripper
compression force, sufficient for holding the object without damaging it.
This operation can be implemented using additional circuit-torque and tactile
sensors in the robot gripper.
In case of the presence of multiple OE objects, the training implies a more
complex process of identification of the objects’ images and the necessity
to train additional motions, such as obstacle avoidance, changing of altitude
convenient for survey in case of any shading, flashing and interference to
image recognition by the camera on the glove, as well as for the camera in the
robot manipulator’s gripper, during the reproduction of movements.

5.3.5 Special Features of the Algorithm for Reproduction of

Movements
As a result of performing the required number of training movements by the
human hand, “motion experience” is formed, which is accumulated in the
form of a frame-structured description in the MFM, stored in the memory of
the intelligent IMI system.
The transfer of the “motion experience” from the human to the robot
occurs, for example, by simply copying the MFM and MEE from the IMI
memory to the RCS memory or even by transferring this data to the remotely
controlled robot over communications channels. Of course, preliminary
checking of training outcomes is performed, for example, on a graphical model
of the robot.
 5.3 Algorithms and Models for Teaching Movements 113

Prior to the robot performing the trained movements, in accordance to the

assigned task and the EE conditions, the MFM is tuned, as already mentioned
(in Part I of the current paper), for example, by masking (searching) among the
total volume of the MFM data for the required types of motions. Descriptions
of motions, selected according to the intended purpose and shape of the
trajectory, are converted by the RCS into typical motion trajectories for their
reproduction by the robot in real EE.
When the robot-manipulator reproduces motions in a real EE, after training
by demonstration, it is possible to execute, for example, the following typical
motion trajectories:
• Survey movement in combination with EE image analysis in order to
identify the object to be taken;
• Shifting of the gripper into the “convenient for observation” position;
• Corrective gripper movement relative to the object based on the signals
from the robot’s sensors;
• Shifting of the gripper to the “convenient for taking” position;
• Motion for grabbing the real object;
• Motion for taking the object.
Before the work starts, a complete check of the TSHP operation and telecontrol
system is carried out. Then, operation of the TSHP is checked using a graphical
model (GM) at the control station without using a robot and exchanging infor-
mation over the communications lines. Checks of the training outcomes are
performed using the surveillance MFM or manipulation MFM, located in the
RCS, without switching on the robot at this moment. If necessary, additional
adjustment of the MFM is performed through task training by demonstration of
the natural human-hand movements and their storage in the MFM of the RCS.
The robot is switched on and it executes motions in terms of the orig-
inal EE inspection, selection of objects, position selection, convenient for
grabbing or convenient for visual control over object grabbing and manipula-
tions, as well as safe obstacle avoidance, before a transition to remote control
mode is performed.
The human operator sits in front of the monitor screen, which displays a
graphical model or a real object image, and controls the robot through natural
movements of the head and hand with the glove.

5.3.6 Some Results of Experimental Studies

The effectiveness of the proposed training technology using demonstrations
of the movements, the algorithms and theoretical calculations was tested on
114 Information Technology for Interactive Robot Task Training

the basis of the “virtual reality” environment at the laboratory of Information

Technology in Robotics, SPIIRAS (St. Petersburg) [20].
The hardware-software environment includes:
• Two six-stage robotic manipulators of the «Puma» class, equipped with
remotely operated stereo cameras and force-torque sensing;
• Models of the fragment of the space station surface, two graphic stations
to work with three-dimensional models of the external environment
(MEE);
• Intelligent multimodal interface (IMI) with a system for tracking hand
movements (TSHP) and a system for tracking the head motions (THM)
of the human operator.
The “Virtual reality” environment enables the performance of experimental
studies of various information technology approaches for remote control and
task training of robots:
• “Immersion technologies” of the human operator in the remote environ-
ment using the robot-like device that moves surveillance stereo cameras
in the room with models of the fragment of the space station surface and
containers for scientific equipment;
• “Virtual observer” technologies using the model of the freely flying
machine (equipped with the surveillance camera, which allows the exam-
ination of the three-dimensional graphical model of the space station),
as well as the simulation of an astronaut’s work in outer space;
• Technologies for training and remotely controlling a space (medical)
robot manipulator with a force-torque sensing system, which provides
operational safety during manipulative operations, reflection of forces
and torques on the control handle, including when working with virtual
objects.
Experimental studies were performed on some algorithms for training by
demonstration and remote control of a robot manipulator, including:
• Training of the robot manipulator to execute survey motions through
motions of the human head;
• Scanning the surroundings by the robot and remotely operated camera
on the glove;
• Using the IMI for training by demonstration of hand movements and
human voice commands;
• Training the robot manipulator to grab items by demonstration of the
operator’s hand movements.
 5.3 Algorithms and Models for Teaching Movements 115

The motion reproduction of the robot manipulator among the real EE objects
based on the use of the virtual graphical models of the EE and the robot
manipulator with force-torque sensing system was also practiced in the
experimental environment.

5.3.7 Overview of the Environment for Task Training by

Demonstration of the Movements of the Human Head
A functional diagram of the equipment for remotely monitoring the EE is
shown on Figure 5.4.
The operator, located in the control room, sets coordinates and orientations
of the manipulator gripper and remotely operated camera on it using the
tracking system for head position (THM). He observes the obtained EE image
on the monitor screen.

Figure 5.4 Functional diagram of robot task training regarding survey motions and object
grabbing motions using THM and RCS.
116 Information Technology for Interactive Robot Task Training

Before starting, the human operator must be able to verify the THM in an
off-line mode. For this purpose, a graphical model (GM) and a special com-
munications module for controlling 6 coordinates were developed. Training
the robot manipulator to execute EE surveillance motions by demonstration
is carried out in the following way (Figure 5.5).
The human operator performs the EE inspection based on his personal
experience in object surveillance. The robot repeats the action, using the
surveillance procedure and shape of the trajectory of the human head move-
ment. In this case, the cursor can first be moved around the obtained panoramic
image, increasing (decreasing) the scale of separate fragments, and then,
after accuracy validation, the actual motion of the robot manipulator can be
executed.

5.3.8 Training the Robot to Grab Objects by Demonstration of

Operator Hand Movements
There are several variations of the implementation of the “sensitized glove”
(Figure 5.6): a remote-operated camera in the bracelet with control points
and laser pointers, bracelet with active control points (infrared diodes),
manipulation object - stick with the active control points [21].
When training by demonstration of human hand movements, through a
sensitized glove with camera and control points, a greater range and closeness
to natural movements is achieved, as compared to the use of joysticks or handle
like “Master-Arm» (Figure 5.7).
This provides for the natural coordination of movements of the hand
and head of the human operator. Using the head, the human controls the
movement of the remotely operated surveillance camera, fixed, for example,

Figure 5.5 Robot task training to execute survey movements, based on the movements of
the operator’s head: Training the robot to execute survey motions to insect surroundings (a);
Training process (b); Reproduction of earlier trained movements (c).
 5.3 Algorithms and Models for Teaching Movements 117

Figure 5.6 Variations of the “Sensitized Glove” construction.

Figure 5.7 Using the special glove for training the robot manipulator.

on an additional manipulator, and with the hand he controls the position and
orientation of the main robot gripper (Figure 5.8).

Figure 5.8 Stand for teaching robots to execute motions of surveillance and grabbing objects.
118 Information Technology for Interactive Robot Task Training

The coordination of simultaneous control using the operator’s hand and

head during training and remote control through natural human operator
movements was put into practice in order to control the complex objects
(Figure 5.9).
A new prototype of the intelligent IMI equipment with recognition of
the operator’s hand without markers, while performing manual control and
training by demonstration of natural hand movements, was experimentally
studied (Figure 5.10). In the future it is planned to continue research on new
algorithms for training and remote robot control of intelligent mechatronic
systems based on the use of advanced intelligent multimodal human-machine
interface systems and new motion modeling principles using frame-structured

Figure 5.9 Training of motion coordination of two robot manipulators by natural movements
of human head and hand.

Figure 5.10 Training with the use of a system for the recognition of hand movements and
gestures without “Sensitized Gloves” against the real background of the operator’s work station.
 References 119

MFM descriptions, including for medical robots, mechatronic systems and

telemedicine [22].

5.4 Conclusions
A new information technology approach for training robots (mechatronic
systems) by demonstration of movement is based on the use of a frame-
structured data representation in the models of the shape of the movements
that makes it easy to adjust the movement’s semantics and topology both for
the human operator and for the autonomous sensitized robot.
Algorithms for training by demonstration of natural movements of the
human operator’s hand using a television camera, fixed on the so-called
“sensitized glove”, allow not only the application during the training process
of graphical models of objects in surroundings but also of full-scale models,
which enables the operator the possibility to practice optimal motions of the
remote-controlled robots under real conditions.
It is sufficient to demonstrate the shape of a human hand movement to
the intelligent system of the IMI and to enter it into the MFM, and then this
movement can be executed automatically, for example, by a robot manipulator
with adjustment and navigation among the surrounding objects based on the
signals from the sensors.

References
[1] G. Herzinger, G. Grunwald, B. Brunner and J Heindl, ‘A sensor-based
telerobot system for the space robot experiment ROTEX’, Proc. 2nd
Internat. Symp. on Experimental Robots (ISER). Toulouse. France.
1991. June 25–27.
[2] G. Herzinger, J. Heindl, K. Landzettel and B. Brunner, ‘Multisensory
shared autonomy – a key issue in the space robot technology experiment
ROTEX’, Proc. IEEE Conf. on Intelligent Robots and Systems (IROS).
Raleigh. 1992. July 7–10.
[3] Klaus H. Strobl, Wolfgang Sepp, Eric Wahl, Tim Bodenmüller, Michael
Suppa, Javier F. Seara, Gerd Hirzinger, ‘The DLR Multisensory
Hand-Guided Device: the Laser Stripe Profiler’, ICRA 2004:
1927–1932.
[4] Michael Pardowitz, Steffen Knoop, Rüdiger Dillmann, R. D. Zollner,
‘Incremental Learning of Tasks From User Demonstrations, Past
120 Information Technology for Interactive Robot Task Training

Experiences, and Vocal Comments’, IEEE Transactions on Systems,

Man, and Cybernetics, Part B 37(2): 322–332 (2007).
[5] R. Dillmann, O. Rogalla, M. Ehrenmann, R. Zoellner and M. Bordegoni,
‘Learning robot behavior and skills based on human demonstration and
advice’, ‘The machine learning paradigm’, 9th International Symposium
of Robots Research (ISSR), pp. 229–238, Oct., 1999.
[6] M. Minsky, ‘Frames for knowledge representation’, M.: “Energia”,
1979.
[7] S. E. Chernakova, F. M. Kulakov, A. I. Nechayev, ‘Simulation of the
outdoor environment for training by demonstration’, Collected works
of the SPII RAS. Issue No. 1.: St. Petersburg: SPIIRAS, 2001.
[8] S. E. Chernakova, F. M. Kulakov, A. I. Nechayev, A. I. Burdygin, ‘Mul-
tiphase method and algorithm for measuring the spatial coordinates of
objects for training the assembly robots’, Collected works of SPIIRAS.
Issue No. 1.: St. Petersburg: SPIIRAS, 2001.
[9] F. M. Kulakov, ‘Technology for the Creation of Virtual Objects in
the Real World’, Workshop Conference, Binghamton University, NY,
4–7 March, 2002.
[10] Advanced System for Learning and Optimal Control of Assembly
Robots, edited by F. M. Kulakov. SPIIRAS, St. Petersburg, 1999,
pp. 76
[11] F. M. Kulakov, V. B. Naumov, ‘Application of fuzzy logic and sensitized
glove for programming robots’, Collected works of the III Interna-
tional Conference “Current Problems of Informatics-98”. Voronezh,
pp. 59–61.
[12] C. E. Chernakova, F. M. Kulakov, A. I. Nechayev, ‘Training robot by
method of demonstration with the use of “sensitized” glove’, Works
of the First International Conference on Mechatronics and Robots.
St. Petersburg, May 29 – June 2, 2000, pp. 155–164.
[13] Chernakova S. E., Timofeev A. V., Nechaev A. I., Litvinov M. V.,
Gulenko I. E., Andreev V. A., ‘Multimodal Man-Machine Interface
and Virtual Reality for Assistive Medical Systems’, International Jour-
nal “Information Theories and Applications” (iTECH-2006). Varna,
Bulgaria, 2006.
[14] F. M. Kulakov, S. E. Chernakova, ‘Information technology of task train-
ing robots by demonstration of motions’, Part I “Concept, principles of
modeling movements”, “Mechatronics, Automation, Control”. Moscow,
No. 6, 2008.
 References 121

[15] F. M. Kulakov, ‘Potential techniques of management supple movement

of robots and their virtual models Part I’, “Mechatronics, Automation,
Control”. Moscow. No. 11. 2003.
[16] F. M. Kulakov, ‘Potential management techniques supple movement of
robots and their virtual models Part II’, “Mechatronics, Automation,
Control”. No. 1. Moscow, 2004. pp. 15–21.
[17] S. Chernakova, A. Nechaev, A. Karpov, A. Ronzhin, ‘Multimodal system
for hands-free PC control’, 13th European signal Processing Conference
EUSIPCO-2005. Turkey. 2005.
[18] F. M. Kulakov, A. I. Nechaev, S. E. Chernakova and oth., ‘Eye Track-
ing and Head-Mounted Display/Tracking Computer Systems for the
Remote Control of Robots and Manipulators’, Project # 1992P, Task
5 with EOARD. St. Petersburg, 2002.
[19] A. I. Nechaev, J. S. Vorobjev, I. N. Corobkov, M. S. Olkov, V. N. Javnov,
‘Structural Methods of recognition for real time systems’, Interna-
tional Conference on Modeling Problems in Bionics “BIOMOD-92”.
St.-Petersburg, 1992.
[20] F. M. Kulakov, ‘Technology of immersion of the virtual object into
real world’, Supplement to magazine “Informatsionnyie Tekhnologii”,
No. 10, 2004, pp. 1–32.
[21] C. E. Chernakova, F. M. Kulakov, A. I. Nechayev, ‘Hardware and soft-
ware means of the HMI for remote operated robots with the application
of systems for tracking human operator’s movements’, Works of the
III International Conf. “Cybernetics and technologies of XXI century”,
Voronezh, 2002.
[22] C. E. Chernakova, A. I. Nechayev, V. P. Nazaruk, ‘Method of recording
and visualization of three-dimensional X-ray images in real-time module
for tasks of nondestructive testing and medical diagnostics’, Magazine
“Informatsionnyie Tekhnologii”, No. 11, 2005, pp. 28–37.
 6
A Multi-Agent Reinforcement Learning
Approach for the Efficient Control
of Mobile Robots

U. Dziomin1 , A. Kabysh1 , R. Stetter2 and V. Golovko1

1 Brest
State Technical University, Belarus
2 University
of Ravensburg-Weingarten, Germany
Corresponding author: R. Stetter <[email protected]>

Abstract
This paper presents a multi-agent control architecture for the efficient control
of a multi-wheeled mobile platform. Such control architecture is based on
the decomposition of a platform into a holonic, homogenous, multi-agent
system. The multi-agent system incorporates multiple Q-learning agents,
which permits them to effectively control every wheel relative to other
wheels. The learning process was divided into two steps: module positioning –
where the agents learn to minimize the error of orientation, and cooperative
movement – where the agents learn to adjust the desired velocity in order to
conform to the desired position in formation. From this decomposition, every
module agent will have two control policies for forward and angular velocity,
respectively. Experiments were carried out with a simulation model and the
real robot. Our results indicate a successful application of the purposed control
architecture both in the simulation and in real robot.

Keywords: control architecture, holonic homogenous multi-agent system,

reinforcement learning, Q-Learning, efficient robot control.

6.1 Introduction
An efficient robot control is an important task for the applications of a mobile
robot in production. The important control tasks are power consumption
optimization and optimal trajectory planning. Control subsystems should

Advances in Intelligent Robotics and Collaborative Automation, 123–146.


c 2015 River Publishers. All rights reserved.
124 A Multi-Agent Reinforcement Learning Approach for the Efficient Control

provide energy consumption optimization in a robot control system. Four

levels of robot power consumption optimization can be distinguished:
• Motor power consumption optimization. Those approaches based on
energy-efficient technologies of motor development that produce sub-
stantial electricity saving and improve the life of the motor drive
components [1, 2];
• Efficient robot motion. Commonly, this is a task of an inverse kinematics
calculation. But the dynamic model is usually far more complex than the
kinematic model [3]. Therefore, intellectual algorithms are relevant for
the optimization of a robot motion [4];
• Efficient path planning. Such algorithms build a trajectory and divide it
into different parts, which are reproduced by circles and straight lines. The
robot control subsystem should provide movement along the trajectory
parts. For example, Y. Mei and others show how to create an efficient
trajectory using knowledge of the energy consumption of robot motions
[5]. S. Ogunniyi and M. S. Tsoeu continue this work using reinforcement
learning for path search [6];
• Efficient robot exploration. When a robot performs path planning
between its current position and its next target in an uncertain environ-
ment, the goal is to reduce repeated coverage [7].
The transportation of cargo is an actual task in modern production. Multi-
wheeled mobile platforms are increasingly being used in autonomous trans-
portation of heavy components. One of these platforms is a production mobile
robot, which was developed and assembled at the University of Ravensburg-
Weingarten, Germany [3]. The robot is illustrated in Figure 6.1(a). The
platform dimensions are 1200cm in length and 800cm in width. The maximum
manufacturer’s payload is 500kg, battery capacity is 52Ah, and all modules
drive independently.
The platform is based on four vehicle steering modules [3]. The steering
module (Figure 6.1(b)) consists of two wheels powered by separate motors
and behaves like a differential drive.
In this paper, we explore the problems of formation control and efficient
motion control of multiple autonomous vehicle modules in circular trajectory
motion. The goal is to achieve a circular motion of a mobile platform around
a virtual reference beacon with optimal forward and angular speeds.
One solution to this problem, [8–10] is to calculate the kinemat-
ics of a one-wheeled robot for circle driving and then generalize it for
multi-vehicle systems. This approach has shown promising modeling results.
 6.2 Holonic Homogenous Multi-Agent Systems 125

Figure 6.1 Production mobile robot: Production mobile platform (a); Driving module (b).

The disadvantage of this technique is its low flexibility and high computational
complexity.
An alternative approach is to use the machine learning theory to obtain
an optimal control policy. The problem of multi-agent control in robotics is
usually considered as a problem of formation control, trajectory planning,
distributed control and others. In this paper we use techniques from multi-
agent systems theory and reinforcement learning to create the desired control
policy.
The content of this paper is the following: Section 6.2 gives a short
introduction to the theory of holonic, homogenous, multi-agent systems
and reinforcement learning. Section 6.3 describes the steering of a mobile
platform in detail. Section 6.4 describes the multi-agent decomposition of a
mobile platform. Using this decomposition, we propose a multi-agent control
architecture based on the model described in Section 6.2. Section 6.5 contains
a detailed description of the multi-agent control architecture. The Conclusion
highlights important aspects of the presented work.

6.2 Holonic Homogenous Multi-Agent Systems

A multi-agent system (MAS) consists of a collection of individual agents,
where each agent displays a certain amount of autonomy about its actions
and perception of domain, and communicates via message-passing with
another agent [11, 12]. Agents act in organized structures which encapsulate
the complexity of subsystems and therefore modularize its functional-
ity. Organizations are social structures with means of conflict resolution
through coordination mechanisms [13]. The overall emergent behavior of a
126 A Multi-Agent Reinforcement Learning Approach for the Efficient Control

multi-agent system is composed of a combination of individual agent behav-

iors determined by autonomous computation within each agent, and by
communication among agents [14]. The field of MAS is a part of dis-
tributed AI, where each agent has a distinct problem solver for a specific
task [12, 14].

6.2.1 Holonic, Multi-Agent Systems

An agent (or MAS) that appears as a single entity to the outside world
but is in fact composed of many sub-agents with the same inherent
structure is called holon, and such sub-agents are called holonic agents
[11, 14]. The transformation of a single entity into a set of interacting
subagents is called holonic decomposition. Holonic decomposition is an
isomorphic transformation. Gerber et al. [15] show that an environment
containing multiple holonic agents can be isomorphically mapped as an
environment in which exactly one agent is represented explicitly, and vice
versa.
For the purposes of this paper and without the loss of generality, we use
terms holon and holonic multi-agent system (Holonic MAS) interchangeably,
meaning that a MAS contains exactly one holon. In the general case, a holonic,
multi-agent system (called holarhy) is a self-organized, hierarchical structure
composed of holons [14].
A holon is always represented as a single entity to the outside world.
From the perspective of the environment, a holon behaves as an autonomous
agent. Only a closer inspection reveals that a holon is constructed from a
set of cooperating agents. It is possible to communicate with a holon simply
by sending messages to them from the environment. The most challenging
problem in this design is the distribution of individual and overall computation
of the holonic MAS [15].
Although it is possible to organize holonic structures in a completely
decentralized manner, it is more efficient to use an individual agent to represent
a holon. Representatives are called the head of the holon; the other agents in
the holon are called the body [11]. In some cases, one of the already existing
agents is selected as the representative of the holon. In other cases, a new
agent is explicitly introduced to represent the holon during its lifetime.
The head agent represents the shared intentions of the holon and negotiates
these intentions with the agents in the holon’s environment, as well as with
the internal agents of the holon. Only the head agent communicates with the
entities outside of the holon.
 6.2 Holonic Homogenous Multi-Agent Systems 127

Figure 6.2 Organizational structure of a Holonic Multi-Agent System. Lines indicate the
communication channels.

The organizational structure of a holonic, multi-agent system is depicted

in Figure 6.2.
When agents join the holon, they surrender some of their autonomy to the
head agent. The binding force that keeps the head and body in a holon together
can be called a commitments [16]. It should be explicitly noted that agents are
not directly controlled by the head agent. The agents remain autonomous
entities within the holon, but they align their individual behavior with the
goals of holon.

6.2.2 Homogenous, Multi-Agent Systems

For the purposes of this paper, we will consider the case when all body agents
are homogenous. In a general, multi-agent scenario with homogeneous agents,
there are several different agents with an identical structure (sensors, effectors,
domain knowledge, and decision functions) [17]. The only differences among
agents are their sensory inputs and the actual actions they take, as they are
situated differently in the world [18]. Having different effector outputs is a
necessary condition for MAS; if the agents all act together as a unit, then
they are essentially a single agent. In order to realize this difference in output,
homogeneous agents must have different sensor input as well. Otherwise, they
will act identically.
128 A Multi-Agent Reinforcement Learning Approach for the Efficient Control

Thus, the formal definition of holonic, homogenous, multi-agent system

(H2 MAS) is a tuple H = B, h, C:
• B = {M1 , M2 , . . . , Mn }– is the set of homogenous body agents. Each
agent is described by a tuple Mi = s, α, πi , where:
• s – is the set of possible agent states, where si ∈ s is the i-th agent current
state;
• α– is the set of possible agent actions, where ai ∈ α current action of
the i-th agent;
• π : s → α is the behavior policy (decision function) which maps it’s
state to actions;
• h – is the head agent representing the holon to the environment and
responsible for coordinating the actions inside the holon:
• S = s×n = {(s1 , s2 , . . . , sn ) |si ∈ s for all 1 ≤ i ≤ n} – is a joint state
of the holon;
• A = α×n = {( a1 , a2 , . . . , an ) |ai ∈ α for all 1 ≤ i ≤ n} is a joint
action of the holon;
• π : S → A – global;
• C – is the commitment that defines the agreement to be inside the holon.
The learning of multi-agent systems composed of homogenous agents has a
few important properties which affect the usage of such systems.

6.2.3 Approach to Commitment and Coordination in H2 MAS

The holon is realised exclusively through cooperation among the constituent
agents. The head agent is required to co-ordinate the work of the body agents
to achieve the desired global behavior of H2 MAS by combining individual
behaviors, resolving collisions, etc. In this way, a head agent serves as a
coordination strategy among agents. The head is aware of the goals of the
holon, and has access to important environmental information which allows
it to act as a central point of coordination for body agents.
Since a body agent has some degree of autonomy, it may perform an
unexpected action, which can lead to uncoordinated behavior within the
Holon. The head agent can observe the states and actions of all subordinate
agents and can fix undesired behavior using simple coordination rule: if the
current behavior of the holon Mi is inconsistent with the head agent’s vision,
then it sends a correction message to Mi . This action by the head is known as
an influence on the body. When the body Mi succumbs to the influence, this
is called making a commitment to the Holon.
 6.2 Holonic Homogenous Multi-Agent Systems 129

6.2.4 Learning to Coordinate Through Interaction

The basic idea of the selected approach for coordination is to use influences
between the head and the body to determine the sequence of correct actions
to coordinate behavior within the holon. The core design question is how to
determine such influences in terms of received messages and how received
messages affect changes of individual policies.
To answer this question we postulate that interacting agents should
constantly learn optimal coordination from scratch. To achieve this, we
can use influence-based, multi-agent reinforcement learning [18–20]. In
this approach, agents learn to coordinate using reinforcement learning by
exchanging rewards with each other.
In reinforcement learning, the ith agent executes an action ai at the current
state si . It then goes to the next state si and receives a numerical reward r
as feedback for the recent action [21], where si , si ∈ s, ai ∈ α, r ∈ R.
Ideally, agents should explore state space (interact with environment) to build
an optimal policy π ∗ .
Let Q (s, a) – represent a Q-function that reflects the quality of the
specified action a in state s. Optimal policy can be expressed in terms of
optimal Q-function Q∗ :
π ∗ (s) = arg max Q∗ (s, a) . (6.1)
a∈α(s)

The initial values of Q-funcions are unknown and equal to zero. The learning
goal is to approximate the Q-function, (e.g. to find true Q-values for each
action in every state using received sequences of rewards).
A model of influence-based multi-agent reinforcement learning depicted
in Figure 6.3.
In this model, a set of body agents with identical policies π acts in
a common, shared environment. The ith body agent Mi in the state si
selects an action ai using current policy π, and then moves to the next
state si . The head agent observes changes resulting from the executed
action and then calculates and assigns a ri to the agent as an evaluative
feedback.
Equation (6.2) is a variation of the Q-learning update rule [21] used to
update the values of the Q-function, and where learning homogeneity and
parallelism are applied. Learning homogeneity refers to all agents building
the same Q-function, and parallelism requires that they can do it in parallel.
The following learning rule executes N times per step for each agent in parallel
over single-shared Q-function:
130 A Multi-Agent Reinforcement Learning Approach for the Efficient Control

Figure 6.3 Model of Influence Based Multi-Agent Reinforcement Learning in the Case of a
Holonic Homogenous Multi-Agent System.

 

ΔQ (si , ai ) = α ri + γ max Q (si , a) − Q (si , ai ) . (6.2)
a∈α(si )

6.3 Vehicle Steering Module

The platform is based on four vehicle steering modules. The steering module
consists of two wheels powered by separate motors and behaves as a differ-
ential drive. It is mounted to the platform by a bearing that allows unlimited
rotation of the module with respect to the platform (Figure 6.4). The platform
may be equipped with three or more modules.
The conventional approach for the platform control is a kinematics
calculation and an inverse kinematics modeling [3]. The inverse kinematics
calculation is known for the common schemes: the differential scheme, car
scheme, and bicycle scheme. In the case of production module platforms,
the four modules are controlled independently. As a consequence, the con-
trol system can only perform symmetric turning. Hence, the platform has
limited maneuverability [3]. The other problem is the limitations of the robot
 6.3 Vehicle Steering Module 131

Figure 6.4 The Maneuverability of one module.

configuration. Previous systems require recalculations if modules are added or

removed from the platform. These recalculations require a qualified engineer.
The problem of steering the robot along the trajectory is illustrated in
Figure 6.5. This trajectory consists of four segments:

Figure 6.5 Mobile Robot Trajectory Decomposition.

132 A Multi-Agent Reinforcement Learning Approach for the Efficient Control

• The turning radius length is R1 , the center of rotation is (x1 ,y1 );

• The straight segment;
• The turning radius length is R2 , the center of rotation is (x2 ,y2 );
• The straight segment.
The steering of the robot also fulfills the following specifications:
• At the starting point, the robot rotates all modules in the direction of the
trajectory;
• A robot cannot stop at any point of the trajectory. The trajectory always
has smooth transitions from one segment to another.

6.4 A Decomposition of Mobile Platform

A platform is composed of identical modules attached to the platform in the
same way as a multi-agent decomposition. This is a prominent way to develop
a distributed control strategy for such platforms. Mobile platforms with four
identical independent driving modules can be represented as homogenous,
holonic, multi-agent systems as described in Section 6.2. The driving modules
are represented as body agents (or module agents) and the head agent (or
platform agent) represents the whole platform. The process of multi-agent
decomposition described above is shown in Figure 6.6.
The whole platform reflects global information, such as the shape of the
platform and the required module topology, including its desired positions
relative to the centroid of the platform. To highlight this information, we can
attach a virtual coordinate frame to the centroid of the platform to create the
virtual structure.
Figure 6.7 shows an illustrative example of the virtual structure approach
with a formation composed of four vehicles capable of planar motions, where

Figure 6.6 Holonic Decomposition of the Mobile Platform. Dashed lines represent the
boundary of a Multi-Agent System (the Holon). Introduction of the Head Agent Leads to
a reduction of communication costs.
 6.4 A Decomposition of Mobile Platform 133

Figure 6.7 Virtual Structure with a Virtual Coordinate Frame composed of Four Modules
with a known Virtual Center.

C0 represents the beacon frame and C represents a virtual coordinate frame

located at a virtual center (xvc , yvc ) with an orientation ϕvc relative to C0 .
Values of ρi = [xi , yi ]T and ρdi = xdi , yid represent, respectively, the -th
vehicle’s actual and desired position. Values of ϕi and ϕdi represent the actual
and desired
 dorientation,
 respectively, of the -th vehicle. Each module’s desired
d d
position xi , yi , ϕi can be defined relative to the virtual coordinate frame.
For a formation stabilization with a static formation centroid, if each vehi-
cle in a group can reach a consensus on the center point of the desired formation
and specify a corresponding desired deviation from the center point, then the
desired motion can be achieved [22]. If each vehicle can track its desired posi-
tion accurately, then the desired formation shape can be preserved
 accurately.
The vectors di = (xc − xi , yc − yi ) and diC = xc − xdi , xc − yid
represent, respectively, the -th vehicle’s desired and actual deviation relative to
C. The deviation vector derr
i of the i-th module relative to the desired position
is defined as:
134 A Multi-Agent Reinforcement Learning Approach for the Efficient Control

derr
i = di − diC . (6.3)
Each module’s desired position can be defined relative to the virtual coordinate
frame. Once the desired dynamics of the virtual structure are defined, the
desired motion for each agent can be derived. As a result, path planning
and trajectory generation techniques can be employed for the centroid
while trajectory tracking strategies can be automatically derived for each
module [23].

6.5 The Robot Control System Learning

The main goal of the control system is to provide the movement of the robot
along the desired circular trajectory. The objective is to create a cooperative
control strategy for any configuration of N modules so that all the modules
within the platform achieve circular motion around the beacon. The circular
motions should have a prescribed radius of rotation ρC defined by the center
of the platform and the distance between neighbors. Further requirements are
that module positioning before movement must be taken into account, and
the adaptation of angular and linear speed during circular movement to reach
optimal values.
We divide the process of learning into two steps:
• Module positioning – a learning of the module to rotate to the trajectory
direction (6.5.1);
• Cooperative movement - a learning of cooperative motion of modules
within platform (6.5.2).
The overall control architecture is depicted in Figure 6.8.
From this decomposition, every module agent will have two control
policies, πv and πω , for both forward and angular velocity, respectively.
Policy πω is responsible for correct module orientation around the beacon.
Each module follows this policy before the platform starts moving. Policy πv
is used during circular motion of the platform along curves. Both policies are
created via reinforcement learning, which allows for generalization.
In the simulation phase, the head agent interacts with the modeling
environment. In experiments with real robots, the head agent interacts with
the planning subsystem. The Environment/Planning subsystem provides infor-
mation about the desired speed of the platform v d and the global state of the
multi-agent S = Ui=1 N s ∪ {s }, where s ∈ s is the state of the i-th module
i h i
 6.5 The Robot Control System Learning 135

Figure 6.8 A unified view of the control architecture for a Mobile Platform.

 N  d d d
is defined by values in Table 6.1 and sh = C, C0 , Ui=1 xi , yi , ϕi is
the state of the head agent describes the virtual coordinate frame.

6.5.1 Learning of the Turning of a Module-Agent

This subsection describes the model for producing an efficient control rule for
the positioning of a module, based on the relative position of the module with
respect to the beacon. This control rule can be used for every module, since
every steering module agent is homogenous.
The agent stays in a physical, 2-D environment with a reference beacon,
as shown in Figure 6.9. The beacon position is defined by coordinates (x0 ,
y0 ). The rotation radius ρ is the distance from the center of the module to the
beacon.
136 A Multi-Agent Reinforcement Learning Approach for the Efficient Control

Figure 6.9 State of the Module with Respect to Reference Beacon.

The angle error is calculated using the following equations:

φc = arctan 2(x0 − xi , y0 − yi ), (6.4)

ϕerr = ϕdi − ϕi . (6.5)

Here, ϕdi and ϕi are known from the environment.
In the simulated model environment, all necessary information about an
agent and a beacon is provided. In a real robotic environment, this information
is taken from wheel odometers and a module angle sensor. The environment
information states are illustrated in Table 6.1.
The full set of actions available to the agent is presented in Table 6.2.
The agent with actions A = {Aω , Av } can change an angle speed by actions
Aω = {ω+ , ω− } and linear speed by actions Av = {v+ , v− }. To turn, an
agent controls the angular speed Aω .

Table 6.1 The Environment Information

No Robot Get Value
1 X robot position, x Coordinate, m
2 Y robot position, y Coordinate, m
3 X of beacon center, xb Coordinate, m
4 Y of beacon center, yb Coordinate, m
5 Robot orientation angle, ϕi Float number, radians
−π < ϕi ≤ π
6 Desired orientation angle relative to robot, ϕdi Float number, radians
−π < ϕdi ≤ π
7 The radius size, r Float number, m
8 The desired radius size, rd Float number, m
 6.5 The Robot Control System Learning 137

Table 6.2 Agent Actions

No Robot Action Value
1 Increase force, v− +0.1, m/s
2 Reduce force, v+ –0.1, m/s
3 Increase turning left, ω+ +0.1, rad/s
4 Increase turning right, ω− –0.1, rad/s
5 Do nothing, ∅ +0 m/s, +0 rad/s

The learning system is given a positive reward when the robot orientation
is closer to the goal orientation (ϕerr → 0) and is using optimal speed ωopt .
A penalty is received when the orientation of the robot deviates from the goal
orientation or the selected action is not optimal for the given position. The
value of the reward is defined as:
rω = Rω (φterr , ω t ). (6.6)
Where Rω – is a reward function, which is represented by the decision tree
depicted in Figure 6.10. Here, ϕstop represents the value of the angle, where

Figure 6.10 A Decision tree of the reward function.

138 A Multi-Agent Reinforcement Learning Approach for the Efficient Control

the robot reduces speed to stop at the correct orientation, ωopt [0.6 .. 0.8]
rad/s, which is the optimal speed to minimize module power consumption.
The parameter ϕstop is used to decrease the search space for the agent. When
the agent angle error becomes smaller than ϕstop , an action that reduces the
speed will receive the highest reward. The parameter ωopt shows the possibility
of power optimization by setting a value function. If the agent angle error is
more than ϕstop and ωoptmin < ω < ω max , then the agent reward will increase.
opt
This coefficient which determines the increase ranges between [0 .. 1]. The
optimization allows the use of the preferred speed with the lowest power
consumption.

6.5.1.1 Simulation
The first task of the robot control is becoming familiar with robot positioning
through simulation. This step is done once for an individual module before
any cooperative simulation sessions. The learned policy is stored and copied
for other modules via knowledge transfer. The topology of the Q-function
trained during 720 epochs is shown in Figure 6.11.

Figure 6.11 Result Topology of the Q-Function.

 6.5 The Robot Control System Learning 139

The external parameters of a simulation are:

• Learning rate α = 0, 4;
• Discount factor γ = 0, 7;
• min = 0, 6 rad/s;
Minimal optimal speed ωopt
• Maximum optimal speed ωopt max = 0, 8 rad/s;

• Stop angle, ϕstop = 0, 16 radians.

Figure 6.12 shows the platform’s initial state (left) and the positioning
auto-adjustment (right) using learned policy [23].

6.5.1.2 Verification
The learning of the agent was executed on the real robot after a simulation with
the same external parameters. The learning process took 1440 iterations. A real
learning process takes more iterations on average because the real system has
noise and sensor errors. Figure 6.13 illustrates the result of execution of a
studied control system used to turn modules to the center, which is on the rear
right side of the images [24].

Figure 6.12 Initial and Final Agente Positions.

140 A Multi-Agent Reinforcement Learning Approach for the Efficient Control

Figure 6.13 Execution of a Learned Control System to turn modules to the center, which is
placed on the rear right relative to the platform.

6.5.2 Learning of the Turning of a Module-Agent

This subsection describes multi-agent learning for producing an efficient
control law in the case of cooperative motion using an individual module’s
speed. The module’s desired linear speed Av should be derived through the
learning process relative to the head agent so that the whole platform is moved
in a circular motion.
Let the state of the module be represented by st = {vt , derri }, where vt
err
is the current value of linear speed, and di is the error vector calculated by
(6.7). Action set Av = {Ø, v+ , v− } is represented by the increasing/decreasing
of the linear speed from Table 6.2 and action at ∈ Aν is a change of forward
speed Δν t for given moment in time t.
The virtual agent receives error information for each module and calculates
the displacement error. This error can be positive (module ahead of the
platform) or negative (module behind of the platform). The learning process
progresses toward the minimization of error derr i for every module. The
err
maximum reward is given for the case where di → 0, and a penalty is
given when the position of the module deviates from the predefined position.
The value of the reward is defined as:
⎧
⎨ 1, di err > dti err

⎪

rv = − 1, di err < dti err .

(6.7)
⎪
⎩ 10, dt
i err = 0

6.5.2.1 Simulation
Figure 6.14 shows the experimental results of the cooperative movement after
learning positioning [23]. It takes 11000 epochs on average. The external
parameters of a simulation are:
• Learning rate α = 0,4;
• Discount factor γ = 0,7.
 6.5 The Robot Control System Learning 141

Figure 6.14 Agents Team Driving Process.

During the modules learning, the control system did not use any stabilization
of the driving direction. This is because a virtual environment has an ideal,
flat surface. In the case of the real platform, stabilization will be provided
by internal controllers of the low-level module software. This allows us to
consider only the linear speed control.

6.5.2.2 Verification
The knowledge base of the learned agents was transferred to the agents of
the control system on the real robot. Figure 6.15 demonstrates the process of
the platform moving by the learned system [25]. At first, modules turn in the
driving direction relative to the center of rotation (the circle drawn on white
paper), as shown in screenshots 1–6 in Figure 6.15. Then, the platform starts
driving around the center of rotation in screenshots 7–9 in Figure 6.15. The
stabilization of the real module orientation is based on a low-level controller
with feedback. This controller is provided by the software control system of
the robot. It helps to restrict the intellectual control system by manipulating
the linear speed of modules.
142 A Multi-Agent Reinforcement Learning Approach for the Efficient Control

Figure 6.15 The Experiment of modules turning as in the Car Kinematics Scheme (1–6
screenshots) and movement around a White Beacon (7–9).

Figure 6.16 The Experiment shows that the radius doesn’t change during movement.

The distance to the center of rotation is always the same on the entire trajec-
tory of the platform. This is confirmed by Figure 6.16. Hence, the robot drives
around in a circle where the coordinates of the center and the radius are known.

6.6 Conclusions
This paper focuses on an efficient, flexible, adaptive architecture for the
control of a multi-wheeled, production, mobile robot. The system is based
on a decomposition into a holonic, homogenous, multi-agent system and on
influence-based, multi-agent reinforcement learning.
 References 143

The proposed approach incorporates multiple Q-learning agents, which

permits them to effectively control every module relative to the platform. The
learning process was divided into two parts:
• Module positioning – where agents learn to minimize the error of
orientation;
• Cooperative movement – where agents learn to adjust the desired velocity
to conform to a desired position in formation.
A head agent is used to coordinate modules through the second step of learning.
From this decomposition, every module agent will have a separate control
policy for both forward and angular velocity.
The reward functions are designed to produce efficient control. During
learning, agents take into account the current reward value and the previous
reward value that helps to find the best policy of agent actions. Altogether,
this provides efficient control where agents must cooperate with each other
and use the policy of least resistance between each other on a real platform.
The advantages of this method are as follows:
• Decomposition means that instead of trying to build a global Q-function,
we can build a set of local Q-functions;
• Adaptability – the platform will adapt its behavior for a dynamically
assigned beacon and will auto-reconfigure its moving trajectory;
• Scalability and generalization – the same learning technique is used
for every agent, for every beacon position, and for every platform
configuration.
In this chapter, we showed successful experiments with the real robot where
the system provides robust steering of the platform. These results indicate that
the application of intellectual adaptive control systems for real mobile robots
have great potential in production.
In future works, we will consider a comparison of the developed approach
to mobile robot steering with existing approaches and will provide further
information about efficiency of the developed control systems relative to real
control systems.

References
[1] J. C. Andreas, ‘Energy-Efficient ElectricMotors, Revised and
Expanded’, CRC Press, 1992.
[2] A. T. de Almeida, P. Bertoldi and W. Leonhard, ‘Energy efficiency
improvements in electric motors and drives’, Springer Berlin, 1997.
144 A Multi-Agent Reinforcement Learning Approach for the Efficient Control

[3] R. Stetter, P. Ziemniak and A. Paczynski, ‘Development, Realiza-

tion and Control of a Mobile Robot’, In Research and Education in
Robotics-EUROBOT 2010, Springer, 2011:130–140.
[4] U. Dziomin, A. Kabysh, V. Golovko and R. Stetter, ‘A multi-agent
reinforcement learning approach for the efficient control of mobile
robot’, In Intelligent Data Acquisition and Advanced Computing Sys-
tems (IDAACS), 2013 IEEE 7th International Conference on, 2,
2013:867–873.
[5] Y. Mei, Y.-H. Lu, Y. C. Hu, and C. G. Lee, ‘Energy-efficient motion plan-
ning for mobile robots’, In Robotics andAutomation, 2004. Proceedings.
ICRA’04. 2004 IEEE International Conference on, 5, 2004:4344–4349.
[6] S. Ogunniyi and M. S. Tsoeu, ‘Q-learning based energy efficient path
planning using weights’, In proceedings of the 24th symposium of the
Pattern Recognition association of South Africa, 2013:76–82.
[7] Y. Mei, Y.-H. Lu, C. G. Lee and Y. C. Hu, ‘Energy-efficient
mobile robot exploration’, In Robotics and Automation, 2006. ICRA
2006. Proceedings 2006 IEEE International Conference on, 2006:
505–511.
[8] N. Ceccarelli, M. Di Marco, A. Garulli and A. Giannitrapani, ‘Collective
circular motion of multi-vehicle systems with sensory limitations’, In
Decision and Control, 2005 and 2005 European Control Conference.
CDC-ECC’05. 44th IEEE Conference on, 2005:740–745.
[9] N. Ceccarelli, M. Di Marco, A. Garulli, and A. Giannitrapani, ‘Col-
lective circular motion of multi-vehicle systems’, Automatica, 44(12):
3025–3035, 2008.
[10] D. Benedettelli, N. Ceccarelli, A. Garulli and A. Giannitrapani, ‘Experi-
mental validation of collective circular motion for nonholonomic multi-
vehicle systems’, Robotics andAutonomous Systems, 58(8):1028–1036,
2010.
[11] K. Fischer, M. Schillo and J. Siekmann, ‘Holonic multiagent sys-
tems: A foundation for the organisation of multiagent systems’, In
Holonic and Multi-Agent Systems for Manufacturing., Springer, 2003:
71–80.
[12] N. Vlassis, ‘A concise introduction to multiagent systems and distributed
artificial intelligence’, Synthesis Lectures on Artificial Intelligence and
Machine Learning, 1(1):1–71, 2007.
[13] L. Gasser, ‘Social conceptions of knowledge and action: DAI foun-
dations and open systems semantics’, Artificial intelligence, 47(1):
107–138, 1991.
 References 145

[14] C. Gerber, J. Siekmann and G. Vierke, ‘Flexible autonomy in holonic

agent systems’, In Proceedings of the 1999 AAAI Spring Symposium
on Agents with Adjustable Autonomy, 1999.
[15] C. Gerber, J. Siekmann and G. Vierke, ‘Holonic multi-agent systems’,
Tech. rep. DFKI Deutsches Forschungszentrum fr Knstliche Intelligenz,
Postfach 151141, 66041 Saarb.
[16] C. Castelfranchi, ‘Commitments: From Individual Intentions to Groups
and Organizations’, In ICMAS, 95, 1995:41–48.
[17] P. Stone and M. Veloso, ‘Multiagent Systems: A Survey from a
Machine Learning Perspective’, Autonomous Robots, 8(3):345–383,
2000. [Online]. http://dx.doi.org/10.1023/A%3A1008942012299
[18] A. Kabysh and V. Golovko, ‘General model for organizing interactions
in multi-agent systems’, International Journal of Computing, 11(3):
224–233, 2012.
[19] A. Kabysh, V. Golovko and A. Lipnickas, ‘Influence Learning for
Multi-Agent Systems Based on Reinforcement Learning’, International
Journal of Computing, 11(1):39–44, 2012.
[20] A. Kabysh, V. Golovko and K. Madani, ‘Influence model and rein-
forcement learning for multi agent coordination’, Journal of Qafqaz
University, Mathematics and Computer Science, 33:58–64, 2012.
[21] A. G. Barto, ‘Reinforcement learning: An introduction’, MIT press,
1998.
[22] W. Ren and N. Sorensen, ‘Distributed coordination architecture for
multi-robot formation control’, Robotics and Autonomous Systems,
56(4):324–333, 2008.
[23] [Online]. https://www.youtube.com/watch?v=MSweNcIOJYg
[24] [Online]. http://youtu.be/RCO-j32-ryg
[25] [Online]. http://youtu.be/pwgmdAfGb40
 7
Underwater Robot Intelligent Control Based
on Multilayer Neural Network

D. A. Oskin1 , A. A. Dyda1 , S. Longhi2 and A. Monteriù2

1 Department of Information Control Systems, Far Eastern Federal University

Vladivostok, Russia
2 Dipartimento di Ingegneria dell’Informazione, Università Politecnica delle

Marche, Ancona, Italy

Corresponding author: D. A. Oskin <[email protected]>

Abstract
The chapter is devoted to the design of an intelligent neural network-based
control system for underwater robots. A new algorithm for intelligent con-
troller learning is derived using the speed gradient method. The proposed
systems provide robot dynamics close to the reference ones. Simulation
results of neural network control systems for underwater robot dynamics with
parameter and partial structural uncertainty have confirmed the perspectives
and effectiveness of the developed approach.

Keywords: Underwater robot, control, uncertain dynamics, multilayer

neural network speed gradient method.

7.1 Introduction
Underwater Robots (URs) promise great perspectives and have a broad
scope of applications in the area of ocean exploration and exploitation. To
provide exact movement along a prescribed space trajectory, URs need a
high-quality control system. It is well known that URs can be considered as
multi-dimensional nonlinear and uncertain controllable objects. Hence, the
design procedure of URs control laws is a difficult and complex problem
[3, 8].

Advances in Intelligent Robotics and Collaborative Automation, 147–166.


c 2015 River Publishers. All rights reserved.
148 Underwater Robot Intelligent Control Based on Multilayer Neural Network

Modern control theory has derived a lot of methods and approaches to

solve appropriate synthesis problems such as nonlinear feedback lineariza-
tion, adaptive control, robust control, variable structure systems, etc [1, 4].
However, most of these methods for control systems synthesis essentially use
information about the structure of the URs mathematical model. The nature
of the interaction of a robot with the water environment is so complicated
that it is hard to get the exact equations of URs motion. A possible way to
overcome control laws synthesis problems can be found in the class of artificial
intelligence systems, in particular, based on multi-layer Neural Networks
(NNs) [1, 2, 5].
Recently, a lot of publications were devoted to the problems of NNs
identification and control, starting from the basic paper [5]. Many papers
are associated, in particular, with applications of NNs to the problems of URs
control [1, 2, 7].
Conventional applications of multi-layer NNs are based on preliminary
network learning. As a rule, this process is the minimization of the criterion
which expresses overall deviations of NN outputs from the desirable values,
with given NN inputs. The network learning results in NN weight coefficients
adjustment. Such an approach supposes the knowledge of teaching input-
output pairs [5, 7].
The feature of NNs application as a controller consists in the fact that
a desirable control signal is unknown in advance. The desired trajectory
(program signal) can be defined only for the whole control system [1, 2].
Thus, the multi-layer NNs application in control tasks demands a develop-
ment of approaches that take into account the dynamical nature of controllable
objects.
In this chapter, an intelligent NNs-based control system for URs is
designed. A new learning algorithm for an intelligent NN controller, which
uses the speed gradient method [4], is proposed. Numerical experiments with
control systems containing the proposed NN controller were carried out in
different scenarios: varying parameters and different expressions for viscous
torques and forces. Modeling results are given and discussed.
Note that the choice of a NN regulator is connected with the principal
orientation of the neural network approach to a priori uncertainty, which
characterizes any UR. In fact, matrices of inertia of the UR’s rigid body are not
exactly known, as well as the added water mass. Forces and torques of viscous
friction are unknown and uncertain functional structure parameters. Hence,
an UR can be considered as a controllable object with partial parameter and
structure uncertainties.
 7.3 Intelligent NN Controller and Learning Algorithm Derivation 149

7.2 Underwater Robot Model

The UR mathematical model traditionally consists of differential equations
describing its kinematics
q̇1 = J(q1 )q2 (7.1)
and its dynamics

D(q1 )q̇2 + B(q1 , q2 )q2 + G(q1 , q2 ) = U, (7.2)

where J (q1 ) is the kinematical matrix; q1 , q2 are the vectors of generalized

coordinates and body-fixed frame velocities of the UR; U is the control forces
and torques vector; D is the inertia matrix taking into account added masses of
water; B is the Coriolis – centripetal term matrix; G is the vector of generalized
gravity, buoyancy and nonlinear damping forces/torques [3].
The lack a priori knowledge of the mathematical structure and the param-
eters of the UR model matrices and the UR model vectors can be compensated
by an intensive experimental research. As a rule, this way is too expensive
and takes a long time. One alternative approach is connected with the usage
of the intelligent NN control.

7.3 Intelligent NN Controller and Learning Algorithm

Derivation
Our objective is to synthesize an underwater robot NN controller in order to
provide the UR movement along a prescribed trajectory qd1 (t) , qd2 (t).
Firstly, we consider the control task with respect to the velocities qd2 (t).
Let us define the error as:
e2 = qd2 − q2 (7.3)
and let’s introduce the function Q as a measure of the difference between
desired and real trajectories:
1
Q = eT2 De2 , (7.4)
2
where the matrix of inertia is D > 0.
Furthermore, we use the speed gradient method developed by A.
Fradkov [4]. According to this method, let compute the time derivative of Q:
1
Q̇ = eT2 Dė2 + eT2 Ḋe2 . (7.5)
2
150 Underwater Robot Intelligent Control Based on Multilayer Neural Network

From
q2 = qd2 − e2 (7.6)
and one has
D(q1 )q̇2 = D(q1 )q̇d2 − D(q1 )ė2 . (7.7)
Using the first term of the dynamics Equation (7.2), one can get the following:
D(q1 )ė2 = D(q1 )q̇d2 + B(q1 , q2 )qd2 −
− B(q1 , q2 )e2 + G(q1 , q2 ) − U, (7.8)
and thus the time derivative of function Q can be written in the following
form:
Q̇ = eT2 (D(q1 )q̇d2 + B(q1 , q2 )qd2 −
1
− B(q1 , q2 )e2 + G(q1 , q2 ) − U ) + eT2 Ḋe2 . (7.9)
2
After mathematical manipulation, one gets
Q̇ = eT2 (D(q1 )q̇d2 + B(q1 , q2 )qd2 + G(q1 , q2 ) − U )−
1
− eT2 B(q1 , q2 )e2 + eT2 Ḋ(q1 )e2 =
2
= eT2 (D(q1 )q̇d2 + B(q1 , q2 )qd2 + G(q1 , q2 ) − U )+
1
+ eT2 ( Ḋ(q1 ) − B(q1 , q2 )e2 ).
2
As known, there is a skew-symmetric matrix in the last term, hence, this term
is equal to zero, and we obtain the following simplified expression:
Q̇ = eT2 (D(q1 )q̇d2 + B(q1 , q2 )qd2 + G(q1 , q2 ) − U ). (7.10)
Our aim is to implement an intelligent UR control [1] based on neural
networks. Without loss of generality of the proposed approach, let’s choose
a two-layer NN (Figure 7.1). Let the hidden and output layers have H and
m neurons, respectively (m is equal to the dimension of e2 ). For the sake
of simplicity, one supposes that only the sum of weighted signals (without
nonlinear transformation) is realized in the neural network output layer. The
input vector has N coordinates.
Let’s define wij as the weight coefficient for the i-th input of the j-th neuron
of the hidden layer. So, these coefficients compose the following matrix
⎡ ⎤
w11 w12 ... w1N
⎢ w21 w22 ... w2N ⎥
w=⎢ ⎣ ...
⎥. (7.11)
... ... ... ⎦
wH1 wH2 ... wHN
 7.3 Intelligent NN Controller and Learning Algorithm Derivation 151

Figure 7.1 Neural network structure.

As a result of the nonlinear transformation f (w, x), the hidden layer output
vector can be written in the following form:
⎡ ⎤
f1 (w1T x)
f (w, x) = ⎣ ... ⎦, (7.12)
T
fH (wH x)
where wk denotes the k-th raw of matrix w and x is the NN input vector.
Analogously, let’s introduce the matrix W whose element Wli denotes the
transform (weight) coefficient from the i-th neuron of the hidden layer to the
l-th neuron of the output layer.
Once the NN parameters are defined, the underwater robot control signal
(NN output) is computed as follows:

U = y(W, w, x) = W · f (w, x). (7.13)

Substitution of this control into (7.10), allows us to get

Q̇ = eT2 (D(q1 )q̇d2 + B(q1 , q2 )qd2 +

+ G(q1 , q2 ) − W · f (w, x)). (7.14)

To derive the NN learning algorithm, we apply the speed gradient method

[4]. For this, we compute the partial derivatives of the time derivative of
152 Underwater Robot Intelligent Control Based on Multilayer Neural Network

function Q with respect to the adjustable NN parameters – matrices and W .

Direct differentiation gives
∂ Q̇
= −e2 f T (w, x). (7.15)
∂W
It is easy to demonstrate that if we choose all activation functions in the usual
form
f ( x ) = 1/(1 + e−τ x ), (7.16)
this implies the following property
∂
fi (wiT x) = fi (wiT x)[1 − fi (wiT x)]xj . (7.17)
∂wij
Let’s introduce the following additional functions
φi (wiT x) = fi (wiT x)[1 − fi (wiT x)] (7.18)
and the matrix
Φ(w, x) = diag(φ1 (w1T x)...φH (wH
T
x)). (7.19)
Hence, direct calculation gives
∂ Q̇
= −ΦW T e2 xT . (7.20)
∂w
As a final stage, one can write the NN learning algorithm in the following
form:
W (k+1) = W (k) + γe2 f T (w, x),
w(k+1) = w(k) + γΦW T e2 xT , (7.21)
where γ is the learning step, k is the number of iterations.
The continuous form of this learning algorithm can be presented as

Ẇ = γe2 f T (w.x),
(7.22)
ẇ = γΦW e2 xT (w.x).
Such an integral law of the NN-regulator learning algorithm may cause
unstable regimes in the control system, as it takes place in adaptive systems
[4]. The following robust form of the same algorithm is also used:

Ẇ = γe2 f T (w.x) − αW,

(7.23)
ẇ = γΦW e2 xT (w.x) − αw,
 7.4 Simulation Results of the Intelligent NN Controller 153

where constant α > 0.

Now, let’s consider which components should be included in the NN
input vector. The NN controller is oriented to compensate an influence of
the appropriate matrix and vector functions, and thus, in the most common
case, the NN input vector must be composed of q1 , q2 , e2 , qd2 and their time
derivative.
The NN learning procedure leads to the reduction of function Q, and thus,
in ideal conditions, the error e2 converges to zero and the UR follows the
desired trajectory
q2 (t) → qd2 (t). (7.24)
If the UR trajectory is given by qd1 (t), one can choose

qd2 (t) = J −1 (q1 )(q̇d1 (t) + k(qd1 (t) − q1 (t))), (7.25)

where k is a positive constant. From the kinematics Equation (7.1), it follows

that
q̇1 (t) → q̇d1 (t) + k(qd1 (t) − q1 (t)) (7.26)
and
ė1 (t) + ke1 (t) → 0, (7.27)
where
e1 (t) = qd1 (t) − q1 (t). (7.28)
Hence, the UR follows to the planned trajectory qd1 (t).

7.4 Simulation Results of the Intelligent NN Controller

In order to check the effectiveness of the proposed approach, different
computer simulations have been carried out. The UR model parameters were
taken from [6]. The UR parameters are the following:

D = DRB + DA ,

where the inertia matrix of the UR rigid body is

⎡ ⎤
1000 0 200
DRB = ⎣ 0 1000 0 ⎦,
200 0 11000
154 Underwater Robot Intelligent Control Based on Multilayer Neural Network

and the inertia matrix of the hydrodynamic added mass is

⎡ ⎤
1000 0 100
DA = ⎣ 0 1100 80 ⎦ .
100 80 9000

Matrices B and G are

⎡ ⎤
210 20 30
B = ⎣ 25 200 70 ⎦ ,
15 33 150
 T
G= 0 0 0 .
Vector q2 consists of the following components (linear and angular UR
velocities):
 T
q2 = vx vz ωy . (7.29)
The NN input is composed by q2 and e2 . The NN output (control forces
and torque) is the vector
 T
U= F x Fz M y . (7.30)
For the NN controller containing 10 neurons in the hidden layer, the
simulation results are given on Figures 7.2 – 7.9.
In the considered numerical experiments, the desired trajectory was taken
as follows:
⎧
⎨ vxd = 0.75m/ sec,
vzd = 0.5m/ sec, 0 ≤ t ≤ 250 sec ,
⎩
ωyd = −0.15rad/ sec,
⎧
⎨ vxd = 0.5m/ sec,
vzd = 0.75m/ sec, 250 ≤ t ≤ 500 sec
⎩
ωyd = 0.15rad/ sec .

7.5 Modification of NN Control

In previous sections, a NN control was designed. Practically speaking, the
synthesis procedure of the NN regulator does not use any information of
the mathematical model of the controlled object. As one can see, differential
 7.5 Modification of NN Control 155

Figure 7.2 Transient processes in NN control system (α = 0.01, γ = 250).

Figure 7.3 Forces and Torque in NN control system (α = 0.01, γ = 250).

156 Underwater Robot Intelligent Control Based on Multilayer Neural Network

Figure 7.4 Examples of hidden layer weight coefficients evolution (α = 0.01,γ = 250).

equations describing the underwater robot dynamics have a particular structure

which can be taken into account for solving the synthesis problem of the control
system.

Figure 7.5 Examples of output layer weight coefficients evolution (α = 0.01, γ = 250).
 7.5 Modification of NN Control 157

Figure 7.6 Transient processes in NN control system (α = 0.01, γ = 200).

Figure 7.7 Forces ant Torque in NN control system (α = 0.01, γ = 200).

158 Underwater Robot Intelligent Control Based on Multilayer Neural Network

Figure 7.8 Examples of hidden layer weight coefficients evolution (α = 0.01, γ = 200).

There exist different ways to solve it. One of the possible approaches is
derived below:

Figure 7.9 Examples of output layer weight coefficients evolution (α = 0.01, γ = 200).
 7.5 Modification of NN Control 159

As mentioned before, the parameters of underwater robots, such as added

masses, moments of inertia, coefficients of viscous friction etc, are not all
exactly known because of the complex hydrodynamic nature of the robot
movement in the water environment.
Let’s suppose that a set of nominal UR parameters can be estimated. Hence,
it is possible to get appropriate nominal matrices D0 (q1 ), B0 (q1 , q2 ) and
G0 (q1 , q2 ) in Equation (7.2). Let’s denote the deviations of the real matrices
from the nominal ones as ΔD (q1 ), ΔB (q1 , q2 ) and ΔG (q1 , q2 ), respectively.
So, the following takes place:
D(q1 ) = D0 (q1 ) + ΔD(q1 ),
B(q1 , q2 ) = B0 (q1 , q2 ) + ΔB(q1 , q2 ), (7.31)
G(q1 , q2 ) = G0 (q1 , q2 ) + ΔG(q1 , q2 ).
Inserting expressions (7.29) into Equation (7.10) gives
Q̇ = eT2 (D0 (q1 )q̇d2 + B0 (q1 , q2 )qd2 + G0 (q1 , q2 )+
+ ΔD(q1 )q̇d2 + ΔB(q1 , q2 )qd2 + ΔG(q1 , q2 ) − U ). (7.32)
Now let’s choose the control law in the form:
U = U0 + UN N , (7.33)
where
U0 = D0 (q1 )q̇d2 + B0 (q1 , q2 )qd2 + G0 (q1 , q2 ) + Γe2 , (7.34)
is the nominal control associated with the known part of the robot dynamics
(matrix Γ > 0 is positively definite) and UN N is the neural network control to
compensate the uncertainty. The scheme of the proposed NN control system
for an underwater robot is given on Figure 7.10.
If the robot dynamics can be exactly determined (and uncertainty does not
take place), the nominal control (7.34) fully compensates undesirable terms
in (7.32) (UN N can be taken as equal to zero) and one has
Q̇ = −eT2 Γe2 < 0. (7.35)
Thus, functions Q (t) and e2 (t) converge to zero for t → ∞.
In the general case, as follows from (7.32) – (7.34), one has
Q̇ = eT2 (ΔD(q1 )q̇d2 + ΔB(q1 , q2 )qd2 + ΔG(q1 , q2 ) − UN N ). (7.36)
As one can expect, the use of the nominal component of the control facilitates
the implementation of the proper NN control.
160 Underwater Robot Intelligent Control Based on Multilayer Neural Network

Figure 7.10 Scheme of the NN control system.

Further steps of the NN controller learning algorithm can be done

practically in the same manner as above (see Equation (7.15, 20 and 21)).
In order to check the derived NN control, mathematical simulations of the
UR control system were carried out. The nominal matrices D0 (q1 ), B0 (q1 , q2 )
and G0 (q1 , q2 ) were taken as follows:

D0 = DRB0 + DA0 ,
⎡ ⎤ ⎡ ⎤
1000 0 0 1000 0 0
DRB0 =⎣ 0 1000 0 ⎦ , DA0 = ⎣ 0 1100 0 ⎦,
0 0 11000 0 0 9000
⎡ ⎤
210 0 0
B0 = ⎣ 0 200 0 ⎦ ,
0 0 150
 T
G0 = 0 0 0 .
and matrix Γ = diag [0.02, 0.02, 0.02].
Note that the matrices D0 , B0 of the nominal dynamics model contain only
diagonal elements which are not equal to zero. This means that the nominal
model is simplified and does not take into account an interaction between
different control channels (of linear and angular velocities). The absence of
these terms in the nominal dynamics results in partial parametric and structural
uncertainty.
Figures 7.11 – 7.18 show the transient processes and control signals
(forces and torque) in the designed system with a modified NN regulator.
The experimental results demonstrated that the robot coordinates converge to
 7.5 Modification of NN Control 161

Figure 7.11 Transient processes with modified NN-control (α = 0, γ = 200).

Figure 7.12 Forces and torque with modified NN control (α = 0, γ = 200).

the desired trajectories. In comparison with the conventional multilayer NN

applications, the weight coefficients of the proposed NN controller are varying
simultaneously with the control processes.
162 Underwater Robot Intelligent Control Based on Multilayer Neural Network

Figure 7.13 Examples of hidden layer weight coefficients evolution (α = 0, γ = 200).

Figure 7.14 Examples of output layer weight coefficients evolution (α = 0, γ = 200).

 7.5 Modification of NN Control 163

Figure 7.15 Transient processes with modified NN control (α =0.001, γ = 200).

Figure 7.16 Forces and Torque with modified NN control (α = 0.001, γ = 200).
164 Underwater Robot Intelligent Control Based on Multilayer Neural Network

Figure 7.17 Examples of hidden layer weight coefficients evolution (α =0.001, γ =200).

Figure 7.18 Examples of output layer weight coefficients evolution (α =0.001, γ = 200).
 References 165

7.6 Conclusions
An approach on how to design an intelligent NN controller for underwater
robots and how to derive its learning algorithm on the basis of a speed
gradient method is proposed and studied in this chapter. The numerical
experiments have shown that high-quality processes can be achieved with the
proposed intelligent NN control. In the study case of producing an UR control
system, the NN learning procedure allows to overcome the parameter and
partial structural uncertainty of the dynamical object. The combination of
the neural network approach with the proposed control, designed using the
nominal model of the underwater robot dynamics, allows to simplify the
control system implementation and to improve the quality of the transient
processes.

Acknowledgement
The work of A.Dyda and D.Oskin was supported by Ministry of Science and
Education of Russian Federation, the State Contract No 02G25.31.0025.

References
[1] A. A. Dyda, ‘Adaptive and neural network control for complex dynamical
objects’, - Vladivostok, Dalnauka. 2007. pp. 149 (in Russian).
[2] A. A. Dyda, D. A. Oskin, ‘Neural network control system for underwater
robots.’ IFAC conference on Control Application in Marine Systems
“CAMS-2004”, - Ancona, Italy, 2004, pp. 427–432.
[3] T. I. Fossen, ‘Marine Control Systems: Guidance, Navigation and Con-
trol of Ships, Rigs and Underwater Vehicles’, Marine Cybernetics AS,
Trodheim, Norway, 2002.
[4] A. A. Fradkov, ‘Adaptive control in large-scale systems’, - M.: Nauka.,
1990, (in Russian).
[5] A. A. Narendra, K. Parthasaraty, ‘Identification and control of dynamical
systems using neural networks’, IEEE Identification and Control of
Dynamical System, Vol.1. No 1. 20, 1990, pp. 1475–1483.
[6] A. Ross, T. Fossen and A. Johansen, ‘Identification of underwater vehicle
hydrodynamic coefficients using free decay tests’, Preprints of Int. Conf.
CAMS-2004, Ancona, Italy, 2004. pp. 363–368.
166 Underwater Robot Intelligent Control Based on Multilayer Neural Network

[7] R. Sutton and A. A. Craven, ‘An on-line intelligent multi-input multi-

output autopilot design study’, Journal of Engineering for the Maritime
Environment, vol. 216. No. M2, 2002, pp. 117–131.
[8] J. Yuh, ‘Modeling and control of underwater robotic vehicles’,
Systems, Man and Cybernetics, IEEE Transactions on, vol. 20, no. 6,
pp. 1475–1483, Nov/Dec 1990. doi: 10.1109/21.61218
 8
Advanced Trends in Design of Slip
Displacement Sensors for Intelligent Robots

Y. P. Kondratenko1 and V. Y. Kondratenko2

1 Petro
Mohyla Black Sea State University, Ukraine
2 University
of Colorado Denver, USA
Corresponding author: Y. P. Kondratenko <y [email protected]>

Abstract
The paper discusses advanced trends in design of modern tactile sensors and
sensor systems for intelligent robots. The main focus is the detection of slip
displacement signals corresponding to object slippage between the fingers of
the robot’s gripper.
It provides information on three approaches for using slip displacement
signals, in particular, for the correction of the clamping force, the identification
of manipulated object mass and the correction of the robot control algorithm.
The study presents the analysis of different methods for the detection of slip
displacement signals, as well as new sensor schemes, mathematical models
and correction methods. Special attention is paid to investigations of sensors
developed by the authors with capacitive, magnetic sensitive elements and
automatic adjustment of clamping force. The new research results on the
determination of object slippage direction based on multi-component capacity
sensors are under consideration when the robot’s gripper collides with the
manipulated object.

Keywords: slip displacement, tactile sensor, gripper, intelligent robot, model,

information processing.

Advances in Intelligent Robotics and Collaborative Automation, 167–192.


c 2015 River Publishers. All rights reserved.
168 Advanced Trends in Design of Slip Displacement Sensors for Intelligent Robots

8.1 Introduction
Updated intelligent robots pose high-dynamic characteristics and effectively
function under a particular set of conditions. The robot control problem is more
complex in uncertain environments, as robots are usually lacking flexibility.
Supplying robots with effective sensor systems provides essential extensions
of their functional and technological feasibility [11]. For example, a robot may
often encounter a problem of gripping and holding i object doing manipulation
processes with the required clamping force Fi r avoiding its deformation or
mechanical injury, i = 1...n. To successfully solve the current tasks, the robots
should possess the capability to recognize the objects by means of their
own sensory systems. Besides, in some cases, the main parameter due to
which robot can distinguish objects of the same geometric shape is their mass
mi (i = 1...n). The robot sensor system should identify the mass mi of each
i-th manipulated object in order to identify a class (set) an object refers to.
The sensor system should develop the required clamping force Fi r corre-
sponding to mass value mi , as Fi r = f (mi ). Such current data may be applied
when the robot functions in dynamic or random environments. For example,
in a case when the robot should identify unknown parameters for any type of
object and location in robot’s working zone. The visual sensor system may
not always be utilized, in particular, in poor vision conditions. Furthermore,
in cases when the robot manipulates with an object of variable mass mi (t), its
sensor system should provide the appropriate change of clamping force value
Fi r (t) = f [mi (t)] for the gripper fingers. This information can also be used
for the robot control algorithm correction, since the mass of the robot arm’s
last component and its summary inertia moment vary.

8.2 Analysis of Robot Task Solving Based on Slip

Displacement Signals Detection
One of the current approaches to solving the mass mi identification problem
of grasped objects and producing the required clamping force Fi r is in
the development of tactile sensor systems based on object slippage regis-
tration [1, 11, 17, 18, 20, 22] while slipping between the gripper fingers
(Figure 8.1).
As usual, the slippage signal detection in robotic systems is accomplished
either in the trial motion or in the regime of continuous lifting of the robot
arm. In some cases, during the process of trial motions, it is necessary to make
a series of trial motions (Figure 8.2) for creating the required clamping forces
 8.2 Analysis of Robot Task Solving Based on Slip Displacement 169

Figure 8.1 Grasping and lifting an object with the robot’s arm: Initial positions of the gripper
fingers (1,2) and object (3) (a); Creating the required clamping force Fob by the gripper fingers
during object slippage in the lifting process (b).

Fob or Fie = kFob , where k is a coefficient which impacts the reliability of the
object motion (by the robot arm) in the realisation of the required path, k > 1.
According to Figure 8.2, the robot creates a minimal value of clamping
force Fmin in time moment t1 . Then step by step robot lifts an object vertical
distance Δl and the robot gripper increases the clamping force (F (t 1 ) + ΔF) if
a slip displacement signal appears. The grasping surface of the object is limited
by value lmax . The first series of trial motions is finished in time moment
t2 , when l = l max (Figure 8.1(b)). After that, the robot decompresses the
fingers (t 2 ...t 3 ), moves the gripper (t 3 ...t 4 ) to the initial position (Figure 8.1(a))
and creates (t 4 ...t 5 ) the initial value of the clamping force F(t5 ) = F (t2 ) +
ΔF = F1 for the beginning of a second stage or second series of trial motions.
Some sensor systems based on the slip displacement sensors were con-
sidered in [24, 25], but random robot environments very often requires the
development of new robot sensors and sensor systems for increasing the speed
of operations, the growth of positioning accuracy or the desired path-following
precision.
Thus, the task of the registration of slippage signals between the robot
fingers for manipulated objects is connected with: a) the necessity of the
required force creation being adequate to the object’s mass value; b) the
recognition of objects; c) robot control algorithm correction.
The idea of a trial motion regime comprises the process of an iterative
increase in the compressive force value if the slippage signal is being detected.
The continuous lifting regime provides the simultaneous object lifting process
and increasing clamping force until the slippage signal disappears. The choice
of the slip displacement data acquisition method depends on a robot’s purpose,
170 Advanced Trends in Design of Slip Displacement Sensors for Intelligent Robots

Figure 8.2 Series of trial motions with increasing clamping force F of gripper fingers based
on object slippage.

the salient features of its functioning medium, the requirements of its response
speed and the performance in terms of an error probability.
Figure 8.3 illustrates the main tasks in robotics which can be solved based
on slip displacement signal detection.
 8.2 Analysis of Robot Task Solving Based on Slip Displacement 171

Figure 8.3 The algorithm for solving different robot tasks based on slip signal detection.
172 Advanced Trends in Design of Slip Displacement Sensors for Intelligent Robots

8.3 Analysis of Methods for Slip Displacement Sensors

Design
Let’s consider the main methods of slip displacement data acquisition, in
particular [3, 4, 10, 11, 14, 17, 18, 20, 22]:
The vibration detection method. This method is based on a principle of
the vibration detection in the sensing element when the object is slipping.
To implement the method mentioned, the following sensing elements may be
adopted: a sapphire needle interacting with the crystal receiver or a rod with
a steel ball, connected to the electromagnetic vibrator.
The pressure re-distribution detection method. The method relies on the
detection of a distribution change in pressure between gripper fingers at object
slippage and is based on the physiological sensitivity function of human skin.
The pressure transducers serve as nerves and are surrounded by an elastic
substance as in the human body.
The rolling motion detection method. The method is characterized by
transducing the object displacements in the vertical direction when slipping
to the rolling motion of a sensitive element. A slip displacement signal is
detected through rolling of a cylinder roller with elastic covering and a large
friction coefficient. The roller’s rolling motions may be converted into an
electric signal by means of photoelectric or magnetic transducers, containing
a permanent magnet on a movable roller, and in the case of a magnetic head
being placed on a gripper.
The impact-sliding vibrations detection method. A core of the method
implies the detection of liquid impact-sliding vibrations when the object is
slipping. An acrylic disk with cylinder holes is used in the slip displacement
sensor, realizing the method under consideration. A rubber gasket made in
the form of a membrane protects one end of the disk, and a pressure gauge
is installed on the other end. The hole is filled with water so that its pressure
slightly exceeds the atmospheric pressure. While the motion of the object is
in contact with a membrane, the impact-sliding vibrations are appearing and,
therefore, inducing impulse changes on the water pressure imposed by a static
pressure.
The acceleration detection method. This method is based on the mea-
surement of accelerations of the sensing element motion by the absolute
acceleration signal separation. The slip displacement sensor, comprising two
accelerometers, can be used in this case. One of the accelerometers senses the
absolute acceleration in the gripper, another responds to the acceleration of the
sensitive plate springing when the detail is slipping. The sensor is attached to
 8.3 Analysis of Methods for Slip Displacement Sensors Design 173

the computer identifying the slip displacement signal by comparing the output
signals of both accelerometers.
The interference pattern change detection method. This method involves
the conversion of the intensity changes reflected from the moving surface of
the interference pattern. The intensity variation of the interference pattern is
converted to a numerical code, the auto-correlation function is computed and
it achieves its peak at the slip displacement disappearance.
The configuration change detection in the sensitive elements method.
The essence of the method incorporates the measurement of the varying
parameters when the elastic sensitive element configuration changes. The
sensitive elements made of conductive rubber afford coating of the object
surface protruding above the gripper before the trial motion. When the object is
displacing from the gripper, the configuration changes, the electrical resistance
of such sensitive elements changes accordingly, confirming the existence of
slippage.
The data acquisition by means of the photoelastic effect method. An
instance representing this method may be illustrated by a transducer, in
which, under the applied effort, the deformation of sensitive leather produces
the appearance of voltage in the photoelastic system. The object slippage
results in the change of the sensitive leather deformation being registered by
the electronic visual system. The photosensitive transducer is a device for the
transformation of interference patterns into the form of a numerical signal. The
obtained image is of binary character, each pixel gives one bit of information.
The binary representation of each pixel enables to reduce the processing time.
The data acquisition based on friction detection method. The method
ensures the detection of the moment when the friction between the gripper
fingers and the object to be grasped goes from friction at rest to dynamic
friction.
Method of fixing the sensitive elements on the object. The method is based
on fixing the sensitive elements on the surface of the manipulated objects
before the trial motions with the subsequent monitoring of their displacement
relative to the gripper at slipping.
Method based on recording oscillatory circuit parameters. The method is
based on a change in the oscillatory circuit inductance while the object slips.
The inductive slip sensor with a mobile core, stationary excitation winding
and solenoid winding being one of the oscillatory circuit branches implements
the method. The core may move due to the solenoid winding. The reduction
of the solenoid winding voltage indicates the process of lowering. The core
is lowering under its own weight from the gripper center onto the object
174 Advanced Trends in Design of Slip Displacement Sensors for Intelligent Robots

to be grasped. The oscillatory circuit induces the forced oscillations with

the frequency coinciding with the frequency of excitation in the excitation
winding.
The video signal detection method. The basis of this method constitutes a
change in detection and ranging of patterns or video pictures as an indication
of object slippage. The slip displacement detection is accomplished by means
of the location sensors or visual sensors based on a laser source that has either
a separated and reflecting beam or a vision with a non-coherent beam of light
conductors for picture lighting and a coherent beam for image transmission.
The choice of a slip displacement detection method involves the mul-
ticriterion approach taking into account the complexity of implementation,
the bounds of functional capabilities, mass values and overall dimensions,
reliability and cost.

8.4 Mathematical Model of Magnetic Slip Displacement

Sensor
8.4.1 SDS Based on “Permanent Magnet/Hall Sensor” Sensitive
Element and Its Mathematical Model
In this chapter, the authors consider a few instances of updating the measure-
ment systems. To suit the requirements of increasing the noise immunity of
the vibration measurement method, a modified method has been developed.
The modified method is founded on the measurement of the sensitive element
angular deviation occurring at the object slippage (Figure 8.4).
Let’s consider the structure and mathematical model (MM) of the SDS
developed by the authors with a magnetic sensitive element which can detect
the bar’s angular deviation appearing at object slippage (Figure 8.4). The
dependence U = f (α) can be used to determine the sensitivity of the SDS and
the minimal possible amplitudes of the robot trial motions.
To construct the mathematical models, consider a magnetic system com-
prising a prismatic magnet with dimensions c × d × l, which is set to
ferromagnetic plane with infinite permeability μ = ∞ (Figure 8.5), where:
c- width, d- length, and l- height of magnet, (d >> l). The point P(XP , YP ) is
the observation point, which is located on the vertical axis and can change its
position relative to the horizontal axis Ox or vertical axis Oy.
A Hall sensor with a linear static characteristic is located at the observation
point P. Let’s form the mathematical model for the determination of the
magnetic induction B and the output voltage Uout (P) of the Hall sensor in
 8.4 Mathematical Model of Magnetic Slip Displacement Sensor 175

Figure 8.4 Magnetic SDS: 1– Rod; 2– Head; 3– Permanent magnet; 4– Hall sensor.

Figure 8.5 Model of magnetic sensitive element.

relation to an arbitrary position of the observation point P under the surface

of the magnet.
The value of magnetic induction (outside the magnet volume) is B̄ = μ0 H̄,
where μ0 is a magnetic constant; H̄ is the magnetic field strength vector.
In the middle of the magnet– the magnetic induction value is determined
by the dependence B̄ = μ0 (J¯ + H̄), where J is a magnetization value.

J = J0 + χH,
176 Advanced Trends in Design of Slip Displacement Sensors for Intelligent Robots

where χ is a magnetic susceptibility; J0 is the residual magnetization

value.
The permanent magnet can be represented [26-28] as a simulation model
of the surface magnetic charges that are evenly distributed across the magnet
pole faces with the surface density JT .
Thus, a y- component of the magnetic field strength Hy of the magnetic
charges can be calculated as:
 
Xp +c/2 Xp −c/2
Hy = − J2πT arctg l−Y p
− arctg l−Yp −
  (8.1)
Xp +c/2 Xp −c/2
arctg −l−Y p
− arctg −l−Yp ,
and the y- component of magnetic induction By can be presented as:
 
Xp +c/2 Xp −c/2
By = − JT2πμ0 arctg l−Y p
− arctg l−Yp −
 
Xp +c/2 Xp −c/2
arctg −l−Yp − arctg −l−Yp ,
To determine the parameter
 JT , it is necessary to measure the induction

of the pole faces center By  . Value (y = l+) indicates that the mea-
x=0,y=l+
surement of Bmes is conducted outside the volume of the magnet. The value
of magnetic induction at a point with the same coordinates on the inside of
the pole faces can be considered equal to the value of induction from the
outside pole faces by virtue of the continuity of the magnetic flux and lines of
magnetic induction, namely:
 
 
By  = By  .
x=0,y=l+ x=0,y=l−
 
 
So, we can write: By  = Bmes = μ0 Jt + μ0 Hy  , where
x = 0, y = l+ x = 0,y = l−
Bmes is the value of magnetic induction measured at the geometric center of
the top pole faces of the prismatic magnet.
On the basis of (8.1), we obtain:
  
2 c/2 c/2
Bmes = μ0 JT 1 − lim 2π arctg l−Yp + arctg l+Yp
(Y p→l−)

1
= μ0 JT 2 + π1 arctg 4lc ,
2πBmes
JT = .
μ0 π + 2arctg 4lc
 8.4 Mathematical Model of Magnetic Slip Displacement Sensor 177

For the y- component of the magnetic induction By (P) at the observation

point P, the following expression was obtained:


Bmes Xp + c/2 Xp − c/2
By (p) = − c arctg − arctg −
π + 2arctg 4l l − Yp l − Yp

Xp + c/2 Xp − c/2
arctg − arctg . (8.2)
−l − Yp −l − Yp

8.4.2 Simulation Results

For the analysis of the existing mathematical model (8.2), let’s calculate the
value of magnetic induction on the surface of the magnet (Barium Ferrite)
with parameters of c=0.02m, d =0.08m, l=0.014m and a value of magnetic
induction Bmes = 40 mT (value measured at the geometric center of the upper
limit of the magnet).
The simulation results for the magnetic induction are represented as
By =fi (XP ),i = 1,2,3 above the magnet for different values of the height YP
of the observation point P (XP ,YP ) , where indicated:

f1 − f or 

By  ;
x∈[−20;20]mm,y=l+1mm
f2 − f or 

By  ;
x∈[−20;20]mm,y=l+5mm
f3 − f or 

By  .
x∈[−20;20]mm,y=l+20mm

As can be seen from Figure 8.6, the magnetic induction By = f1 (XP )

above the surface of the magnet is practically constant for the coordinate
XP ∈ [–5; 5] mm, which is half of the corresponding size of the magnet. If
the distance between the observation point P and magnet increases (f2 (XP ),
f3 (XP ) in Figure 8.6), the curve shape changes become more gentle, with
a pronounced peak above the geometric center of the top pole faces of the
prismatic magnet (at the point XP = 0).
For the Hall sensor (Figure 8.4) in the general case, the dependence of
the output voltage Uout (P) on the magnitude of the magnetic induction By is
defined as:
178 Advanced Trends in Design of Slip Displacement Sensors for Intelligent Robots

Figure 8.6 Simulation results for By (P) based on the mathematical model (8.2).

Uout (P) = Uc + kBy (P), (8.3)

where k is the correction factor, that depends on the type of Hall sensor;
Uc is a constant component of the Hall sensor output voltage.
For Hall sensors with linear dependence of the output signal Uout on the
magnetic induction By , k = const and k = f{By (P)} for nonlinear dependence
(8.3). For Hall sensor SS490 (Honeywell) with linear dependence (8.3), the
values of the parameters are: Uc = 2.5V and k = 0.032 (according to the static
characteristic of the Hall sensor). The authors present the mathematical model
of the Hall sensor output voltage Uout (YP ) at its vertical displacement above
the geometric center of the top pole faces of magnet (XP = 0):

0, 01
Uout (YP ) = 2, 5 + 7, 4 × 10−3 arctg +
0, 014 − YP
0, 01
arctg . (8.4)
0, 014 + YP
The comparative results for dependences Uout (Yp ), UE (Yp ) and UR (Yp )
are presented in Figure 8.7, where Uout (Yp ) was calculated using MM (8.4),
UE (Yp ) are the experimental results according to [7] and UR (Yp ) is a nonlinear
regressive model according to [8].
The comparative analysis (Figure 8.7) of the developed mathematical
model Uout (Yp ) with the experimental results UE (Yp ) confirms the correctness
and adequacy of the synthesized models (8.1)–(8.4).
 8.5 Advanced Approaches for Increasing the Efficiency 179

Figure 8.7 Comparative analysis of modeling and experimental results.

8.5 Advanced Approaches for Increasing the Efficiency

of Slip Displacement Sensors
This study presents a number of sensors for data acquisition in real time
[5, 11, 13, 14]. The need for rigid gripper orientation before a trial motion has
caused the development of the slip sensor based on a cylinder roller with a
load which has two degrees of freedom [3, 14].
The sensitive element of the new sensor developed by the authors has the
form of a ball (Figure 8.8) with light-reflecting sections disposed in a staggered
order (Figure 8.9), thus providing slippage detection by the photo-method.
The ball is arranged in the sensor’s space through spring-loaded slides.
Each slide is connected to the surface of the gripper’s space by an elastic
element made of conductive rubber.
The ball motion is secured by friction-wheels and is measured with the aid
of incremental transducers in another modification of the slip sensor with the
ball acting as a sensitive element. The ball contacts with the object through the
hole. In this case, the ball is located in the space of compressed air dispensed
through the hole.
For the detection of the sensitive element angular deviation during object
slippage in any direction, the authors propose [12] a slip displacement sensor
with a measurement of changeable capacitance (Figure 8.10).
180 Advanced Trends in Design of Slip Displacement Sensors for Intelligent Robots

Figure 8.8 The ball as sensitive element of SDS: 1– Finger of robot’s gripper; 2– Cavity for
SDS instalation; 3– Guides; 4– Sensitive element (a ball); 5– Spring; 6– Conductive rubber; 7,
8– Fiber optic light guides; 9– a Sleeve; 10– Light; 11– Photodetector; 13–Cover; 14–Screw;
15–Hole.

Figure 8.9 Light-reflecting surface of the sensitive ball with reflecting and absorbing
portions (12) for light signal.

The structure of the intelligent sensor system developed by the authors

which can detect the direction of object slippage based on the capacitated
SDS (Figure 8.10) is represented in Figure 8.11 with channels of information
processing and electronic units that implement the intellectual base of produc-
tion rules to identify the direction of displacement of the object in the gripper
(if there is a collision with an obstacle).
 8.5 Advanced Approaches for Increasing the Efficiency 181

Figure 8.10 Capacitated SDS for the detection of object slippage in different directions:
1– Main cavity of robot’s gripper; 2– Additional cavity; 3– Gripper’s finger; 4– Rod; 5– Tip;
6– Elastic working surface; 7– Spring; 8– Resilient element; 9, 10– Capacitor plates.

The SDS is placed on at least one of the gripper fingers (Figure 8.10). The
recording element consists of four capacitors distributed across the conical
surface of the additional cavity (2). One plate (9) of each capacitor is located
on the surface of the rod (4) and the second plate (10) is placed on the inner
surface of the cavity (2).
The intelligent sensor system (Figure 8.11) provides an identification of
signals corresponding to object slippage direction {N, NE, E, SE, S, SW, W,
NW} in the gripper in the cases of contacting obstacles.

Figure 8.11 Intelligent sensor system for identification of object slippage direction:
3– Gripper’s finger; 4– Rod; 9, 10– Capacitor plates; 11– Converter “capacitance-voltage”;
12– Delay element; 13, 18, 23– Adders; 14, 15, 21, 26– Threshold elements; 16– Multi-Inputs
element OR; 17– Computer information-control system; 19, 20, 24, 25– Channels for sensor
information processing; 22, 27– Elements NOT; 28–39– Elements AND.
182 Advanced Trends in Design of Slip Displacement Sensors for Intelligent Robots

The implementation of relation (Figure 8.10)

Lb 1
=
Ladd 5

allows to increase the sensitivity of developed sensor system.

Initially (Figure 8.10), the tip (5) is held above the surface of the gripper’s
finger (3) by a spring (7), and a resilient element (8) holds a rod (4) in such a
position that its longitudinal axis is perpendicular to the surface of the finger
and coincides with the axis MN of the rod (4). When gripping a manipulation
object, its surface comes in contact with the tip (5), the spring (7) is compressed
and the tip (5) is immersed in a cavity (2). At this moment, the value of
compressive force corresponds to the minimal pre-determined value F min
that eliminates distortion or damage of the object.
The object begins to slip in the gripper if an inequality between compres-
sive force and the mass of the object exists during the trial motion. In this case,
the rod (4) is deflecting along the sliding direction on the angle α as the result
of the friction forces between the object’s contacting surface and the working
surface (6) of the tip (5).
Thus, the longitudinal axis of the rod (4) coincides with the axis M N .
Reciprocal movement of plates (9) and (10) with respect to each other in
all capacitative elements leads to value changes of the capacities C 1 , C 2 ,
C 3 and C 4 depending on the direction of rod’s movement. The changes of
the capacities lead to the voltage changes at the outputs of the respective
convertors “capacitance-voltage” in all sensory processing channels.
From time to time, the robot’s gripper may be faced with an obstacle when
the robot moves a manipulation object in dynamic environments according
to preplanned trajectories. The obstacles can appear randomly in the dynamic
working area of the intelligent robot. As a result of the collision between the
robot gripper and the obstacle, object’s slippage may appear if the clamping
force F is not enough for reliable fixation of object between the gripper’s
fingers. The direction {N, NE, E, SE, S, SW, W, NW} of such slip displacement
of the object depends on the position of the obstacle in the desired trajectory.
In this case, the output signal of element OR (16) is equal to 1 and
this is a command signal for the computer system (17) which constrains the
implementation of the planned trajectory. At the same time, the logical 1 signal
appears at the output of one of the AND elements {29, 30, 31, 33, 34, 35, 37,
39} that corresponds to one of the object’s slippage direction {N, NE, E, SE,
S, SW, W, NW} in the robot gripper.
 8.5 Advanced Approaches for Increasing the Efficiency 183

Let’s consider, for example, the determination of slippage direction {N}

after the contact between the obstacle and robot gripper (with object). If the
object slips in direction {N} (Figure 8.11), then:
• the capacity C 1 in first channel (19) increases and a logical 1 signal
appears at the output of threshold element (14) and at the first input of
the AND element (28);
• the capacity C 3 in the third channel (20) decreases and a logical 1 signal
appears at the output of threshold element (15), at the second input and
output ofAND element (28) and at the first input of theAND element (29);
• the capacities C2 , C4 of the second (24) and fourth (25) channels of the
sensor information processing are equivalent C 2 = C 4 and in this case a
logical 0 signal appears at the outputs of adder (23) and threshold element
(26), a logical 1 signal appears at the output of NOT element (27), at the
second input and output of AND element (29) and at the second input of
the computer information-control system (17). It means that the direction
of the object’s slippage is {N}, taking into account that output signals of
the AND elements {30, 31, 33, 34, 35, 37, 39} equal 0.
The production rules “IF-THEN” base are represented in Table 8.1. This
rule base determines the functional dependence between the direction of
object slippage (Figure 8.11), the current state of each capacitor {C 1 , C 2 ,
C 3 , C 4 } and the corresponding output signals {U 1 , U 2 , U 3 , U 4 } of the
multi-capacitor slip displacement sensor, where: Ui , (i = 1...4)– output signal
of i-th converter “capacitance - voltage” (11); (>)– indicator of the corre-
sponding signal Ui ,(i = 1...4) increases during the object slippage process;
(<)– indicator of the corresponding signal Ui ,(i = 1...4) decreases during
the object slippage process; (=)– pair’s indicator of equivalence according to

Table 8.1 The base of production rules “IF-THEN” for indetification of the slip displacement
direction
Number of Antecedent Consequent
Production Rule U1 U2 U3 U4 Direction of Slippage
1 > = < = N
2 > > < < NE
3 = > = < E
4 < > > < SE
5 < = > = S
6 < < > > SW
7 = < = > W
8 > < < > NW
184 Advanced Trends in Design of Slip Displacement Sensors for Intelligent Robots

conditions Ui = Uj , (i = 1...4), ( j = 1...4), i = j in the antecedents of the

production rules.
The mathematical models of different types of slip displacement sensors
with a measurement of changeable capacity are presented in [5, 19, 21, 23].

8.6 Advances in Development of Smart Grippers for

Intelligent Robots
8.6.1 Self-Clamping Grippers of Intelligent Robots
The slip displacement signals, which are responsible for the creation of the
required compressive force adequate to the object mass, provide the condi-
tions for the correction of the gripper trajectory-planning algorithm, which
identifies an object mass as a variable parameter [9]. The object mass identi-
fication is carried out in response to the final value of the compressive force,
recorded at the slippage signal disappearance. It is of extreme importance to
employ slip sensors with uprated response when the object mass changes in
the functioning process.
In those cases when the main task of the sensing system is the com-
pression of the object without its deformation or damage, it is expedient in
future research to project the advanced grippers of a self-clamping design
(Figure 8.12), excluding the gripper drive for the compressive force growth
(at slippage) up to the required value.
The information processing system of a self-adjusting gripper of an
intelligent robot with angle movement of clamping rollers consists of:
24– control unit; 25– delay element; 26, 30, 36– adder; 27, 32, 37, 41– thresh-
old element; 28, 29, 31, 33, 34, 35, 38, 40, 42– switch; 39– voltage source.
In such a gripper (Figure 8.12), the rollers have two degrees of freedom
and during object slippage they have a compound behavior (rotation and
translation motions). This gripper (Figure 8.12) is adaptive since object self-
clamping is being accomplished with a force adequate to the object mass up
to the moment of the slippage disappearance [6, 14].
Another example of a developed self-clamping gripper [15] is represented
in Figure 8.13, where: 1– finger; 2– roller axis; 3– roller; 4– sector element;
5– guide gear racks; 6– pinion; 7– travel bar; 8, 9– axis; 10– object; 11–
elastic working surface; 12, 13– spring; 14, 15– clamping force sensor; 16–
electroconductive contacts; 17, 18– fixator; 19, 20– pintle.
 8.6 Advances in Development of Smart Grippers for Intelligent Robots 185

Figure 8.12 Self-adjusting gripper of an intelligent robot with angle movement of clamp-
ing rollers: 1, 2– Finger; 3, 4– Guide groove; 5, 6– Roller; 7, 8– Roller axis; 9, 15,
20– Spring; 10– Object; 11, 18– Elastic working surface; 12– Clamping force sensor; 13,
14– Electroconductive contacts; 16, 19– Fixator; 17– Stock; 21– Adjusting screw;
22– Deepening; 23– Finger’s drive.

The experimental self-clamping gripper with plane-parallel displacement

of the clamping rollers and intelligent robot with 4 degrees of freedom for
experimental investigations of the developed grippers and slip displacement
sensors are represented in Figure 8.14 and Figure 8.15, correspondingly.
186 Advanced Trends in Design of Slip Displacement Sensors for Intelligent Robots

Figure 8.13 Self-clamping gripper of an intelligent robot with plane-parallel displacement

of the clamping roller: Front view (a); Top view (b).

8.6.2 Slip Displacement Signal Processing in Real Time

Frequent handling operations require a compressive force being exerted
through the intermediary of the robot’s sensing system in the continuous
hoisting operation. This regime shows a simultaneous increase in the compres-
sive force while the continuous hoisting operation and lifting of the gripper
in the vertical direction take place accompanied by the slip displacement
signal measurement. When the slippage signal disappears, the compressive
force does not increase and, therefore, the operations with the object are
 8.6 Advances in Development of Smart Grippers for Intelligent Robots 187

Figure 8.14 Experimental self-clamping gripper with plane-parallel displacement of the

clamping rollers.

Figure 8.15 Intelligent robot with 4 degrees of freedom for experemental investigations
of SDS.

accomplished according to the robot’s operational algorithm. To realize the

trial motion regime and the continuous hoisting operation being controlled in
real time, stringent requirements to the parameters should be met, in particular:
• the response time between the moment of slippage emergence and the
moment when the gripper fingers begin to increase the compressive force;
188 Advanced Trends in Design of Slip Displacement Sensors for Intelligent Robots

• the time of the sliding process including the moments between the
emergence of sliding and its disappearance;
• the minimal object displacement detected by the slip signal.
The problem of raising the sensors response in measuring the slip displacement
signals is tackled by improving their configuration and using the measuring
circuit designs with high resolving power.

8.7 Conclusions
The slip displacement signal detection method considered in the present paper
furnishes an explanation of the main detection principles and allows robot
sensing systems to obtain broad capabilities. The authors developed a wide
variety of SDS schemes and mathematical models for capacitative, magnetic
and light-reflecting sensitive elements with improved characteristics (accuracy
time response, sensitivity). It is very important that the developed multi-
component capacity sensor allows identifying the direction of object slippage
based on slip displacement signal detection which can appear in the case
of intelligent robot collisions with any obstacle in a dynamic environment.
The design of self-clamping grippers is also a very appropriate direction for
intelligent robot development.
The results of the research are applicable in the automatic adjust-
ment of the clamping force of robot’s gripper and robot motion correction
algorithms in real time. The methods introduced by the authors may be
also used in random operational conditions, within problems of automatic
assembly, sorting, pattern and image recognition in the working zones of
robots. The proposed sensors and models can be used for the synthesis of
intelligent robot control systems [2, 16] with new features and for solving
orientation and control tasks during intelligent robot contacts with obstacles.

References
[1] Ravinder S. Dahiya, Giorgio Metta, Maurizio Valle and Giulio Sandini.
Tactile Sensing From Humans to Humanoids, volume 26 of Issue 1,
pages 1–20. IEEE Transactions on Robotics, 2010.
[2] A. A. Kargin. Introduction to Intelligent Machines. Book 1: Intelligent
Regulators. Nord-Press, DonNU, Donetsk, 2010.
 References 189

[3] Y. P. Kondratenko. Measurements methods for slip displacement signal

registration. In Proc. of Intern. Symposium on Measurement Technol-
ogy and Intelligent Instruments, pages 1451–1461, Chongqing-Wuhan,
China, 1993. Published by SPIE. USA.
[4] Y. P. Kondratenko and X. Y. Huang. Slip displacement sensors of robotic
assembly system. In Proc. of 10-th Intern. Conference on Assembly
Automation, pages 429–436, Kanazava, Japan, 23–25 Oct 1989. IFS
Publications. Kempston, United Kingdom.
[5] Y. P. Kondratenko and I. L. Nazarova. Mathematical model of capacity
sensor with conical configuration of sensitive element. In Proceedings of
the Donetsk National Technical University, No. 11(186), pages 186–191,
Donetsk: DNTU.
[6] Y. P. Kondratenko and E. A. Shvets. Adaptive gripper of intelligent robot.
Patent No. 14569, Ukraine, 2006.
[7] Y. P. Kondratenko and O. S. Shyshkin. Experimental studies of the
magnetic slip displacement sensors for adaptive robotic systems. In
Proceedings of the Odessa Polytechnic University, pages 47–51, Odessa,
2005. Special Issue.
[8] Y. P. Kondratenko and O. S. Shyshkin. Synthesis of regression models
of magnetic systems of slip displacement sensors. Radioelectronic and
Computer Systems, (No. 6(25)): 210–215, Kharkov, 2007.
[9] Y. P. Kondratenko, A. V. Kuzmichev and Y. Z. Yang. Robot control
system using slip displacement signal for algoritm correction. In
ROBOT CONTROL (SYROCO91). Selected papers from the 3-rd
IFAC/IFIP/IMACS Symposium, pages 463–469. Pergamon Press,
Vienna, Austria. Oxford-NewYork-Seoul-Tokyo, 1991.
[10] Y. P. Kondratenko, E. A. Shvets and O. S. Shyshkin. Modern sensor
systems of intelligent robots based on the slip displacement signal
detection. In Annals of DAAAM for 2007 & Proceedings of the 18th
International DAAAM Symposium, pages 381–382, Vienna, Austria,
2007. DAAAM International.
[11] Y. P. Kondratenko, L. P. Klymenko, V. Y. Kondratenko, G. V. Kondratenko
and E. A. Shvets. Slip displacement sensors for intelligent robots: Solu-
tions and models. In Proceedings of the 2013 IEEE 7th International
Conference on Intelligent Data Acquisition and Advanced Computing
Systems, Vol. 2, pages 861–866. IDAACS, 2013.
[12] Y. P. Kondratenko, N. Y. Kondratenko and V. Y. Kondratenko. Intelligent
sensor system. Patent No. 52080, Ukraine, 2010.
190 Advanced Trends in Design of Slip Displacement Sensors for Intelligent Robots

[13] Y. P. Kondratenko, O. S. Shyshkin and V. Y. Kondratenko. Device for

detection of slip displacement signal. Patent No. 79155, Ukraine, 2007.
[14] Y. P. Kondratenko, V. Y. Kondratenko, E. A. Shvets and O. S. Shyshkin.
Adaptive gripper devices for robotic systems. In Mechatronics
and Robotics (M&R-2007): Proceeding of Intern. Scientific-and-
Technological Congress (October 2–5, 2007), pages 99–105. Polytech-
nical University Press, Saint-Petersburg, 2008.
[15] Y. P. Kondratenko, V. Y. Kondratenko, I. V. Markovsky, S. K. Chernov,
E. A. Shvets and O. S. Shyshkin. Adaptive gripper of intelligent robot.
Patent No. 26252, Ukraine, 2007.
[16] Y. P. Kondratenko, Y. M. Zaporozhets, G. V. Kondratenko and
O. S. Shyshkin. Device for identification and analysis of tactile signals
for information-control system of the adaptive robot. Patent No. 40710,
Ukraine, 2009.
[17] M. H. Lee. Tactile sensing: New directions, new challenges. Int. J.
of Robotics Research. 19(7), Jul 2000, vol. 19 no. 7, pp. 636–643.,
Jul 2000.
[18] Mark H. Lee and Howard R. Nicholls. Tactile sensing for mechatronics -
a state of the art survey, volume 9, pages 1–31. Mechatronics, 1999.
[19] H. B. Muhammad, C. M. Oddo, L. Beccai, C. Recchiuto, C. J. Anthony,
M. J. Adams, M. C. Carrozza, D. W. L. Hukins and M. C. L. Ward.
Development of a bioinspired MEMS based capacitive tactile sensor for
a robotic finger. Sensors and Actuators A-165, pages 221–229, 2011.
[20] Howard R. Nicholls and Mark H. Lee. A survey of Robot Tactile Sensor
Technology. The International Journal of Robotic Research, (Vol. 8,
No. 3):3–30, June 1989.
[21] E. P. Reidemeister and L. K. Johnson. Capacitive acceleration sensor for
vehicle applications. In Sensors and actuators, pages 29–34. SP-1066,
1995.
[22] Johan Tegin and Jan Wikander. Tactile Sensing in Intelligent Robotic
Manipulation-A Review. Industrial Robot: An International Journal,
(Vol. 32, No. 1):64–70, 2005.
[23] M. I. Tiwana, A. Shashank, S. J. Redmond and N. H. Lovell. Character-
ization of a capacitive tactile shear sensor for application in robotic and
upper limb prostheses. Sensors and actuators, A-165, pages 164–172,
2011.
[24] M. Ueda and K. Iwata. Adaptive grasping operation of an industrial robot.
In Proc. of the 3-rd Int. Symp. Ind. Robots, pages 301–310, Zurich,
1973.
 References 191

[25] M. Ueda, K. Iwata and H. Shingu. Tactile sensors for an industrial robot
to detect a slip. In 2-nd Int. Symp. on Industrial Robots, pages 63–76,
Chicago, USA, 1972.
[26] Y. M. Zaporozhets. Qualitative analysis of the characteristics of direct
permanent magnets in magnetic systems with a gap. Technical electro-
dynamics, (No. 3):19–24, 1980.
[27] Y. M. Zaporozhets, Y. P. Kondratenko and O. S. Shyshkin. Three-
dimensional mathematical model for calculating the magnetic induction
in magnetic-sensitive system of slip displacement sensor. Technical
electrodynamics, (No. 5):76–79, 2008.
[28] Y. M. Zaporozhets, Y. P. Kondratenko and O. S. Shyshkin. Mathematical
model of slip displacement sensor with registration of transversal con-
stituents of magnetic field of sensing element. Technical electrodynamics,
(No. 4):67–72, 2012.
 9
Distributed Data Acquisition and Control
Systems for a Sized Autonomous Vehicle

T. Happek, U. Lang, T. Bockmeier, D. Neubauer and A. Kuznietsov

University of Applied Sciences, Friedberg, Germany

Corresponding author: T. Happek <[email protected]>

Abstract
In this paper, we present an autonomous car with distributed data processing.
The car is controlled by a multitude of independent sensors. For lane detection,
a camera is used, which detects the lane marks using a Hough transformation.
Once the camera detects these, one of them is selected to be followed by the
car. This lane is verified by the other sensors of the car. These sensors check
the route for obstructions or allow the car to scan a parking space and to park
on the roadside if the gap is large enough. The car is built on a scale of 1:10
and shows excellent results on a test track.

Keywords: Edge detection, image processing, microcontrollers, camera.

9.1 Introduction
In modern times, the question of safe traveling becomes more and more
important. Most accidents are caused by human failure, so that in many sectors
of industry the issue of “autonomous driving” is of increasing interest. An
autonomous car will not have problems like being in a bad shape that day
or tiredness and will suffer less from reduced visibility due to environmental
influences. A car with laser sensors to detect objects on the road, sensors that
measure the grip of the road, that calculate speed based on the signals of these
sensors and with a fixed reaction time will reduce the number of accidents and
related costs.

Advances in Intelligent Robotics and Collaborative Automation, 193–210.


c 2015 River Publishers. All rights reserved.
194 Distributed Data Acquisition and Control Systems

This chapter describes the project of an autonomous vehicle on a scale

of 1:10, which was developed based on decentralized signal routing. The
objective of the project is to build an autonomous car that is able to drive
autonomously on a scaled road, including the detection of stopping lines,
finding a parking area and parking autonomously.
The project is divided into three sections. The first section is about the car
itself, the platform of the project. This section describes the sensors of the car,
the schematic construction and the signal flow of the car.
The second section is about lane detection, the most important part
of the vehicle. Utilizing a camera with several image filters, the lane marks of
the road can be extracted from the camera image. This section also describes
the calculation of the driving lane for the car.
The control of the vehicle is the matter of the third section. The car runs
based on a mathematical model of the car, which calculates the speed and the
steering angle of the car in real time, based on the driving lane provided by
the camera.
The car is tested on a scaled indoor test track.

9.2 The Testing Environment

Since the car is based on a scaled model car, the test track has to be scaled,
too. Therefore, the test track has the same dimensions as the scaled road that
is used in a competition for cars on a scale of 1:10 that takes place in Germany
every year. As you can see in Figure 9.1, the road has fixed lane marks.
This is important, because it’s about a prototype. On a real road, several

Figure 9.1 Dimensions of the test track.

 9.3 Description of the System 195

types of lane marks exist. From white lines, as in the test track, to pil-
lars and a missing center line, every type of lane marks is expected to
show up.
In order to simplify the lane detection on the test track, it is assumed, that
at all times only one type of lane marks exists. The road has a fixed width and
the radius of the curves measure at least 1 meter. The test track has no slope,
but a flat surface.

9.3 Description of the System

The basic idea for the vehicle is related to a distributed data acquisition strategy.
That means, that all peripherals are not managed by a single microcontroller,
but each peripheral has its own microcontroller, which handles the data
processing for a specific task. All together are used to analyze the data of the
different sensors of the car. Smaller controllers, for instance for the distance
sensors or for the camera, are managed by one main controller. The input of
the smaller controllers is provided simultaneously via CAN.
The base of the vehicle is a model car scaled 1:10. It includes each
mechanical peripheral of a car, like the chassis or the engine. A platform
for the control system is added. Figure 9.2 shows the schematic topview of
the car with the added platform.
The vehicle itself is equipped with two front boards, a side board, a rear
board and the motherboard. All of these boards have one microcontroller for

Figure 9.2 Schematic base of the model car.

196 Distributed Data Acquisition and Control Systems

data analysis that is positioned next to the sensors. The front boards provide
infrared sensors for object tracking in the distance.
The infrared sensors of the side board have the task of finding a parking
space and transmitting the information via CAN bus to the main controller,
which undertakes the control of parking supported by the information it gets
from the sensors in the front and back of the car.
The rear board is equipped with infrared sensors, too. It serves the
back of the vehicle only. That guarantees a safe distance to all objects in
the back. The microcontrollers are responsible for the data processing of
each board and send the information to the main controller via CAN bus.
Each of the microcontrollers reacts on the incoming input signals of the
corresponding sensors according to its implemented control. The infrared
sensors are distributed alongside the car as you can see in Figure 9.3.
The motherboard with the integrated main controller is the main access
point of the car. It provides the CAN bus connection, the power supply for
the other boards and external sensors of the car. But primarily it’s the central
communications point of the car and manages the information that comes from
the peripheral boards, including the data from the external sensors, the control
signals for the engine and the servo for the starring angle.
The motherboard gets its power supply from three 5 V batteries. With these
three batteries, the model car is able to drive about one hour autonomously.
The main task for the main controller is the control of the vehicle. It
calculates the speed of the car and the starring angle based on a mathematical
model of the car and the information of the sensors. The external engine driver
sets the speed via PWM. The starring angle of the car is adjusted by the front
wheels. An additional servo controls the wheel’s angle.

Figure 9.3 The infrared sensors distributed alongside the car.

 9.3 Description of the System 197

The camera and its lane detection is the most important component of the
vehicle. It is installed in the middle of the front of the car, see Figure 9.4.
The viewing angle is important for the position of the camera. If the viewing
angle is too small, the pictures of the camera show a near area in front of the
car only, but not the area in the middle distance. If the viewing angle is too
big, the camera shows a big area in front of the car indicating near and far
distances, but the information of the road is so condensed, that an efficient lane
detection isn’t possible. The angle depends also on the height of the camera
and the numerical aperture of the lens. The higher the camera is positioned,
the smaller the viewing angle. For this project, the camera has a height of
30 cm and a viewing angle of 35 degrees. The height and the angle of the
camera are based on experimental research.
Figure 9.5 shows the reduced signal flow of the vehicle. The information
from the infrared sensors is sent to a small microcontroller, as it is visualized
by the spotted lines. In reality, each sensor has its own microcontroller, but to
reduce the complexity of the graphic, they were shown as one. The camera
has its own microcontroller. This controller must be able to accomplish
the necessary calculations for lane detection in time. For the control of the
vehicle by the main controller, it is necessary that all information from all
other controllers are actualized in one calculation step, this is needed for
the mathematical model of the car. The main controller gathers the analyzed
data provided by the smaller microcontrollers, the data from the camera

Figure 9.4 Position of the camera.

198 Distributed Data Acquisition and Control Systems

Figure 9.5 Schematic signal flow of the vehicle.

about the driving lane and the information from other sensors like gyroscope
and accelerometer for its calculation. The essential signal flow of all these
components to the main controller is visualized by the solid lines in Figure 9.5.
After its calculation, the main controller sends control signals to the engine
and the servo, which controls the starring angle of the car.
Incremental encoders on the rear wheels detect the actual speed and
calculate the path the vehicle has traveled during the last calculation step
of the mathematical model. The sensors send the data via CAN bus to the
main controller. The vehicle is front-engined, so traction of the rear wheels is
ensured. Potential error in measurement through spinning is avoided.
There are two modules that do not communicate via CAN bus with the
main controller: the first one is the camera, ensuring that the vehicle keeps
the track, the second is a sensor module, which includes the gyroscope
and accelerometer. Both modules do not have a CAN interface, but they
communicate via an USART interface with the main microcontroller.
In the future, the focus will be on an interactive network of several inde-
pendent vehicles based on radio transmission. This will allow all vehicles to
communicate with each other and share information like traction and behavior
of the road, actual position from GPS, or speed. The radio transmission is
carried out with the industry standard called “Zigbee”. An XBEE module of
the company “Digi” undertakes the radio transmission. The module uses an
UART interface for the communication with the main microcontroller on the
vehicle. Via this interface, the car will get information from other cars nearby.
A detailed overview of the data processing system, including the XBEE
module, is shown in Figure 9.6.
 9.4 Lane Detection 199

Figure 9.6 Overview of the data processing system.

9.4 Lane Detection

There are several steps needed to accomplish the lane detection.
First, the image has to be analyzed with an In-Range filter. In the second
step, the points that the Hough-transformation has identified as lane marks, are
divided into left and right lane marks. Next, the least squares method is used
to transform the lane marks into a second-degree polynomial, thus providing
the base to calculate the driving lane. Subsequently, the points of the driving
lane are transformed into world coordinates.
Two types of filters are used to get the needed information from the
image. Both are functions from the OpenCV-library. An In-Range filter
is used to detect the white lane marks on the defined test track. The
Hough-transformation calculates the exact position of the lane marks prepar-
ing them for the next steps.

9.4.1 In-Range Filter

The In-Range filter transforms an RGB-image into an 8-bit binary image. It’s
made for the detection of pixels in a variable color range. The transformed
picture has the same resolution as the original picture. Pixels belonging to
the chosen color range are white. All other pixels in the image are black. The
function works with the individual values of the RGB format. The chosen
color is defined by two critical values of this format.
Figure 9.7 shows the result of the In-Range filter.
200 Distributed Data Acquisition and Control Systems

Figure 9.7 Comparison between original and in-range image.

9.4.2 Hough-Transformation
The Hough-transformation is an algorithm to detect lines or circles in images,
which in this case means that it investigates the binary image from the In-Range
filter in order to find the lane marks.
The Hessian normal form converts individual pixels, so that they can be
recognized as lines in the Hough space. In this state, space lines are expressed
by the distance to the point of origin and the angle to one of the axes. Due to
the fact that the exact angle of the marks is unknown, the distance to the point
of origin is calculated based on Equation (9.1), utilizing the most probable
angles:
r = x · cos (a) + y · sin (a) . (9.1)
The intersection of the sinusoidals provides an angle and the distance of
the straight line from the origin of coordinates. These parameters create a
new line, so that the majority of the pixels can be detected. Furthermore, the
function from the OpenCV-library returns the start and the endpoint of each
Hough-line. As Figure 9.8 shows, the lines of the Hough-transformation are
precisely mapped on the lane marks of the road.

Figure 9.8 Original image without and with Hough-lines.

 9.4 Lane Detection 201

9.4.3 Lane Marks

To provide a more precise calculation, all points along the line are included.
These points are stored in two arrays and then sorted.As a first sorting criterion,
the position of the last driving lane is used. The second criterion for sorting
derives from their position in the image.
As mentioned before, the information in the image regarding long distances
can be critical depending on the viewing angle and height of the camera. In
order to concentrate on noncritical information only, points in the middle area
of the image are used. Figure 9.9 shows the sorted points on the right and the
corresponding Hough-lines on the left side.

9.4.4 Polynomial
To describe the lane marks more efficiently, a second-degree polynomial is
used. The coefficients of the parable are derived by the least-squares method.
A polynomial of a higher degree isn’t needed, because the effort to calculate
the coefficients is too high to make sense in this context, for the speed of the
image processing is one of the critical points of the project. Furthermore, the
area of the road, which is pictured by the camera, is too small. The road is
unable to clone the typical form of a third-degree polynomial.
As visible in Figure 9.10, the parables derived from the sorted points are
mapped precisely on the lane marks of the road. The algorithm to calculate
the coefficients derived from the points of the lane marks is handwritten.

9.4.5 Driving Lane

The driving lane for the car lies between the parables mentioned in the last
chapter. To calculate the position of the points of the driving lane, the average

Figure 9.9 Hough-Lines and sorted points along the Hough-Lines.

202 Distributed Data Acquisition and Control Systems

Figure 9.10 Sorted Points and Least-Square Parable.

of two opponent points of the two parables is taken. According to 9.2, the
average for the x- and y-coordinates is calculated.
       
xm x1 1 x2 x1
= + − . (9.2)
ym y1 2 y2 y1

In order to simplify the transformation from pixel-coordinates to

world-coordinates, the driving lane is described by a fixed number of points
in the image. The essential feature of these points is that they lie in predefined
rows in the image. So, there is only the need to calculate the horizontal position
of the parable for these points.
Theoretically it is possible that the program delivers an incorrect driving
lane. Mistakes can occur because of flash lights, reflections on the road,
missing lane marks due to different reasons or extreme light conditions, which
are much faster than the auto white balance of the camera can bear. So in
order to avoid mistakes that occur within a short time period, some kind of
stabilization is required. Short time in this case means shorter than one second.
For the purpose of stabilization, the different driving points are stored. The
stabilization works with these stored points in combination with four defined
edge points in the image. First, the algorithm checks if the edge points of the
new image differ from the edge points in the old image.
If the difference between the old points and the new points is low, the
driving lane is calculated and the driving points are stored. In case that the
difference between the points is too big, the driving lane is not updated and
the driving lane is calculated by using the stored points. The algorithm works
with the changes of the stored points. The new points are calculated by using
the difference between the last image and the current one. This difference is
 9.4 Lane Detection 203

derived from the change of the difference between the third and second image,
that have been taken before the current one, and the difference between the
second and the first image before the current one.
The critical values for the difference also depend on this calculation. That
means that in curvas, the critical values are higher. If not, only the last three
images are used for the calculation, in order to reduce the noise of the driving
lane. However, in this case, the reaction time of the algorithm is lesser.
The reaction time also depends on the fps (frames per second) of the
camera. For this project, a camera with 100 fps is used and the last fifteen
driving lanes are stored. The number of stored driving lanes for 100 fps is
based on experimental research.
Figure 9.10 shows the driving lane in red color. The four edge points mark
the edge points of the rectangle.

9.4.6 Stop Line

One of the main tasks of the camera is to detect stop lines. Figure 9.11 shows
the dimensions of the stop lines for this test track.
In order to detect stop lines, the algorithm is searching for the main
characteristics. First, the stop line is a horizontal line in the image. If the
angle of a vertical line in the image is defined as zero degrees, that means,
that the perfect stop line has an angle of 90 degree. The algorithm not only
searches for 90 degree lines. The angle for a potential stop line is smaller
than –75 degree and bigger than +75 degree.
The next criterion of a stop line is that it lies on the car traffic lane. So,
the algorithm does not need to search in the complete image for stop lines,
but only in the area of the cars traffic lane. This area is marked by the four
edge points of the rectangle mentioned in the last chapter. Once the algorithm

Figure 9.11 Parables and driving lane.

204 Distributed Data Acquisition and Control Systems

finds a potential stop line in the right area with a correct angle, the algorithm
checks the next two characteristics of a stop line: the length and the width of
the stop line.
The length of the stop line is easy to check. The stop line must be as long
as the road is wide, so the algorithm only needs to check the endpoints of the
line. On the left side, the endpoint of the stop line must lie on the middle road
marking. On the right side, the stop line borders on the left road marking from
the crossing road. The stop line and the road marking differ in just one point:
the width.
Since it is not possible to perceive the differences of the width in each
situation, the stop line has no defined end point on this side. So, the algorithm
checks if the end point of the potential stop line lies on or above the right road
marking. It is hard to measure the width of a line in an image that has constant
width and length in reality. The width of the line in the image in pixels depends
on the camera position in relation to the line, the numerical aperture of the
camera lens and the resolution of the camera. So, because in this project the
position of the camera changes from time to time, measuring the width is not
reliable to perceive the stop line. Therefore, the width is not used as a criterion
for stop lines.
Figure 9.12 shows a typical crossing situation. The left image visualizes
the basic situation and the middle image shows the search area as a rectangle.
Here you can see that the stop line on the left side is not covered by the research
area so the algorithm doesn’t recognize the line as a stop line. On the right
image, the stop line ends correctly on the middle road marking. The line in
the image shows that the algorithm has found a stop line. Due to the left road
marking from the crossing road, the line ends outside the real stop line.

9.4.7 Coordinate Transformation

To control the car, the lateral deviation and the course angle are needed.
Both are calculated by the controller of the camera. The scale unit for the

Figure 9.12 Detection of stop lines.

 9.4 Lane Detection 205

lateral deviation is meters and degrees for the course angle. Course angle
means the angle of the driving lane which is calculated by the camera.
The lateral deviation is the distance of the car’s center of gravity to the
driving lane when they are at the same level. Since the lateral deviation
is needed in meters, the algorithm has to convert the pixel coordinates
from the image into meters in the real world. The course angle can be
calculated from the pixel coordinates in the image, but this method is
error-prone.
There are two different methods to convert the pixels into meters.
Pixels can be converted via Equations (9.3) and (9.4).

h 2α
x(u, v) =    · cos (γ̄ − α) + u + l, (9.3)
tan θ̄ − α + u n−12α n−1

h 2α
x(u, v) =    · cos (γ̄ − α) + u + l. (9.4)
tan 2α
θ̄ − α + u n−1 n −1

In the equations, x and y are the coordinates in meters. γ stands for the drift
angle of the camera in the plane area and θ stands for the pitch angle of the
camera. α is the numerical aperture of the camera, u and v are the coordinates
of one pixel in the image.
Using this equation, the complete image can be converted into
real-world coordinates. The drawback of this method is that all parameters
of the camera have to be known exactly; every difference between the
numerical aperture in the equation and the exact physical aperture of the
camera lens can cause massive failure in the calculation. Furthermore, this
method needs more calculation time on the target hardware. A big plus of
this method is that the camera can be re-positioned during experimental
research.
The second method is to store references to some pixels in lookup tables.
For these pixels, the corresponding values in meters can be calculated or can be
measured. This method expends much less calculation time but is also much
less precise. With this method, the camera cannot be re-positioned during
experiment research. Every time the camera is re-positioned the reference
tables must be re-calculated.
The method to prefer depends on the project requirements regarding
accuracy and the projects hardware. For this project, the second method is
used. To meet the demands on accuracy, for each tenth pixel of the camera, a
reference is stored.
206 Distributed Data Acquisition and Control Systems

9.5 Control of the Vehicle

The driving dynamic of the vehicle is characterized by the linear track model
of Ackermann. As Figure 9.13(a) shows, the model is composed of a rear and
front wheel which are connected by an axe. In order to rotate the vehicle on
its main axe, the steering angle can be set with the front wheel.
To reduce the complexity of vehicle dynamics, three simplifications
are made.
These are:
• Neglect the air resistance, because the vehicle speed is very low;
• Lateral forces on the wheels are linearized;
• No roll of the vehicle about the x and y axis.
Using these simplifications, the created model should differ only marginally
from reality. Linking the transverse dynamics of the vehicle with the driving
dynamics, you can derive the following relation:
⎡ ⎤ ⎡ ⎤ ⎡ ⎤ ⎡ ⎤
ψ̈ (t) a22 0 0 ψ̇ (t) b2
r ⎣θ̇Δ (t) ⎦ = ⎣ −1 0 0 ⎦ · ⎣θΔ (t) ⎦ + ⎣ 0 ⎦ · δ (t) . (9.5)
γ̇ (t) 0 V 0 γ (t) 0

Because of the equation in the state space, a controller can be designed using
tools such as Matlab Simulink or Scilab X-cos.

Figure 9.13 Driving along a set path: Track model (a); Lateral deviation and heading
angle (b).
 9.7 Conclusions 207

In order to keep the vehicle on the track, the conditions such as heading
angle, the yaw rate and lateral deviation must be known. A gyroscope is
used to detect the yaw rate of the vehicle. The lateral deviation and the
course angle are calculated from the camera. The camera sends the lateral
deviation and the curse angle. Until the next image is analyzed, the coordinates
on the microcontroller stay the same as before. Between two pictures the
lock angle and the transverse deviation were recalculated after each motion.
This is possible because the velocity and yaw rate are known at any time.
Figure 9.13(b) illustrates the relationship of lateral deviation (ΔY) and heading
angle (α).

9.6 Results
This section gives an overview of the project results.
The autonomous car was built with the hardware suggested before.
Experiments on scaled test roads show that the car can drive autonomously.
However, the tests also showed the limitations of this prototype. The effort
for the image processing was undervalued. The on-board processor of the
camera isn’t able to accomplish the necessary calculations in time. In terms
of reaction to this fact, the maximum speed of the car has to be very slow. If
it isn’t, the control of the vehicle gets unstable with a more or less random
driving path. In addition, the car has problems with too sharp curves. The
process to divide the image isn’t dynamic, so in curves the preset section
becomes incorrect and the algorithm isn’t able to calculate a correct driving
path. Thanks to its laser sensors, the car is able to avoid collisions with
baffles.
To improve the performance of the car, the hardware for the image
processing has to be improved. The image processing works stable. Problems
derive from the calculation algorithm of the driving path. At this point in time,
the algorithm doesn’t contain the necessary interrupts for every situation on
the road, but this drawback will be corrected in the second prototype.

9.7 Conclusions
In this chapter, an autonomous vehicle with distributed data acquisition and
control systems has been presented. For control, the vehicle has a number
of independent sensors. The main sensor is a camera with a lane tracking
algorithm, which contains edge detection and Hough transformation. The lane
is verified by laser sensors in the front and side of the vehicle. It is planned
208 Distributed Data Acquisition and Control Systems

to build a superordinate control system, which leads a group of autonomous

vehicles using a wireless communication protocol.

References
[1] J. Canny, ‘A Computational Approach to Edge Detection’, IEEE 1986.
[2] A. Erhardt, ‘Einführung in die Digitale Bildverarbeitung’, Offenburg
2008.
[3] W. Heiden, ‘Kanten in Bildern - Filterung und Kantenerkennung’,
St. Augustin 2009.
[4] B. Jähne, ‘Digitale Bildverarbeitung und Bildgewinnung’, Berlin 2013.
[5] M. Jauernig, ‘Einsatz von Algorithmen der Photogrammmetrie und
Bildverarbeitung zur Einblendung spezifischer Lichtraumprofile in
Videosequenzen’, Hochschule Leipzig 2006.
[6] A. Kant, ‘Bildverarbeitungsmodul zur Fahrspurerkennung für ein
autonomes Fahrzeug’, Hochschule Hamburg 2007.
[7] L. F. Kirk, ‘Schätzung der Brennweite mit Hilfe der Hough-
Transformation’, Fachhochschule, Köln 2011.
[8] N. Kruse, ‘Kameragestützte Fahrspurerkennung für autonome Modell-
fahrzeuge’, Universität Hamburg 2008.
[9] G. Linß, ‘Praktische Ausbildung und Training Qualitätsmanagement
Objekterkennung mit Hough-Transformation’, Technische Universität
Ilmenau.
[10] R. Maini, H. Aggarwal, ‘Study and Comparison of Various Image Edge
Detection Techniques’, Punjabi University, India.
[11] P. Schöley, Kantendetektoren, Technische Universität Dresden, 2011.
[12] J. Unger, ‘Untersuchung von Linien und Kantenextraktionsalgorith-
men im Rahmen der Verifikation von Ackerlandobjekten’, Universität
Hannover, 2009.
[13] C. Wagner, Kantenextraktion – Klassische Verfahren, Universität, Ulm,
2006.
[14] G. M. Wagner, G. M., ‘Bestimmung der Kameraverzerrung mit Hilfe
der Hough-Transformation’, Fachhochschule Köln, 2011.
[15] L. Bergen, H. Burkhardt, Morphological image processing.
[16] J. Wohlfeil, ‘Detection and tracking of vehicles with a moving
camera’ Humbolt-Universität zu Berlin Institut für Informatik;
Deutsches Zentrum für Luft- und Raumfahrt Institut für Verkehrs-
forschung, 2006.
 References 209

[17] B. Hulin ‘Video-based obstacle detection clearance of the pantograph

electric railways’, Fakultät für Elektrotechnik und Informationstech-
nick, Technische Universität München, 2003.
[18] J. Alberts, ‘Vision for the project FAUST, Focus of Expansion/optical
flow’, Hamburg University of Applied Sciences, 2007.
[19] Tutorial Filter; Vision&Control System Lighting Optics, 2007.
[20] Prof. Dr. –Ing. Habil G. Linß, ‘Practical education and training quality
management, object detection with Hough transform’, Technische Uni-
versität Ilmenau, Department of Mechanical Engineering, Department
of Quality Assurance.
[21] C. Wagner, Edge extraction, Classical methods; Seminar presentation
on “Image segmentation and computer Vision”, 2006.
[22] I. Nikolov, ‘Adaptive camera parameters for optimum lane detection and
tracking’, Hamburg University of Applied Sciences, 2007.
[23] V. Schaefer, A. Zipser, ‘Algorithmic Applications Hough transform’,
2006.
[24] C.Rathemacher, ‘GPU-based detection of algebraic structures writable’,
Fachhochschule Wiesbaden, Fachbereich Design Informatik Medien,
2007.
[25] M. Pelkofer, ‘Behavioral decision for autonomous vehicles with
gaze control’, Universität der Bundeswehr München, Fakultät für
Luft- und Raumfahrttechnik, Institut für Systemdynamik und
Flugmechanik, 2003.
[26] D.Berger, ‘Lane inference system for lane algorithm Three Feature
Based Lane Detection Algorithm (TFALDA)’, Hamburg University of
Applied Sciences, 2008.
[27] Alberto Broggi, ‘A Massively Parallel Approach to Real-Time
Vision-Based Road Markings Detection’, Dipartimento di ingengneria
dell’Information Universita di Parma.
[28] Andreas Weide, ‘Entwicklung einer Kameragestützten Fahrspurerken-
nung für ein autonomes Fahrzeug’, University of Applied Sciences
(Friedberg, Germany), 2013.
 10
Polymetric Sensing in Intelligent Systems

Yu. Zhukov, B. Gordeev, A. Zivenko and A. Nakonechniy

National University of Shipbuilding named after admiral Makarov, Ukraine

Corresponding author: Yu. Zhukov <[email protected]>

Abstract
The authors examine the up-to-date relationship between the theory of poly-
metric measurements and the state of the art in intelligent system sensing.
The chapter contains commentaries about: concepts and axioms of polymetric
measurements theory, corresponding monitoring information systems used in
different technologies and some prospects for polymetric sensing in intelligent
systems and robots. The application of the described concepts in technological
processes ready to be controlled by intelligent systems is illustrated.

Keywords: Polymetric signal, polymetric sensor, intelligent sensory system,

control, axiomatic theory, modelling.

10.1 Topicality of Polymetric Sensing

For some time it has been widely recognized [1] that the future factory should
be a smart facility where design and manufacture are strongly integrated into a
single engineering process that enables ‘right first time’ every time production
of products to take place. It seems completely natural to use intelligent robots
and/or control systems at such factories to provide on-line, opportune and
precise assessment of the quality of the production process and to guarantee
the match and fit, performance and functionality of every component of the
product prototype that is created.
Wide deployment of sensor-based intelligent robots at the heart of future
products and technology-driven systems may substantially accelerate the

Advances in Intelligent Robotics and Collaborative Automation, 211–234.


c 2015 River Publishers. All rights reserved.
212 Polymetric Sensing in Intelligent Systems

process of their technological convergence and lead to their introduction in

similar processes and products.
At the same time it is well known [2] that intelligent robot cre-
ation and development is a rather fragmented task and there are different
progress-restricting problems in each particular field of research. Numerous
advanced research teams are doing their best to overcome these limitations
for intelligent robot language, learning, reasoning, perception, planning and
control.
Nowadays, a few researchers are still developing the Ideal Rational Robot,
as it is quite evident that the computations necessary to reach ideal rationality
in real operating environments require much more time and much more
productive processors. Productivity is permanently growing but being still
limited for IRR tasks. Furthermore, the robotic perception of all the neces-
sary characteristics of real operating environments has not been sufficiently
provided yet.
The number of corresponding sensors and instruments necessary for these
robots is growing, but their sensing remains too complicated, expensive, time-
consuming and IRR remains far from being realistic for actual application.
F.H. Simons’ concept of Bounded Rationality and the concept of Asymp-
totic Bounded Optimality based on S.J. Russell’s and D. Subramanian’s
approach of provably bounded-optimal agents have formed a rather pragmatic
and fruitful trend for the development of optimal programs.
One has to take into account the fact that time limitations are even stricter
if we are trying to overcome them for both calculation and perception tasks.
While immense progress has been made in each of these subfields in the last
few decades, it is still necessary to comprehend how they can be integrated to
produce a really effective intelligent robot.
The introduction of the concepts of agent-based methods and systems
[3, 4] including holonic environment [5] and multi-agent perceptive agencies
[6] jointly with the concept of polymetric measurements [7–9] has engendered
a strong incentive to evaluate all the potentialities of using polymetric sensing
for intelligent robots, as this may be a precondition for generating real benefits
in the field.
The matter is that obvious practical success has been achieved during the
wide deployment of SADCO polymetric systems for monitoring a variety
of multiple quantitative and qualitative characteristics of different liquid
and/or loose cargoes using a single polymetric sensor for measuring more
than three characteristics of cargoes simultaneously within a single chronos-
topos framework. Similar information systems were also a successful tool for
 10.2 Advanced Perception Components of Intelligent Systems or Robots 213

the online remote control of complex technological processes of production,

storage and consumption of various technological media [8].
But, as indicated above, there are very specific requirements for sensing in
intelligent robots. In fact, one of the most important restrictions is connected
with very limited time-consumption for the input of real-time information
concerning an intelligent robot and/or multi-agent control systems operational
environment.
Thus, we face an actual and urgent need to integrate and combine these
advanced approaches within the calculation and perception components of
intelligent robots and/or multi-agent monitoring and control system design,
starting from different (nearly diametrically opposite) initial backgrounds of
each component and, by means of irrefutable arguments, arriving at jointly
acceptable conclusions and effective solutions.

10.2 Advanced Perception Components of Intelligent

Systems or Robots
10.2.1 Comparison of the Basics of Classical and Polymetric
Sensing
Classical or industrial metrology is based on the contemporary Axiomatic
Theory [10, 11].
In classical industrial metrology, it is presupposed that for practical
measurement processes it is necessary to have some set Ω = {σ, λ,. . . } of
different instruments with various sensor transformation and construction
systems.
But every particular instrument has a custom-made uniform scale for
the assessment of the actual value of the object specific characteristic under
control. Let aσ (i) be a numerical function of two nonnumeric variables – a
physical object i∈ℵ and a specific instrument, i.e.
a : Ω × ℵ →  and (σ, i) → aσ (i). (10.1)
This function is called a quantitative assessment of a physical quantity of an
object.
In the general case for an industrial control system or for intelligent robot
sensing process, it is necessary to have N instruments for each characteristic
under monitoring at M possible locations of the components of the object
under control. The more practical the application, the quicker we face the
curse of multidimensionality for our control system.
214 Polymetric Sensing in Intelligent Systems

Polymetric measurement in general is the process of getting simultaneous

assessments of a set of object physical quantities (more than two) using one
special measuring transformer (a sensor). The first successful appraisal of the
corresponding prototype instruments was carried out in 1988–1992 on-board
three different vessels during full-scale sea trials. After the successful tests
and the recognition of the instruments by customers and the classification
societies, the theoretical background was generalized and presented to the
scientific community [12–14].
The latest results [7–9] seem to be prospective for intelligent robot sensing
due to reduced time and financial pressure, simplified design and reduced
general number of sensory components.
Summarizing the comparison of the basics of classical and polymetric
measurement theories, it is essential to comment another consequence of their
axioms and definitions. The introduction of the principle of the simultaneous
assessment of a physical quantity and its measurement from the same poly-
metric signal Polymetric signal is one of the key provisions of the theory and
the design practice for developing appropriate instruments. The structure of
an appropriate perceptive intelligent control or monitoring system should be
changed correspondingly.

10.2.2 Advanced Structure of Multi-Agent Intelligent Systems

In order for multi-agent control or monitoring systems and intelligent robots
to satisfactorily fulfil the potential missions and applications envisioned for
them, it is necessary to incorporate as many recent advances in the above
described fields as possible within the real-time operation of the intelligent
system or robot-controlled processes.
This is the challenge for the intelligent robotics engineer, because many
advanced algorithms in this field still require too much time for computation,
despite improvements made in recent years in microelectronics and algo-
rithms. Especially, it concerns the problem of intelligent robot sensing (timely
and precise perception).
The well-known variability versus space (topos) and versus time (chronos)
is related to similar variability of the measured data [6, 7, 10] engendering the
advantages of using polymetric measurements.
That is why, in contrast to the multi-sensor perceptive agency concept [6]
based on the use of several measuring transformers, each one of them being
the sensing part of each particular agent within the distributed multi-sensor
control system (i.e. several perceptive agents in complex perceptive agency),
 10.2 Advanced Perception Components of Intelligent Systems or Robots 215

the use of one equivalent polymetric transformer for an equivalent perceptive

agency is proposed.
The concept of Polymetric Perceptive Agency (PPA) for intelligent system
and robot sensing is schematically illustrated in Figure 10.1. Such simplified
structure of PPA sub-agency of the Decision-making Agency (DMA) is
designated to be used in different industries and technologies (maritime,
on-shore, robotics, etc. [9, 15]).
There are some practical optimistic examples of the successful deploy-
ment and long-term (more than 15 years) operation of industrial polymetric
monitoring and control systems/agencies based on polymetric measurement-
Polymetric measurement technique in different fields of manufacturing and
transportation [8].

Figure 10.1 The main idea of the replacement of the distributed multi-Sensor system by a
polymetric perceptive agent.
216 Polymetric Sensing in Intelligent Systems

10.3 Practical Example of Polymetric Sensing

Here we describe a practical case from the research of the Polymetric Systems
Laboratory at the National University of Shipbuilding and LLC AMICO
(Mykolaiv, Ukraine). The practical goal is to ensure the effective and safe
control of the water level in the cooling pond of nuclear power stations
at the spent nuclear fuel storage during normal operation and emergency
post-accident operation. There are many level sensors, which are used by
the control systems in the normal operation mode: floating-type, hydrostatic,
capacitive, radar, ultrasonic, etc. [16]. But there exist many problems con-
cerning their installation and functioning under real post-accident conditions.
The matter is that high pressure and extremely high temperature, saturated
steam and radiation, vibration and other disturbing factors are expected in the
cooling pond in emergency mode.
Thus, high reliability and radiation resistance are the most important
requirements for such level-sensing equipment. One of the most suitable
sensing techniques in this case is the proposed modified Time Domain
Reflectometry (TDRTdr) – see Figure 10.2.

Figure 10.2 TDR Coaxial probe immersed into the liquid and the corresponding polymetric
signal.
 10.3 Practical Example of Polymetric Sensing 217

In this type of level measurement, microwave pulses are conducted along

a cable or rod probe 1 partially immersed in the liquid 2 and reflected by the
product surface.
Sounding and reflected pulses are detected by the transducer. The transit
time of the pulse t 0 is a function of the distance from the level sensor to the
surface of the liquid L0 . This time is measured and then the distance from
the level sensor to the surface of the liquid is calculated according to the
calibration function:
L0 = f (t0 ), (10.2)
which is usually presented as a linear function:
L0 = b0 t0 + b1 , (10.3)
where b0 and b1 – coefficients which are obtained during a calibration
procedure.
From the physical point of view, the slope of this function stands for
the speed of the electromagnetic wave which propagates forward from the
generator on the pcb-board, along the cable and the specially designed
measuring probe, and back to the generator. It can be defined as:
c
b0 = √ , (10.4)
2 ε0
where c – speed of light in vacuum; ε0 – dielectric constant of the material
through which the electromagnetic pulse propagates (it is close to 1 for air); the
coefficient of ½ stands for the fact that the electromagnetic pulse propagates
along double the length of the probe (forward and backward).
The fact of the presence and the value of the intercept b1 are caused by
many reasons. One of them is that the electromagnetic pulse actually passes
the distance greater than the measuring probe length. In the general case, the
real calibration function is not linear [8].

10.3.1 Adding the Time Scale

In this practical case, the basic approaches for constructing the polymetric
signal are described based on the TDR technique. The concept of forming an
informative pulse polymetric signal includes the following: generation of short
pulses and sending them to a special measuring probe, receiving the reflected
pulses and signal pre-processing for its final interpretation. It is worth men-
tioning that in terms of polymetrics, each of these steps is specially designed
to increase the informativity and interpretability of the resulting signal.
218 Polymetric Sensing in Intelligent Systems

Therefore, to calculate the distance from the sensor to the surface

of the liquid, it is necessary to carry out the calibration procedure. The
level-sensing procedure can be simplified by adding additional information to
the initial «classic» signal to obtain the time scale of this signal. A stroboscopic
transformation of the real signal is one of the necessary signal pre-processing
stages during the level measurement procedure.
This transformation is required to produce an expanded time sampled
signal for its future conversion to digital form. This transformation can be
carried out with the help of a stroboscopic transformer based on two oscillators
with frequencies f 1 (the frequency of input signal) and f 2 (the local oscillator
frequency) that are offset by a small value Δf = f 1 – f 2 [17]. The duration of
the input signal is:
1
T1 = . (10.5)
f1
As a result of this transformation, we have expanded signal duration:
1 1
TT S = = . (10.6)
Δf f1 − f2
In this case, the relationship between the duration of transformed and the
original signal is expressed by the stroboscopic transformation ratio:
TT S f1
KT S = = . (10.7)
T1 f1 − f2
The next processing step of the transformed signal is the conversion of this
inherently analog signal to digital form with the help of the analog-to-digital
converter (ADC).
The time scale and delays during analog-to-digital conversion are known.
Therefore, it is possible to count ADC conversion cycles and to calculate the
time scale of the converted signal. It is not convenient because conversion
cycle duration is connected with the ADC parameters, clock frequency value
and stability, etc. In order to exclude the use of the conversion cycles number
and ADC parameters, it is possible to «add» the additional information about
the time scale of the converted signal.
In this case, we can add a special marker which helps to measure the
reference time interval – see Figure 10.3.
The main idea is that ADC conversion time TADC must be greater than the
duration of the signal from the output of the stroboscopic transformer TT S to
obtain at least 2 sounding pulses in the resulting digitized signal. It is necessary
 10.3 Practical Example of Polymetric Sensing 219

Figure 10.3 Time diagrams for polymetric signal formation using additional reference time
intervals.

to count the delays between two sounding pulses τ ZZ and between the first
sounding pulse and the reflected pulse τ ZS (in terms of the number of ADC
readings).
The delay τ ZZ , expressed in ADC readings count, corresponds to the time
delay TT S (in seconds). The next equation should be used to calculate the time
scale of the transformed and digitized signal:
TT S
KDT = s/reading. (10.8)
τZZ
It is possible to calculate the time value of the delay between the pulses t0
(see Figure 10.2):
KDT τZS τZS
t0 = = s. (10.9)
KT S f1 τZZ
Finally, to calculate the distance to the surface of the liquid, it is necessary
to use the equation:
c τZS
L0 = √ + b1 , (10.10)
2 ε0 f1 τZZ
where b1 (zero shift) is calculated using information on the PCB layout and
generator parameters or during a simple calibration procedure.

10.3.2 Adding the Information about the Velocity of the

Electromagnetic Wave
In the previous paragraph, a special marker was added to the signal to obtain
the time scale of the signal and to easily calculate the distance to the surface
of the liquid in normal operating mode.
220 Polymetric Sensing in Intelligent Systems

But, as it was mentioned above, the emergency operating mode can be

characterized by high temperatures and pressure, the presence of saturated
steam in the cooling pond. Steam under high pressure and temperature will
slow down the propagation speed of a radar signal which can cause an
additional measurement error. It is possible to add the additional information
to the signal for automated correction of measurement. A special reflector or
several reflectors are to be used in this case [8]. Almost any step of wave
impedance discontinuities can be used as a required reflector (probes with
stepwise change of the impedance, fixing elements, etc.). The reflector is
placed at a fixed and known distance LR from the generator and the receiver
of short electromagnetic pulses GR in the vapour dielectric – presented in
Figure 10.4.
If there is vapour in the tank, the pulses reflected from the special reflector
and from the surface of the liquid are shifted according to the change of the
dielectric constant of vapour (as compared to the situation when there is no
vapour in the tank).

Figure 10.4 Disposition of the measuring probe in the tank, position of the reflector and
corresponding signals for the cases with air and vapour.
 10.4 Efficiency of Industrial Polymetric Systems 221

The delays between the sounding pulse and the pulse reflected from the
special reflector τ R and between the sounding pulse and the pulse reflected
from the surface of the liquid τ ZS for cases with air and vapour are different.
The dielectric constant ε0 can be calculated using the known distance LR .
Therefore, the distance to the surface of the liquid L0 can be calculated using
the corrected dielectric constant value:
(LR − b1 )τZS
L0 = + b1 . (10.11)
τZR
As it can be seen from the equation, the result of the measurement depends on
the time intervals between the sounding and reflected pulses and the reference
distance between the generator and the reflector.
The above-described example of polymetric signal formation showed
that a single signal carries information about the time scale of this signal,
the velocity of propagation of electromagnetic waves in dielectrics and the
linear dimensions of these layers. This list of measurable parameters can be
easily continued by using additional information in the existing signals and
building new «hyper-signals» [9] for measuring some other characteristics of
controllable objects (e.g. on the basis of the spectral analysis of these signals
the controllable liquid can be classified and some quality characteristics of the
liquid can be calculated).

10.4 Efficiency of Industrial Polymetric Systems

10.4.1 Naval Application
One of the first polymetric sensory systems was designated for on-board
loading and safety control (LASCOS) of tankers, fishing, offshore, supply
and research vessels. The prototypes of these systems were developed,
industrialized and tested during full-scale sea trials in the early 90-es of the
last century [18]. These systems have more developed polymetric sensing
subsystems (from the “topos-chronos” compatibility and accuracy points of
view). That is why they are also successfully providing commercial control
of cargo handling operations.
The structure of the hardware part of LASCOS for a typical offshore
supply vessel is presented in Figure 10.5. It consists of the following:
operator workplace (1); a radar antenna (2); an on-board anemometer (3);
the radar display and a keyboard (4); a set of sensors for ship draft monitoring
(5); a set of polymetric sensors for fuel-oil, ballast water and other liquid
 222 Polymetric Sensing in Intelligent Systems

Figure 10.5 Example of the general structure of lascos hardware components and elements.
 10.4 Efficiency of Industrial Polymetric Systems 223

cargo quantity and quality monitoring and control (6); a set of polymetric
sensors for liquefied LPG or LNG cargo quantity and quality monitoring and
control (7); switchboards of the subsystem for actuating devices and operating
mechanisms control (8); a basic electronic block of the subsystem for liquid,
liquefied and loose cargo monitoring and control (9); a block with the sensors
for real-time monitoring of ship dynamic parameters (10).
The structure of the software part of the typical sensory intelligent
LASCOS is presented in Figure 10.6.
It consists of three main elements: a sensory monitoring agency (SMA)
which includes three other sensory monitoring agencies – SSM (sea state,
e.g. wind and wave model parameters), SPM (ship parameters) and NEM
(navigation environment parameters); an information environment agency
(INE) including fuzzy holonic models of ship state (VSM) and weather
conditions (WCM), and also data (DB) and knowledge (KB) bases; and
last but not least – an operator interface agency (OPIA) which provides the
decision-making person (DMP) with necessary visual and digital information.
Unfortunately, until now, the above-mentioned challenges have not been
combined together in an integrated model, which applies cutting-edge and
novel simulating techniques. Agent-based computations are adaptive to
information changes and disruptions, exhibit intelligence and are inherently
distributed [4]. Holonic agents inherently may help design and operational

Figure 10.6 The general structure of lascos software elements and functions.
224 Polymetric Sensing in Intelligent Systems

control processes in self-recovery and react to environmental real-time

perturbations.
The agents are vital in a ship operations monitoring and control context, as
ship safety refers to the inherently distributed and stochastic perturbation of its
own state parameters and external weather excitations. Agents are welcome in
the on-board LASCOS system design because they provide properties such as
autonomy, responsiveness, distributiveness, openness and redundancy [15].
They can be designed to deal with uncertain and/or incomplete information
and knowledge and this is extremely topical for fuzzy LASCOS as a whole. On
the other hand, the problem of their sensing has never been under systematic
consideration yet.

10.4.1.1 Sensory monitoring agency SMA

As mentioned above, all initial real-time information for the LASCOS as well
as the sensory system is provided to DMP by the holonic agent-based sensory
monitoring agency (SMA) which includes three other sensory monitoring
agencies – SSM, SPM and NEM (see Figure 10.7).
The simultaneous functioning of all the structures presented in Figure 10.5,
Figure 10.6 and Figure 10.7, i.e. hardware, logic and process algorithms, the
corresponding software components and elements of LASCOS provide online
forming of information environment for the system operator or seafarer – the
decision-maker.
The polymetric sensing of the computer-aided control and monitor-
ing system for cargo and ballast tanks is designed to ensure effective
and safe control of ship operations in remote mode. This system is one
of the most critical systems on-board a ship or any other ocean-floating
vehicle. The operation of the system must be finely adjusted with the
requirements of the marine industry. The polymetric system is a high-
tech solution for the task of on-line monitoring and control of the ship
liquid cargoes and ballast water tanks state. These systems are also des-
ignated for ship safety parameter monitoring and for docking operations
control.
The solution provides safe and reliable operations in real-life harsh con-
ditions, reduces risks for vessels and generates both operational and financial
benefits for customers providing:
• The manual, automated or automatic remote control of cargo handling
operations;
 10.4 Efficiency of Industrial Polymetric Systems 225

Figure 10.7 The General Structure of LASCOS Holonic agencies functions.

• The on-line monitoring of level, volume, temperature, mass, centre of

mass for liquid, loose cargo or ballast water in all tanks;
• The monitoring and control of list and trim, draft line, bending and
sagging lines;
• The monitoring of the vessel position during operation or docking; level
indication and alarm of all monitored tanks;
• The remote control of cargo valves and pumps according to the actual
operational conditions; the feeding of control signals to actuators, valves
and pumps; the monitoring of hermetic dryness of dry tanks and
conditions of water ingress in the possible damage conditions;
• The audible and visual warning about risky process parameters deviation
from the stipulated values;
• The registration and storage of retrospective information on operations
process parameters, equipment state and operator actions (“Black Box”
functions);
226 Polymetric Sensing in Intelligent Systems

• Comprehensive data and knowledge bases management;

• The testing and the diagnostics of all equipment operability.

10.4.1.2 Information Environment Agency INE

The information environment agency (INE) includes the fuzzy holonic models
of ship state (VSM), the weather conditions model (WCM), and also data (DB)
and knowledge (KB) bases. The central processor of the LASCOS system
processes all the necessary initial information concerning ship design and
structure in all possible loading conditions using the digital virtual ship model
(VSM).
The information concerning the actual distribution of liquid and loose
cargoes (including their masses and centres of gravity) are provided by the
sensory monitoring agency (SMA). This information is combined with the
information concerning actual weather and navigation environment details
that are converted into the actual weather condition model (WCM). The
information, permanently refreshable by SMA, is stored in the database and
it is used in real-time mode by the VSM and WCM models [18].

10.4.1.3 Operator Interface Agency OPI

The operator interface agency (OPI) of the LASCOS system provides the
decision-maker with all the necessary visual and digital information in the
most convenient way for decision-making support on quality, efficiency and
timeliness. Examples of some of these interfaces are presented in Figure10.8
and Figure 10.9.
One of the functions of the LASCOS system is to predict safe combinations
of ship speed and course in rough sea actual weather conditions. On the
safe-storming diagram, the seafarer may see three coloured zones of “speed-
course” combinations: the green zone of completely safe ship speed and
course combination (grey zone in Figure 10.8); the dangerous red zone with
“speed-course” combinations leading to ship capsizal with up to more than
95% probability (the black zone in Figure 10.8); and the yellow zone with
“speed-course” combinations with intensive ship motions, preventing normal
navigation and other operational activities (the white zone in Figure 10.8).
The centralized monitoring and control of ship’s operational processes is
performed via the operator interface from the main LASCOS system processor
screen located in the wheelhouse. The advanced operator interface is playing
an increasingly important role in the light of the visual management concept.
The proposed operator interface for LASCOS provides all traditional man-
machine control functionality. Moreover, this interface is clearly structured,
 10.4 Efficiency of Industrial Polymetric Systems 227

Figure 10.8 Safe Storming Diagram Interface of LASCOS.

powerful, ergonomic and easy to understand; it also makes the immediate

forecast of ship behaviour under the control action chosen by the operator,
thus preventing too risky and unsafe decisions. It also ensures all control
processes rules and requirements consolidating all the controllable factors for
better efficiency of operations to exclude human factor influence on decision-
making. The interface also ensures system security via personal password
profiles for nominated responsible and thoroughly trained persons to prevent
the incorrect usage of equipment and to avoid the unsafe conditions of ship
operations.

10.4.1.4 Advantages of the polymetric sensing

The described example of a sensory system for typical offshore supply vessel
(presented in Figure 10.5) can be used for the demonstration of the polymetric
sensing advantages.
The structure of the typical cargo sensory system for the offshore supply
vessel consists of the following parts: a set of the sensors for the fuel-oil,
liquefied LPG or LNG, ballast/drinking water and other liquid cargo quantity
and quality monitoring; a set of sensors for the bulk cargo quantity and quality
monitoring.
 228 Polymetric Sensing in Intelligent Systems

Figure 10.9 The Main Window of Ballast System Operations Interface.

 10.4 Efficiency of Industrial Polymetric Systems 229

Each tank is designated for the particular cargo type and equipped with
the required sensors. For example, for the measurement of the diesel oil
parameters, the corresponding tanks are equipped with the level sensors
(level sensors with the separation level measurement feature) and temperature
sensors. The total number of the tanks, corresponding sensors for the tanks
of a typical supply vessel, are shown in Table 10.1 (in case of the traditional
sensory system).
All the information acquired from the sensors must be pre-processed
for the final calculation of the required cargo and ship parameters in the
computing system. Each of the sensors requires power and communication
lines, acquisition devices and/or interface transformers (e.g. current loop into
RS-485 MODBUS) and so on.
In contrast to the classical systems, the polymetric sensory system requires
only one sensor for the measurement of all required cargo parameters in the
tank. Therefore, if we assume that traditional and polymetric sensory systems
are equivalent in the measurement information quality and reliability (systems
are interchangeable without any loss of measurement information quality),
it is obvious that polymetric system has the advantage in the measurement
channels number.
The cost criterion can be used for the comparison of the efficiency
of traditional and polymetric sensory systems. Denoting the cost of one

Table 10.1 Quantity and Sensor Types for the Traditional Cargo Sensory System (Example)
Total
Measureable Tanks Sensors Sensors
Cargo Type Parameters Number Number/Tank Number
Diesel oil Level in the tank, 6 3 18
presence of water in the
tank, temperature
LPG Level in the tank, 1 3 3
quantity of the liquid
and vapor gas,
temperature
Ballast water Level in the tank 6 1 6
Drinking Level in the tank, 6 2 12
water presence of other
liquids in the tank
Bulk cargo Level in the tank, 2 2 4
quality parameter
(e.g. moisture content)
Total 21 43
230 Polymetric Sensing in Intelligent Systems

measurement channel (the sensor + communication/power supply lines +

transformers/transmitters) as CT M S and CP M S , the cost of the processing
complex as CT P C and CP P C , the costs of the sensory systems for traditional
sensory system CT SS and for polymetric sensory system CP SS can be
estimated:
CT SS = CT P C + NT SS · CT M S ,
CP SS = CP P C + NP SS · CP M S ,
where NT SS , NP SS – the numbers of sensors used in the traditional and
polymetric sensory system correspondingly.
Assuming that
CT P C ≈ CP P C ≈ CT M S ≈ CP M S ,
the efficiency comparison of the polymetric and traditional sensory system
ET SS/P SS can be roughly estimated using the costs of the equivalent sensory
systems [8]:
NT SS + 1
ET SS/P SS = CT SS /CP SS = .
NP SS + 1
As can be seen from the above-described supply vessel example, the polymet-
ric sensory system efficiency is two times greater than the efficiency of the
traditional sensory system (ET SS/P SS = 44/22 = 2). It is worth mentioning
that this comparison of the system efficiency is very rough and it is used
only for the demonstration of the main advantage of the polymetric sensory
system.

10.4.1.5 Floating dock operation control system

Another example of naval polymetric sensory systems is a computer-aided
floating dock ballasting process control and monitoring system (CCS DBS)
with the main interface window presented in Figure 10.9.
These systems were designated to ensure effective and safe control of
docking operations in the automatic remote mode. The ballast system is
one of the most critical systems on the floating dock. The operation must
be finely coordinated with the requirements of the marine industry. The
polymetric system enables to take high-tech solutions to support safe control
and monitoring of the dock ballast system. This solution provides the safe
and reliable operations of dock facilities in real-life harsh conditions, reduces
risks for vessels and generates both operational and financial benefits to
customers.
The main system interface window (see Figure 10.9) contains the techno-
logical equipment layout and displays the actual status of a rudder, valves,
 10.4 Efficiency of Industrial Polymetric Systems 231

pumps, etc. The main window consists of the main menu, the toolbar,
the information panel and also the technological layout containing control
elements.
It is possible to start, change parameters or stop technological processes by
clicking a particular control element. The user interface enables the operator
to efficiently supervise and control every detail of any technological process.
All information concerning ship safety and operation monitoring and control,
event and alarm management, database management or message control is
structured in functional windows.

10.4.1.6 Onshore applications

Another example of the successful polymetric sensing of computer-aided
control and monitoring systems is connected with different marine terminals
(crude oil, light and low-volatility fuel, diesel, liquefied petroleum gas, grain)
and even with bilge water cleaning shops. Such sensing allows simultaneous
monitoring of practically all the necessary parameters using one sensor for
each tank, silo or other reservoir.
Namely, a set of the following characteristics is taken by a single
polymetric sensor:
• The level, separation level of the non-mixed media and the online control
of upper/lower critical levels and volume of the corresponding cargo in
each tank;
• The temperature field in the media, the temperature of the product at
particular points;
• Density, octane/cetane number of a fuel, propane-butane proportion in
petroleum gas, presence and percentage of water in any mixture or
solution (including aggressive chemicals – acids, alkalis, etc.);
As a result, a considerable increase of the measuring performance factor
ET SS/P SS was achieved in each application system (up to 4–6 times [8]) with
the essential concurrent reduction of number of instruments and measuring
channels of the monitoring system. All the customers (more than 50 objects
in Ukraine, Russia, Uzbekistan, etc.) reported commercial benefits after serial
SADCOTM systems deployment.

10.4.1.7 Special applications

Special applications of polymetric sensing of control and monitoring systems
were developed for the following: real-time remote monitoring of motor
lubrication oil production process (quantitative and qualitative control of the
232 Polymetric Sensing in Intelligent Systems

components depending on time and real temperature deviation during the pro-
duction process); real-time remote control of aggressive chemicals at a nuclear
power station water purification shop (quantitative and qualitative control of
the components depending on time and real temperature deviation during
the production process); water parameters control in the primary coolant
circuit at a nuclear power station in normal and post-emergency operation
state (fluidized bed-level control, pressure and temperature monitoring – all
in the conditions of increased radioactivity).

10.5 Conclusions
This paper has been intended as a general presentation of a rapidly developing
area of the promising transition from traditional on-board monitoring systems
to intelligent sensory decision support and control systems based on neoteric
polymetric measuring, data mining and holonic agent techniques. The area is
especially attractive to those researchers who are attempting to develop the
most effective intelligent systems by squeezing the maximum information
from the simplest and most reliable sensors by means of sophisticated and
effective algorithms.
In order for monitoring and control systems to become intelligent, not only
for exhibition demonstrations and show presentations, but for real industrial
applications, one needs to implement the leading-edge solutions for each and
every component of such a system.
Combining polymetric sensing, data mining and holonic agencies tech-
niques into one incorporated approach seems to be rather prospective if more
and more research will develop appropriate theory models and integrate them
into the practice as well.

References
[1] K. Hossine, M. Milton and B. Nimmo, ‘Metrology for 2020s’, Middle-
sex, UK: National Physical Laboratory, 2012, p. 28.
[2] S. Russell and P. Norvig, ‘Artificial Intelligence: A Modern Approach’,
Upper Saddle River, NJ: Prentice Hall, 2003.
[3] L. Monostori, J. Vancza and S. Kumara, ‘Agent-Based Systems for
manufacturing’, Annals of the CIRP, no. 55, 2 2006.
[4] P. Leitao, ‘Agent–based distributed manufacturing control: A state-
of-the-art survey’, Engineering Applications of Artificial Intelligence,
no. 22, pp. 979–991, 2009.
 References 233

[5] D. McFarlane and S. Bussman, ‘Developments in Holonic Production

Planning and Control’, Int. Journal of Production Planning and Control,
vol. 6, no. 11, pp. 522–536, 2000.
[6] F. Amigoni, F. Brandolini, G. D’Antona, R. Ottoboni and M. Somal vico,
‘Artificial Intelligence in Science of Measurements: From Measurement
Instruments to Perceptive Agencies’, IEEE Transactions on Instrumen-
tation and Measurements, vol. 3, no. 52, pp. 716–723, 6 2003.
[7] Y. Zhukov, ‘Theory of Polymetric Measurements’, in Proceedings of
1st International Conference: Innovations in Shipbuilding and Offshore
Technology, Nikolayev, 2010.
[8] Y. Zhukov, B. Gordeev and A. Zivenko, ‘Polymetric Systems: Theory
and Practice’, Nikolayev: Atoll, 2012, p. 369.
[9] Y. Zhukov, ‘Concept of Hypersignal and it’s Application in Naval
Cybernetics’, in Proceedings of 3-rd International Conference: Inno-
vations in Shipbuilding and Offshore Technology, Nikolayev, 2012.
[10] B. Marchenko and L. Scherbak, ‘Foundations of Theory of Mea-
surements’, Proceedings of Institute of Electrodynamics of National
Academy of Sciences of Ukraine, Electroenergetika, pp. 221–230, 1999.
[11] B. Marchenko and L. Scherbak, ‘Modern Concept for Development of
Theory of Measurements’, Reports of National Academy of Sciences of
Ukraine, no. 10, pp. 85–88, 1999.
[12] Y. Zhukov, ‘Fuzzy Algorithms of Information Processing in Systems
of Ship Impulse Polymetrics’, in Proceedings of 1st International
Conference: Problems of Energy saving and Ecology in Shipbuilding,
Nikolayev, 1996.
[13] Y. Zhukov, ‘Solitonic Models of Polymetric Monitoring Systems’, in
Proceedings of International Conference “Shipbuilding problems: state,
ideas, solutions”, USMTU, Nikolayev, 1997.
[14] Y. Zhukov, B. Gordeev and A. Leontiev, ‘Concept of cloned Polymetric
Signal and it’s application in Monitoring and Expert Systems’, in
Proceedings of 3-rd International Conference on Marine Industry, Varna,
2001.
[15] Y. Zhukov, B. Gordeev and A. Zivenko, ‘Polymetric sensing of intelli-
gent robots’, in Proceedings of IEEE 7th Int. Conf. on IDAACS, Berlin,
2013.
[16] A. Zivenko, A. Nakonechniy and D. Motorkin, ‘Level measurement
principles & sensors’, in Materialy IX mezinarodni vedecko-prackticka
conference “Veda a technologie: krok do budoucnosti - 2013”, Prague,
2013.
234 Polymetric Sensing in Intelligent Systems

[17] A. Zivenko, ‘Forming and pre-processing of the polymetric sig-

nal’, Collection of Scientific Publications of NUS, vol. 11, no. 439,
pp. 114–122, 2011.
[18] Y. Zhukov, ‘Instrumental Ship Dynamic Stability Control’, in Proceed-
ings of 5th IMAEM, Athens, 1990.
 11
Design and Implementation of Wireless
Sensor Network Based on Multilevel
Femtocells for Home Monitoring

D. Popescu, G. Stamatescu, A. Măciucă and M. Struţu

Department of Automatic Control and Industrial Informatics, University

“Politehnica” of Bucharest, Romania
Corresponding author: G. Stamatescu <[email protected]>

Abstract
An intelligent femtocell-based sensor network is proposed for home monitor-
ing of elderly or people with chronic diseases. The femtocell is defined as a
small sensor network which is placed into the patient’s house and consists of
both mobile and fixed sensors disposed on three layers. The first layer contains
body sensors attached to the patient that monitor different health parameters,
patient location, position and possible falls. The second layer is dedicated for
ambient sensors and routing inside the cell. The third layer contains emer-
gency ambient sensors that cover burglary events or toxic gas concentration,
distributed by necessities. Cell implementation is based on the IRIS family
of motes running the embedded software for resource-constrained devices,
TinyOS. In order to reduce energy consumption and radiation level, adaptive
rates of acquisition and communication are used. Experimental results within
the system architecture are presented for a detailed analysis and validation.

Keywords: wireless sensor network, ambient monitoring, body sensor

network, ambient assisted living.

11.1 Introduction
Recent developments in computing and communication systems applied to
healthcare technology give us the possibility to implement a wide range
of home-monitoring solutions for elderly or people with chronic diseases

Advances in Intelligent Robotics and Collaborative Automation, 235–256.


c 2015 River Publishers. All rights reserved.
236 Design and Implementation of Wireless Sensor Network Based

[1]. Thus, people may perform their daily activities while being constantly
under the supervision of the medical personnel. The indoor environment is
optimized such that the possibility of injury is minimal. Alarm triggers and
smart algorithms sent data to the right intervention units in regard to the
detected emergency [2].
When living inside closed spaces, small variables may be significant to
the entire person’s well-being. Therefore, the quality of the air, temperature,
humidity or the amount of light inside the house may be important parameters
[3]. Reduced costs, size and weight and energy-efficient operation of the
monitoring nodes, together with the more versatile wireless communications,
make the daily usage of the systems monitoring health parameters more
convenient. By wearing them, patients are free to move at their own will
inside the monitored perimeter, practically forgetting their presence. The goal
is to design the entire system operation for a long period of time without human
intervention and at the same time, triggering as few false alarms as possible.
Many studies investigated the feasibility of using several sensors placed
on different parts of the body for continuous monitoring [4]. Home care for the
elderly and chronic disease persons becomes an economic and social necessity.
With a growing population of ageing people and the health care prices rising
all over the world, we expect a great demand for home care systems [5, 6].
An Internet-based topology is proposed in [7] for the remote home-monitoring
applications that use a broker server, managed by a service provider. The
security risks from the home PC are transferred to the broker server and
removed, as the broker server is located between the remote-monitoring
devices and the patient’s house. An early prototype of a mobile health service
platform that was based on Body Area Networks is MobiHealth [8]. The
most important requirements of the developer for an e-health application are
size and power consumption, as considered in [9]. Also, in [10], a thorough
comprehensive study of the energy conservation challenges in wireless sensor
networks is carried out.
In [11], a wireless body area network providing long-term health moni-
toring of patients under natural physiological states without constraining their
normal activities is presented.
Integrating the body sensors with the existing ambient monitoring network
in order to provide a complete view of the monitored parameters is one of the
issues discussed in this paper. Providing a localization system and a basic
algorithm for event identification is also part of our strategy to fulfill all
possible user requests. Caregivers also value information about the quality
 11.2 Network Architecture and Femtocell Structure 237

of air inside the living area. Many false health problems are usually related to
the lack of oxygen or high levels of CO or CO2 .
The chapter is organized as follows: Section 2 provides an overall view
on the proposed system architecture and detailed insight into the operation
requirements for each of the three layers for body, ambient and emergency
monitoring. Section 3 introduces the main criteria for efficient data collection
and a proposal for an adaptive data rate algorithm for both the body sensor
network and the ambient sensor network. This has the aim of reducing the
amount of data generated within the networks, considering processing, stor-
age, and energy requirements. Implementation details and experimental results
are evaluated in Section 4, where the path is set for long-term deployment and
validation of the system. Section 5 concludes the chapter and highlights the
main directions for future work.

11.2 Network Architecture and Femtocell Structure

The proposed sensor network architecture is based on hybrid femtocells. A
hybrid femtocell contains sensors which are grouped based on their functional
requirements, mobility and energy consumption characteristics in three layers:
the body sensor network (BSN), the ambient sensor network (ASN) and the
emergency sensor network (ESN), as presented in Figure 11.1. Coordination
is implemented through a central entity called the femtocell management node
(FMN) which aggregates data from the three layers and acts as a interface to
the outside world by means of the internet. Communication between different
components can be made using wireless technology and radio compatible
fiber optic. Table 11.1 lists the high-level characteristics of the two low-power
wireless communication standards often used in home-monitoring scenarios:
IEEE 802.15.1 and IEEE 802.15.4. This highlights an important design
trade-off in the deployment of a wireless sensor network for home monitoring.

Table 11.1 Main characteristics of IEEE 802.15.1 and 802.15.4

IEEE Standard 802.15.1 802.15.4
Frequency ISM - 2.4GHz ISM/868/915 MHz
Data rate 1 Mbps 250 kbps
Topology Star Mesh
Scalability Medium High
Latency Medium High
Interference mit. FHSS/DHSS CSMA/CA
Trademark Bluetooth ZigBee
238 Design and Implementation of Wireless Sensor Network Based

Figure 11.1 Hybrid femtocell configuration.

While IEEE 802.15.4 and ZigBee enable large dense networks with complex
mesh topologies, the use of Bluetooth can become an advantage in applications
with higher data-rate requirements and low latency.
The layer characteristics and functionalities are further elaborated upon.

11.2.1 Body Sensor Network

The body sensor network functional design includes battery charging nodes
along with user-friendly construction and operation. In an idealised case, the
size and weight would go unnoticed, immediately or after a short accomoda-
tion period, not disturbing the patient or elderly person when wearing them.
Nodes communicate using a low-power wireless communication protocol for
very short distance data transmission and reception e.g. ZigBee or Bluetooth,
depending on the data streaming rate of the application and energy resources
on the node. Very low-energy consumption is an essential design criteria as
changing the batteries on a daily basis becomes stressful on the long term
and ideally the nodes would be embedded into wearable technology. Some
on the sensed parameters for the BSN include dual- or tri- axial accelerom-
eters, blood pressure, ECG, blood oxygen saturation, heart rate and body
temperature.
 11.2 Network Architecture and Femtocell Structure 239

The main events that should be detected by the BSN cover fall detection,
activity recognition and variations in investigated parameters corresponding
to alert and alarm levels.

11.2.2 Ambient Sensor Network

The ambient sensor network is comprised of a series of fixed measurement
nodes, placed optimally in the target area/environment as to maximize sensing
coverage and network connectivity through a minimum number of nodes.
The low-power communication operates on longer communication links
than the BSN and has to be robust to main phenomena affecting indoor
wireless communication, such as interference, reflections and assymetric links.
Though fixed node placement provides more flexibility when choosing the
energy source, battery operation and low maintenance, enabled by low-energy
consumption is preferred to mains power. For example, in the ASN, the router
nodes which are tasked with redirecting much of the network traffic in the
multi-hop architecture can be operated from the main power line.
Within the general framework, the ASN can serve as an external system
for patient monitoring through a localization function based on link quality
and signal strength. The monitored parameters include ambient temperature,
humidity, barometric pressure and light. These can be evaluated individually
or can serve as input data to a more advanced decision support system which
can correlate the evolution of indoor parameters with the BSN data from the
patient to infer the conditions for certain diseases. Some nodes of the ASN
might include complex sensors like audio and video capture giving a more
detailed insight into the patient’s behaviour. Their current use is somewhat
limited by the additional computing and communication ressources needed to
accomodate the sensors into current wireless sensor network architectures as
well as by privacy and data-protection concerns.

11.2.3 Emergency Sensor Network

The multi-level femtocell reserves a special role for the emergency sensor
network which can be considered as a particular case of ASN tasked with quick
detection and reaction to life or property threats through embedded detection
mechanisms. The sensing nodes are fixed and their placement is well suited to
the specifics of the measurement process. As an example, the house could be
fitted with gas sensors in the kitchen next to a stove, carbon dioxide sensors
would be placed in the bedroom and passive infrared sensor and pressure
sensors would be fitted next to doors and windows. As the operation of the
240 Design and Implementation of Wireless Sensor Network Based

ESN is considered critical, energy-efficient operation becomes a secondary

design criteria. The nodes should have redundant energy sources, both batteries
and mains power supply, and redundant communication links. For example, a
wireless low latency communication protocol with simple network topology
to minimize overhead and packet loss can be used as a main interface with the
possibility of switching to a wired interface or using the ASN infrastructure
in certain situations.
The main tasks of the ESN are to detect dangerous gas concentrations
posing threats of explosion and/or intoxication and to detect intruders into the
home such as burglars.

11.2.4 Higher-level Architecture and Functional Overview

One of the features of the hybrid femtocell is that the body sensor always
interacts with the closest ambient sensor node, Figure 11.2, in order to send
data to the gateway. This function reassures us that the sensors attached to the
monitored person are always connected and the energy consumption is optimal
because the distance between the person and the closest router is minimal. This
feature can also be used as a mean of localization. The person wearing the

Figure 11.2 Data and information flow within and outside the femtocell.
 11.2 Network Architecture and Femtocell Structure 241

body sensor will be always connected to the closest router. By using fixed
environmental sensors with own ID and previously known positions, we can
determine which room is presently used by the inhabitant. In order to have
an accurate localization of the patient, an attenuation map of the sensors from
each room must be created. Considering that patient is localization in a certain
room of the home, by the closest ASN, is not accurate this could happen due
to the following scenario: lets suppose that we have an ASN located in the
bedroom, situated in the left side of the room. In the right part of the room,
we have the door to the living room, and near this door we have the ASN for
the living room. If the patient is located in the bedroom, but very close to the
door, it will have as closest ASN the one from the living room, but he/she is
situated in the bedroom. In order to avoid this localization error, we introduce
the attenuation map of the sensors. Every ASN that localizes the BSN on the
patient will transmit an attenuation factor. This way, using the attenuation
factors from each sensor, we can localize the patient very accurately. In our
example, if the bedroom ASN has a 10% factor, and the living room ASN has
a 90% factor, using the attenuation map, we localize the patient as being in
the bedroom, but very close to the door between the two rooms.
The position of a hybrid femtocell in the large wireless sensor network
system is presented in Figure 11.3. Its main purpose is to monitor and interpret
data, sending specific alarms when required. The communication between the
femtocells and the network is based on internet. The same method is used for

Figure 11.3 System architecture.

242 Design and Implementation of Wireless Sensor Network Based

the communication between the network administrator and hospital, police,

fire station or relatives.
In Figure 11.4, the ambient sensors’ spatial placement and reference
deployment inside the patients apartment is showcased. This fixed topology
is consistent with the routing role of the ambient sensors in regard to the
considered mobile body sensors. The goal of the system is to collect relevant
data for reporting and processing. Therefore, achieving a very high sensory
and communication coverage is among our main objectives.
The entire monitoring system benefits from a series of gas sensor modules
strategically placed throughout the monitored space. Thus, the Parallax-
embedded gas-sensing module for CO2 is designed to allow a microcontroller
to determine when a preset carbon dioxide gas level has been reached or
exceeded. Interfacing with the sensor module is done through a 4-pin SIP
header and requires two I/O pins from the host microcontroller. The sensor
module is mainly intended to provide a means of comparing carbon dioxide
sources and being able to set an alarm limit when the source becomes
excessive [4].
A more advanced system for indoor gas monitoring, MICARES, is shown
in Figure 11.5. It consists of an expansion module for a typical wireless sensor
network platform with the ability to measure CO, CO2 and O3 concentrations
and perform on board duty-cycling of the sensors for low energy as well as
measurement compensation and calibration. Using this kind of platform, data

Figure 11.4 Sensor deployment example.

 11.3 Data Processing 243

Figure 11.5 Embedded carbon dioxide sensor node module [12].

can be reliably evaluated locally and relayed to the gateway as small packets
of timely information. The module can be either powered by mains or draw
power from the energy source of the host node. Also, it can operate as a
stand-alone device through ModBus serial communication with other infor-
mation processing systems.
In case an alert is triggered, the information is automatically processed
and the alarm is sent to the specific intervention factor (hospital, police, fire
department, etc.). These have the option of remotely accesing the fem-tocell
mangement node in order to verify that the alarm is valid and act by dispatching
an intervention team to solve the issue. Subsequently, all the alarms generated
over time are classified and stored in a central database for reporting purposes.

11.3 Data Processing

Data harvesting and processing is performed at the femtocell level, so the
whole network is not flooded by useless data from all femtocells. Considering
the main goals of conserving node energy, while taking into account the limited
storage and processing capabilities of each node, we proposed a strategy to
dynamically adapt the sampling rate of the investigated parameters for the
BSN and ASN. The ESN is considered crtical home infrastructure and it is
not affected by this strategy, with its own specific challenges for reliability,
periodic sensor maintenance and calibration.
244 Design and Implementation of Wireless Sensor Network Based

Adaptive data acquisition is implemented by framing collected data into

a safety interval given by [Vmin , Vmax ]. While the raw values are relatively
close to the interval center V = 12 (Vmax – Vmin ) and the recent variation of
the time series, given by the derivative, is low over a exponential weighted
time horizon, then the data acquisition rate is lowered periodically towards a
lower bound. When collected measurements start to vary, data acquisition is
increased steeply in order to capture the significant evene, then a stability
period is observed and it begins to be lowered again. Though local data
storage on the node can be exploited, we prefer to synchornize collection with
transmission of the packets through the radio interface in order to conserve data
freshness which can prove important in order to effectively react to current
events. Energy level must be considered in order to prevent the reception
of corrupted data and to avoid unbalanced node battery depletion e.g. in the
case of routing nodes in the network. This represents a validation factor of
the data and is automatically transmitted together with the parameter value.
Energy-aware policies can be implemented at the femotcell level to optimize
battery recharge, in the BSN case, and battery replacement, for the ASN.
Evaluation of the parameter derivate takes place for attaining a variable
rate of acquisition. The values obtained by calculating the derivate also
help us to decide what type of event has happened. This can be done by
building a knowledge base during system commissioning and initial operating
period which can be used to relate modifications in observed parameter
to trained event classes. These can go from trivial associations like high
temperature and low oxygen/smoke meaning a fire, to subtle patient state of
health detection by changing vital signs and aversion to light. The algorithm
described can be summarized as follows:
Data: measured values, measurement derivative, energy level
Result: adaptive data rate initialize;
while
 network is operational do
 collect measurement values;

 check distance from interval center;

 compute time series derivative;

 if high sudden variation then

  increase data rate;

 else
 
  lower data rate;

 end
end
Algorithm 1: Adaptive Data Collection for Home Monitoring
 11.4 Experimental Results 245

11.4 Experimental Results

In order to bring value to the theoretical system architecture proposed, two
experiments have be devised and implemented. They cover the body and
ambient sensor layers of the multi-level femtocell for home monitoring. The
main operational scenario that was considered involves an elderly patient,
living alone at home in an appartment or house with multiple rooms. A
caregiver is available on call as well as a permanent connection to the
emergency services exists, with the possibility on alerting close relatives
in the process. As functional requirements, we target activity monitoring
and classification by body-worn accelerometers and ambient measurement
of humidity, temperature, pressure and light, along with their correlations.
The first experiment is performed as part of the body sensor network.
Therefore, two accelerometers with two axes were placed on the patient, one
on the right knee and the other one on his left hip, as shown in Figure 11.6.
The sensors are automatically assigned a unique value ID. Therefore, in our

Figure 11.6 Accelerometer node placement on the patient’s body.

246 Design and Implementation of Wireless Sensor Network Based

experiment, the sensor situated on the knee has ID104, and the other one,
placed on the left hip, ID103. In order to overcome the hardware limitations,
in a tridimensional representation, the representation of the 3 axis using those
two accelerometers is the following:
• X axis: front back, X axis of node 104;
• Y axis: left right, X axis of node 103;
• Z axis: bottom up, Y axis of both nodes.
The following activities were executed during the experiment: standing, sitting
on a chair, standing again and slow walking from bedroom to living. Data
acquisition has been performed using MOTE-VIEW [13]. This is an interface,
client layer, between a user and a deployed network of wireless sensors.
Besides this main function, the user can change or update the individual
node firmware, switch from low-power to high-power mode and set the radio
transmit power. Collected data is stored in a local database and can be accessed
remotely by authorized third parties.
Multiple experiments have been conducted in order to determine
the trust values. Therefore, in the X and Y axis charts presented in
Figures 11.7 and 11.8, the readings outside the green lines represent that an
event occurred, obtained by thresholding. The following events have been
monitored during our experiment in this order: standing, sitting on a chair,

Figure 11.7 X Axis experiment acceleration with thresholding.

 11.4 Experimental Results 247

Figure 11.8 Y Axis experiment acceleration with thresholding.

standing again and slow walking from bedroom to living. The used sensors
are ADXL202E [14].
These are low-cost, low-power, complete 2-axis accelerometer with a
digital output, all on a single monolithic IC. They are an improved version
of the ADXL202AQC/JQC. The ADXL202E will measure accelerations
with a full-scale range of 2 g. The ADXL202E can measure both dynamic
acceleration (e.g., vibration) and static acceleration (e.g., gravity). The outputs
are analog voltage or digital signals whose duty cycles (ratio of pulse width
to period) are proportional to acceleration. The duty cycle outputs can be
directly measured by a micro- processor counter, without an A/D converter
or glue logic. The duty cycle period is adjustable from 0.5 ms to 10 ms via a
single resistor (RSET).
The radio base station is made up of an IRIS radio/processor board con-
nected to a MIB520 USB interface board via the 51-pin expansion connector.
The interface board uses a FTDI chip and provides two virtual COM ports
to the host system. COMx is used for programming the connected mote and
COMx+1 is used by middleware applications to read serial data. The used
network protocol is called XMesh, owned by Crossbow, based on the standard
802.15.4. The Crossbow sensor networks can run different power strategies,
each of these strategies being a trade-off between power, data rates and
latency.
XMesh-HP [15] is the best approach for systems that have continuous
power, offering the highest message rate, usually proportional to the band
rate of the radio. Radios and processors are continually powered, consuming
248 Design and Implementation of Wireless Sensor Network Based

Table 11.2 XMesh power configuration matrix

Power Mode MICA2 MICA2DOT MICAz
XMesh-HP X X X
XMesh-LP X X Async. only
XMesh-ELP X X X

Table 11.3 XMesh performance summary table

Parameter XMesh-HP XMesh-LP XMesh-ELP
Route update interval 36 sec. 360 sec. 36 sec. (HP) / 360 sec.
(LP)
Data message rate 10 sec. typ. 180 sec., typ. N/A
Mesh formation time 2–3 RUI X X
Average current usage 20–30 mA 400 uA 50 uA

between 15 and 30 mA depending on the type of the Mote. Route Update

Messages and Health Messages are sent at a faster rate which decreases the
time it takes to form a mesh or for a new mote to join the mesh.
XMeshLP [15] is used for battery-operated systems that require
multi-month or multi-year life. It can run either time-synchronized or asyn-
chronous. The best power efficiency is achieved with time synchronization
within 1 msec. The motes wake-up, typically 8 times/second, time synchro-
nized, for a very short interval to see if the radio is detecting any signal over the
noise background. If this happens, the radio is kept alive to receive the signal.
This action usually results in a base level current of 80 µA. The total average
current depends on the number of messages received and transmitted. If data is
transmitted every three minutes, the usual power in a 50-node mesh is around
220 µA. XMesh-LP can be configured for even lower power by reducing the
wake-up periods and transmitting at lower rates. Also, route update intervals
are set at a lower rate to conserve power, resulting in a longer mesh formation
time.
XMesh-ELP [15] is only used for leaf nodes that communicate with parent
nodes running XMesh-HP. A leaf node is defined as a node that does not
participate in the mesh; it never routes messages from child motes to parent
motes. The results of the ELP version are very low power because the mote
does not need to use the time synchronized wake-up mode to check for radio
messages. The mote can sleep for very long times, this way maintaining its
neighborhood list to remember which parents it can select. If it does not get a
link-level acknowledgement when it transmits to a parent, it will find another
parent and so on. This operation can happen very quickly or might take some
time if the RF environment or mesh configuration has changed considerably.
 11.4 Experimental Results 249

Because of their small size, nodes can be easily concealed into the
background, interfering as little as possible with the user’s day-to-day routine.
We have also the possibility to set the sampling rate at a suitable level in order
to achieve low-power consumption and by this a long operating range without
human intervention [16].
Our infrastructure also offers routing facilities increasing the reliability
of our network by self-configuring into a multihop communication system
whenever direct links are not possible. After experimenting with different
topologies, we achieved a working test scenario which consisted in a four-
level multihop communication network which is more than we expect to be
necessary in any of our deployment locations.
Extensive experimental work has been carried out for the ambient sensor
layer of the system based on MTS400 IRIS sensor nodes. One of the reasons
for choosing this specific board has been that it provided the needed sensors
in order to gather a variety of environmental parameters, like temperature,
humidity, relative pressure and ambient light. The experimental deployment
consists of three measurement nodes organized in a true mesh topology in a
testing indoor environment. These aim at modeling a real implementation in
the patient’s home and were taken over the course of a week-long deployment.
In order to assure accounting to uneven sampling from the sensor nodes, we
use as reference time MATLAB serial time units which are converted from
conventional time stamps entries into the MOTE-VIEW database of the form
dd.mm.yyyy HH:MM:SS.
In Figure 11.9(a), the evolution of the humidity parameter measured by
indoor deployed sensors can be seen. The differences account for node place-
ment in the different rooms and exposure to windows and doors. Subsequent
processing can lead to computing average values and to other correlations with
ambient and body parameters and an intelligent information system which
can associate variations in ambient humidity and temperature to influences
on chronic disease. Figure 11.9(b) illustrates temperature variations obtained
from the ambient sensor network. These reflect the circadian evolution of the
measured parameter and show the importance of correct node placement and
data aggregation within the sensor network.
Barometric pressure (Figure 11.10(a)) is also observed by the sensor
network over the testing period. This is the parameter that is least influenced
by node placement and more by general weather trends. As differences
between individual sensor node values of a few percentage points, these can be
attributed to sensing element calibration or local temperature compensation.
Aggregating data also in this case can lead to higher-quality measurements.
250 Design and Implementation of Wireless Sensor Network Based

Figure 11.9 Measurement data: humidity (a); temperature (b).

 11.4 Experimental Results 251

Figure 11.10 Measurement data: barometric pressure (a); light (b).

252 Design and Implementation of Wireless Sensor Network Based

Ambiental light values, suitable for feeding data to an intelligent mood lighting
system, are shown in Figure 11.10(b). The light sensor saturates in full daylight
at around 1850 lux and quickly responds to variations in the measured light.
The most important period of the day are dawn and dusk where the information
provided can assure a smooth transition from artificial to natural light and
reverse.
Data flow coming from the ambient sensor network is processed in multiple
steps: at the source node, in the network e.g. through aggregation or sensor
fusion, and at the home server or gateway level. Each steps converts raw data
to higher-level pieces of information which can be more efficiently operated
with and become meaninful through correlation and interactive visualization.
Efficient management of this infomation is critical to correct operation of the
home-monitoring system. Alerts and alarms have to be reliable and build trust
among the end user leading to widespread acceptance whilst assuring a high
level of integrity, security and user privacy.
Table 11.4 summarizes the aggregated experimental deployment values
for the three nodes over the investigated period. Minimum, maximum, average
and the standard deviations of the collected time series for each of the measured
ambiental parameters are listed.
Making effective use of the large quantities of data generated by the layers
of the femtocell structure, represented by the individual sensor networks,
poses a significant challenge. One idea is to apply computational intelligence

Table 11.4 Ambiental Monitoring Summary

Node ID 6692 6782 6945
Humidity [%] min 22.5 27.8 35.8
max 41 42.8 50.5
avg 33.39 34.62 42.26
stdev 4.76 3.22 2.58
Temperature [degC] min 23.08 23.78 23.4
max 32.98 29.01 27.48
avg 26.24 25.45 24.66
stdev 2.7 1.1 0.75
Pressure [mbar] min 997.83 997.34 998.24
max 1009.6 1009.3 1010.2
avg 1003.3 1002.2 1003.5
stdev 2.85 2.7 2.7
Light [lux] min 0 0 0
max 1847.1 1847.1 1847.1
avg 952.18 705.87 414.05
stdev 880.32 801.35 517.96
 11.5 Conclusion 253

techniques, either in a centralized manner at the gateway level or in a dis-

tributed fashion where computing taks are spread among the nodes according
to a dynamic strategy. An example is given for time series prediction on
the temperature data collected by the ambient sensor network using neural
networks. As a tool, the MATLAB technical computing environment can be
used for modeling, testing and validation and embedded code generation at
the gateway level.

11.5 Conclusion
The chapter introduced a system architecture composed of a smart hybrid
sensor network for indoor monitoring using a multilayer femtocell for ubiq-
uitous intelligent home monitoring. The main components of the system are
three low-level wireless sensor networks: body, ambient and emergency, along
with a central coordination entity named the femtocell management node.
This also acts as gateway towards the internet and the interested stakeholders
in an ambient-assisted living scenario. It has been argued that efficient data
collection and processing strategies along with robust networking protocols
can enable seemless integration into the patient’s home and bring added value
to home care whilst reducing overall medical and assitance costs. Recent
advances in miniaturization of discrete electronic components and systems,
along with enhanced computing and communication capabilities of intelligent
home device offer a good opportunity in this application area. This can be
exploited for both reasearch and development in the field of ambient-assisted
living to increase quality of life while dealing with increased medical costs.
The main experimental focus was on the body sensor network layer and
the ambient sensor layer and experimental deployment and implementation
has been illustrated. First, body-worn wireless accelerometers were used
to detect and classify human activity based on time-domain thresholding.
Second, extensive results from a medium-term deployment of an ambient
sensor network were illustrated and discussed. This had, as a main purpose,
the collection of ambient parameters, like temperature, humidity, barometric
pressure and ambient light while observing network protocol behaviour. The
main conclusion is that wireless sensor network systems and protocols offer
a reliable option for deployment in home monitoring, given the specific
challenges of indoor environments.
Plans for future development have been established on three main paths.
One direction includes extending the system with more measured parameters
through additional sensor integration with the wireless nodes. The focus here
254 Design and Implementation of Wireless Sensor Network Based

would be on the body sensor network side where a deep insight into the
patient’s well-being and health status can be gained. Also, while raw data and
machine-learning algorithms can provide high-confidence recommendations
and alerts to the caregivers, data visualization in the home and for the
patient should not be neglected. This can be done by developing adaptive
and intuitive interfaces for the patient or elderly person which enhance
acceptance of the system. The quality and accuracy of the expected results has
to be increased by integrating state-of-the-art sensing, signal processing and
embedded computing hardware along with the implementation of advanced
methods for experimental data processing.

References
[1] A. Măciucă, D. Popescu, M. Struu, and G. Stamatescu. Wireless
sensor network based on multilevel femtocells for home monitoring.
In 7th IEEE International Conference on Intelligent Data Acquisition
and Advanced Computing Systems: Technology and Applications,
pages 499–503, 2013.
[2] M. Fahim, I. Fatima, Sungyoung Lee, and Young-Koo Lee. Daily life
activity tracking application for smart homes using android smart-
phone. In Advanced Communication Technology (ICACT), 2012 14th
International Conference on, pages 241–245, 2012.
[3] M. Smolen, K. Czopek, and P. Augustyniak. Non-invasive sensors based
human state in nightlong sleep analysis for home-care. In Computing in
Cardiology, 2010, pages 45–48, 2010.
[4] Yibin Hou, Na Li, and Zhangqin Huang. Triaxial accelerometer-based
real time fall event detection. In Information Society (i-Society), 2012
International Conference on, pages 386–390, 2012.
[5] Andrew D. Jurik and Alfred C. Weaver. Remote medical monitoring.
Computer, 41 (4):96–99, 2008.
[6] P. Campbell. Population projections: States, 1995–2025. Technical
report, U.S. Census Bureau, 1997.
[7] Chao-Hung Lin, Shuenn-Tsong Young, and Te-Son Kuo. A remote
data access architecture for home-monitoring health-care applications.
Medical Engineering and Physics, 29(2):199–204, 2007.
[8] Aart Van Halteren, Dimitri Konstantas, Richard Bults, Katarzyna Wac,
Nicolai Dokovsky, George Koprinkov, Val Jones, and Ing Widya. Mobi-
health: ambulant patient monitoring over next generation public wireless
 References 255

networks. In in Demiris, G. (Ed.): E-health: Current Status and Future

Trends, IOS, pages 107–122. Press, 2004.
[9] DD Vergados, D Vouyioukas, NA Pantazis, I Anagnostopoulos, DJ
Vergados, and V Loumos. Decision support algorithms and optimization
techniques for personal homecare environment. IEEE IntÄôl. Special
Topic Conf. Information Technology in Biomedicine (ITAB 2006), 2006.
[10] S. Tarannum. Energy conservation challenges in wireless sensor net-
works: A comprehensive study. Wireless Sensor Network, 2010.
[11] Emil Jovanov, Aleksandar Milenkovic, Chris Otto, and PietC de Groen.
A wireless body area network of intelligent motion sensors for com-
puter assisted physical rehabilitation. Journal of NeuroEngineering and
Rehabilitation, 2(1):1–10, 2005.
[12] G. Stamatescu, D. Popescu, and V. Sgarciu. Micares: Mote expansion
module for gas sensing applications. In Intelligent Data Acquisition and
Advanced Computing Systems (IDAACS), 2013 IEEE 7th International
Conference on, volume 01, pages 504–508, Sept 2013.
[13] M. Turon. Mote-view: A sensor network monitoring and management
tool. In Embedded Networked Sensors, 2005. EmNetS-II. The Second
IEEE Workshop on, May 2005.
[14] T. Al-ani, Quynh Trang Le Ba, and E. Monacelli. On-line automatic
detection of human activity in home using wavelet and hidden markov
models scilab toolkits. In Control Applications, 2007. CCA 2007.,
Oct 2007.
[15] Crossbow Inc. XMesh User’s Manual, revision c edition, March 2007.
[16] G. Stamatescu and V. Sgarciu. Evaluation of wireless sensor network
monitoring for indoor spaces. In Instrumentation Measurement, Sensor
Network and Automation (IMSNA), 2012 International Symposium on,
Aug 2012.
[17] Parallax Inc. Co2 gas sensor module (27929) datasheet. Technical report,
Parallax Inc., 2010.
 12
Common Framework Model
for Multi-Purpose Underwater Data
Collection Devices Deployed with Remotely
Operated Vehicles

M.C. Caraivan1 , V. Dache2 and V. Sgarciu2

1 Faculty of Applied Sciences and Engineering, University Ovidius
of Constanta, Romania
2 Faculty of Automatic Control and Computers, University Politehnica

of Bucharest, Romania
Corresponding author: M.C. Caraivan <[email protected]>

Abstract
This paper is following the development of real-time applications for marine
operations focusing on modern modelling and simulation methods and
presents a common framework model for multi-purpose underwater sensors
used for offshore exploration. It is addressing deployment challenges of
underwater sensor networks called by the authors “Safe-Nets” by using
Remotely Operated Vehicles (ROV).

Keywords: Remotely Operated Vehicles, ROV, simulation, testing, object

modelling, underwater component, oceanographic data collection, pollution.

12.1 Introduction
The natural disaster following the explosion of BP Deepwater Horizon
offshore oil-drilling rig in the Gulf of Mexico has raised questions more
than ever about the safety of mankind’s offshore oil-quests. For three months
in 2010, almost 5 million barrels of crude oil formed the largest accidental

Advances in Intelligent Robotics and Collaborative Automation, 257–294.


c 2015 River Publishers. All rights reserved.
258 Common Framework Model for Multi-Purpose Underwater Data Collection

marine oil spill in the history of the petroleum industry. The frequency of
maritime disasters and their effects appear to have dramatically increased
during the last century [1], and this draws considerable attention from decision
makers in communities and governments. Disaster management requires the
collaboration of several management organizations resulting in heterogeneous
systems. Interoperability of these systems is fundamental in order to assure
effective collaboration between different organizations.
Research efforts in the exploration of offshore resources have increased
more and more during the last decades, thus contributing to greater global
interest in the area of underwater technologies. Underwater sensor networks
are going to become in the nearby future the background infrastructure for
applications which will enable geological prospection, pollution monitor-
ing, and oceanographic data collection. Furthermore, these data collection
networks could in fact improve offshore exploration control by replacing
the on-site instrumentation data systems used today in the oil-industry
nearby well heads or in well-control operations, e.g. using underwater
webcams which can provide important visual data aid for surveys or for
offshore drilling explorations. These facts lead to the idea of deploying
multi-purpose underwater sensor networks along-side with oil companies’
offshore operations. The study is trying to show the collateral benefits of
deploying such underwater sensor networks and we address state-of-the-art
ideas and possible implementations of different applications like military
surveillance of coastal areas, assisting navigation [2] or disaster prevention
systems – including earthquakes and tsunami detection warning alarms in
advance – all in order to overcome the biggest challenge of development: the
cost of implementation.
It is instructive to compare current terrestrial sensor network practices
to underwater approaches: terrestrial networks emphasize low-cost nodes
(around a maximum of US$100), dense deployments (at most a few 100m
apart) and multi-hop short-range communication. By comparison, typical
underwater wireless communications today are expensive (US$10.000 per
node or even more), sparsely deployed (a few nodes, placed kilometres apart),
typically communicating directly to a “base-station” over long-distance ranges
rather than with each other. We seek to reverse the design points which make
land networks so practical and easy to expand and develop, so underwater
sensor nodes that can be inexpensive, densely deployed, and communicating
peer-to-peer [3].
Multiple Unmanned orAutonomous Underwater Vehicles (UUVs,AUVs),
equipped with underwater sensors, will also find application in exploration
 12.1 Introduction 259

of natural undersea resources and gathering of scientific data in collaborative

monitoring missions. To make these applications viable, there is a need
to enable underwater communications among underwater devices. Ocean
Sampling Networks have been experimented in the Monterey Bay area,
where networks of sensors and AUVs, such as the Odyssey-class AUVs,
performed synoptic, cooperative adaptive sampling of the 3D coastal ocean
environment [4]. While offshore constructions’ number grows, we should be
able to implement auxiliary systems that allow us to better understand and
protect the ocean surface we are building on. We will be using Remotely
Operated Vehicles (ROVs) and a VMAX-Perry Slingsby ROV Simulator
with scenario development capabilities to determine the most efficient way
of deploying our underwater sensor networks, which we called “Safe-Nets”,
around offshore oil drilling operations, including all types of jackets, jack-ups,
platforms, spars or any other offshore metallic or concrete structure.
The ability to have small devices physically distributed near offshore
oil-fields’ operations brings new opportunities to observe and monitor
micro-habitats [5], structural monitoring [6] or wide-area environmental
systems [7]. We even began to imagine a scenario where we can expand these
sensor networks in order to slowly and steadily develop a global “WaterNet”,
which could be an extension of the Internet on land. In the same manner which
allowed Internet networks on land to develop by constantly adding more and
more nodes to the network, we could allow information to be transmitted
from buoy to buoy in an access-point like system. These small underwater
“Safe-Nets” could be joined together and the network could expand into a
global “Water-Net” in the future, allowing data to be sent and received to and
from shore bases. Of course, today, we can see considerable less kilobytes
of data to be sent and received at first, but the main advantages would be
in favour of disaster prevention systems. “Safe-Nets” for seismic activity
and tsunami warning systems alone can represent one of the reasons for
underwater network deployment, which are quite limited today compared to
their counterparts on land. We propose a model of interoperability in case of
a marine pollution disaster for a management system based upon Enterprise
Architecture Principles.
If we keep in mind that the sea is a harsh environment, where reliability,
redundancy and maintenance-free equipment are most desirable objectives, we
should seek the methods and procedures for keeping the future development
in a framework that should be backwards compatible with any other sensor
nodes already deployed. In order to comply with the active need for upgrading
to future technologies, we have thought of a common framework model with
260 Common Framework Model for Multi-Purpose Underwater Data Collection

multiple layers and drawers for components which can be used for different
purposes, but mainly for underwater data collection and monitoring. This
development using Enterprise Architecture Principles is sustainable through
time, as it is backed up by different solutions to our research challenges,
such as power supply problem, fouling corrosion, self-configuration, self-
troubleshooting protocols, communication protocols and hardware methods.
Two-thirds of the surface of Earth is covered by water and as history
proved it, there is a constantly increasing number of ideas to use this space.
One of the most recent is perhaps moving entire buildings of servers - Google’s
Data-Centres [8] overseas, literally, because of their cooling needs which
nowadays are tremendous. These produce a heat footprint clearly visible even
from satellites and by transferring them to the offshore environment, their
overheating problems would have cheaper cooling methods which could be
satisfied by the ocean’s seawater almost constant temperature.Also, we discuss
the electrical power supply possibilities further in the following chapter.

12.2 Research Challenges

We seek to overcome each of the design challenges that prohibited underwater
sensor network development, especially by designing a common framework
with different option modules available to be installed. If having a hardwire
link at hand, by attaching these devices to offshore construction sites or to
autonomous buoys, we could provide inexpensive sensors by using the power
supply or communication means from that specific structure. We are looking
forward to develop a variety of option modules for our common framework
to be used for all types of underwater operations, which can include the
instrumentation necessities nearby wellheads and drill strings or any type of
oceanographic data collection, therefore becoming a solution at hand for any
given task. This could provide the financial means of deploying underwater
Safe-Nets, especially by tethering to all the offshore structures or exploration
facilities which need different underwater data collection by their default
nature.

12.2.1 Power Supply

Until now, only battery power was mainly used in underwater-based sensor
deployments. The sensors were deployed and shortly afterwards were recov-
ered. In our case, the close proximity to oil-rig platforms or other offshore
constructions means already existing external power sources: diesel or gas
 12.2 Research Challenges 261

generators, wind turbines, gas pressure turbines. We can overcome this design
issue with cable connections to jackets or to autonomous buoys with solar
panels which are currently undergoing research [9, 10].
Industrial applications such as oil-fields and production lines use extensive
instrumentation, sometimes with the need of a video-feedback from the
underwater operations site. Considering the depths at which these cameras
should operate, there is also an imperative need for proper lighting of the
area; therefore we can anticipate that these nodes will be tethered in order to
have a power source at hand.
Battery power problems which in our case can be overcome not only by
sleep-awake energy efficient protocols [11–13], but also by having connectiv-
ity at hand to other future system types of producing electricity from renewable
resources, like wave energy converter units according to the European project
Aquatic Renewable Energy Technologies (Aqua-RET) [14]:
• Attenuator-type Figure 12.1: Pelamis Wave Energy Converter [15];
• Axial symmetric absorption points as in Figure 12.2: WaveBob [16],
AquaBuoy, OE Buoys [17] or Powerbuoy [18];
• Wave-level oscillation converters: completely submerged Waveroller or
surface Oyster [19];
• Overtopping devices Figure 12.3: Wave Dragon [20];

Figure 12.1 Pelamis wave converter Orkney, U.K.

262 Common Framework Model for Multi-Purpose Underwater Data Collection

• Submersible differential pressure devices Figure 12.4: Archimedes

Waveswing [21];
• Oscillating Water Column (OWC) devices.

Figure 12.2 Axial symmetric absorption buoy.

Figure 12.3 Wave Dragon - Overtopping devices principle.

Figure 12.4 Archimedes Waveswing (AWS).

 12.2 Research Challenges 263

Figure 12.5 Wind farms in North Sea.

In addition, we are considering also other types of clean energy technology

production systems at sea:
• Wind farms: usually the wind speed at sea is far greater than on land,
however, by comparison to its land counterpart, offshore wind turbines
are harder to install and need more technical and financial efforts. The
distance to land, water depth and sea floor structure are factors that need
to be taken into consideration for Aeolian projects at sea. The first project
for an offshore wind farm was developed in Denmark in 1991;
• Oceans’ thermic energy by using the temperature difference between
surface and depth waters, which needs to be at least 20o C at less than
100m from sea surface. These desiderates are usually full-filled nearby
in Equatorial regions;
• Tidal waves and ocean currents such as Gulf Stream, Florida Straits,
North Atlantic Drift possess energy which can be extracted with
underwater turbines.
Besides the power supply facilities, all these devices themselves could in fact
be areas of interest for deployment of our Safe-Net sensors.

12.2.2 Communications
Until now, there were several attempts to deploy underwater sensors that
record data during their mission, but they were always recovered afterwards.
This did not give the flexibility needed for real-time monitoring situations
like surveillance or environmental and seismic monitoring. The recorded
264 Common Framework Model for Multi-Purpose Underwater Data Collection

data could not be accessed until the instruments were recovered. It was also
impossible to detect failures before the retrieval and this could easily lead to
the complete failure of a mission. Also, the amount of data stored was limited
by the capacity of the devices on-board the sensors (flash memories, hard
disks).
Two possible implementations are buoys with high-speed RF-based com-
munications, or wired connections to some sensor nodes. The communication
bandwidth can be provided also by satellite connections which are usually
present on offshore facilities. If linked to an autonomous buoy, the device
provides GPS telemetry and has communication capabilities of its own.
Therefore, once the information gets to the surface, radio communications
are considered to be already provided as standard. Regarding underwater
communications, usually the typical physical layer technology implies acous-
tic communications. Radio waves have long-distance propagation issues
through sea water and can only be done at extra low frequencies, below
300 Hz [22]. This requires large antennae and high transmissions power,
which we would prefer avoiding. Optical waves do not suffer from such
high attenuation but are affected by scattering. Moreover, transmission of
optical signals requires high precision in pointing the narrow laser beams.
The primary advantage of this type of data transmission is the higher the-
oretical rate of transmission, while the disadvantages are the range and the
line-of-sight operation needed. We did not consider this as a feasible solution
due to marine snow, non-uniform illumination issues and other possible
interferences.
We do not intend to mix different communication protocols with different
physical layers, but we analyze the compatibility of each with existing
underwater acoustic communications, state-of-the-art protocols and routing
algorithms. Our approach will be a hybrid system, like the one in Figure 12.6
that will incorporate both tethered sensors and wireless acoustic where abso-
lutely no other solution can be implemented (e.g.: a group of bottom sea floor
anchored sensor nodes are implemented nearby an oil pipe, interconnected to
one or more underwater “sinks”, which are in charge of relaying data from
the ocean bottom network the a surface station [23].
Regarding the propagation of acoustic waves in the frequency gamma
we are interested in, for the multi-level communication between Safe-Net
sensor nodes, we are looking into already known models [24]. One of the
major problems related to the fluid dynamics are the non-linear movement
equations, which imply the fact that there isn’t a general exact solution.
Acoustics represent the first order of approximation in which the non-linear
 12.2 Research Challenges 265

Figure 12.6 Possible underwater sensor network deployment nearby Jack-up rig.

Figure 12.7 Sound pressure diagram:1– Equilibrium; 2– Sound; 3– Environment Pressure;

4– Instantaneous Pressure of Sound.

effects are neglected [25]. Acoustic waves propagate because of the medium
compressibility and the acoustic pressure or the sound pressure represents the
local deviation of the pressure whose root cause can be traced back to a sound
wave generated against the local environment. In air, the sound pressure can
be measured using a microphone, while in water it can be measured using a
hydrophone.
Considering the case of acoustic waves propagation in real fluids for our
mathematic general formalism, we have made the following assumptions:
gravity forces can be neglected, so equilibrium pressure and density get
uniform values all over the fluid’s volume (p0 and ρ0 ); the dissipative effects
such as viscosity and thermic conductibility are negligible; the medium is
homogenous, isotropic and has perfect elasticity, as well as the fluid particles
266 Common Framework Model for Multi-Purpose Underwater Data Collection

speed is slow (the “small amplitudes” assumption). Therefore, we can write a

Taylor development for the pressure and density fluctuation relationship:

∂p 1 ∂2p
p = p0 + |ρ = ρ0 (ρ − ρ0 ) + |ρ = ρ0 (ρ − ρ0 )2 + . . . , (12.1)
∂ρ 2 ∂ρ2
where the partial derivatives are constant for the adiabatic process around
the ρ0 equilibrium density of the fluid.
If the density fluctuations are small, meaning ρ̄  ρ0 , then the high-order
terms can be reduced and the adiabatic state equation becomes linear:
ρ − ρ0
p − p0 = K . (12.2)
ρ0
The pressure generated by the sound p (12.4) is directly related with the
particle movement and the amplitude ξ through equation (12.3):
v v p p
ξ= = = = , (12.3)
2πf ω Zω 2πf Z

aZ  Pac Z
p = ρc2πf ξ = ρcωξ = Zωξ = 2πfξZ = = Zv = c ρE = ,
ω A
(12.4)
where the symbols together with the I.S. measurement units are presented
in the following table:
The fundamental attenuation describes the power loss of a tone at a
frequency f, during its movement across a distance d. The first level of our

Table 12.1 Symbols Definition and Corresponding I.S. Measurement Units

Symbol Measurement Unit Description
p Pascal Sound Pressure
f Hertz Frequency
ρ kg/m3 Environment Density (constant)
c m/s Sound Speed (constant)
v m/s Particle Speed
ω rad / s Angular Speed
ξ m Particle Movement
Z N·s/m3 Acoustic Impedance
a m/s2 Particle Acceleration
I W/m2 Sound Intensity
E W·s/m3 Sound Energy Density
Pac Watt Acoustic Power
A m2 Surface
 12.2 Research Challenges 267

summary description takes into consideration this loss which occurs on the
transmission distance d. The second level calculates the specific loss of one
location caused by reflexions and refractions of upper and lower surfaces,
i.e. sea surface and bottom and also, the sound speed variations due to depth
differences. The result is a better prediction model of a specific transmitter.
The third level addresses the apparently random power shifts of the signal
received, by considering an average during a period of time. These changes
are due to slow variations of the propagation environment, e.g. tidal waves.
All these phenomena are relevant for determining the transmission power
needs in order to accomplish an efficient and successful underwater commu-
nication. We can also think at a separate model which could address much
faster changes of the instantaneous signal power at any given time, but at a far
smaller scale. The Signal Noise Ratio for different transmission distances as a
frequency function can be viewed in Figure 12.8. The sound absorption limits
the bandwidth which can be used for transmission and becomes dependent on
the distance:
By evaluating the entity A(d,f) N(f) as a function of ideal propagation of
the attenuation A(d,f) and as a consequence of tipical spectral power of the
background noise N(f), which drops 18dB per decade, we find the combined

Figure 12.8 Signal-Noise Ratio (SNR).

268 Common Framework Model for Multi-Purpose Underwater Data Collection

effect of attenuation and noise in underwater acoustics. This characteristic

describes the observation of SNR around the frequency bandwidth f. This
shows that high frequencies suffer fast attenuation on long distances, which
forces most modems to operate on a narrow bandwidth few kHz at most,
by suggesting the optimal frequency for a specific transmission [26]. It also
shows that the available bandwidth and implicitly the effective transmission
rate reduces with higher distances; therefore, the development of any big
network should start by determining its specific frequency and reserving a
bandwidth around it [27].

12.2.3 Maintenance
Ocean can be a very harsh environment and underwater sensors are prone to
failures because of fouling and corrosion. The sensor’s construction method
could include one miniaturized copper-alloy anode for anti-corrosion, as well
as one miniaturized aluminum-alloy anode which could fight fouling. Modern
anti-fouling systems already installed on rigs use microprocessor controlled
anodes and the current flowing to each anti-fouling and anti-corrosion anode
is quite low and the technology could be adapted by miniaturization of the
existing anodes. Although we are considering the environmental impact of
deploying such a high number of underwater devices, our primary concerns
are the feasibility and the durability of the network and how we can address
these factors in order to be able to expand our network through time and its
enlargement to be backwards compatible to already deployed nodes. Besides
the communication protocols being backwards compatible, underwater Safe-
Net nodes must possess self-configuration capabilities, i.e. must be able to
coordinate their operation, location or movement and data communication
handshake protocols by themselves. So, we state the obvious, that this can only
be possible if the Safe-Net nodes are resistant enough in the salt, corrosive
water of the sea.

12.2.4 Law and Finance

At the end of 2010, the European Commission issued a statement concerning
the safety regulations for offshore oil and gas drilling operations, with the
declared purpose of developing new laws for the European Union concerning
oil rigs. The primary objective of these measures will be the enforcement in
this domain of the highest safety standards in the world until present time in
order to prevent ecological disasters like the one in the Gulf of Mexico.
 12.2 Research Challenges 269

Moreover, during March – May 2011, following a public consultation

session regarding the European Union legal frame for current practices of
marine exploration and production safety standards, the community experts
have drawn the line saying that although generally speaking the activities meet
high standards of safety, these vary from one company to another, because of
the different laws which apply in each country. Therefore, the next legislative
proposal should enforce a common ground for all E.U. members concerning
the laws for prevention, reaction times and measures in case of emergencies,
as well as the financial liability.
According to the top 10 list of companies by largest revenues in the fiscal
year 2010–2011, 7 are oil and gas industry companies, which summed up
$2.54 billion USD revenues out of a total $3.43 billion USD. This means
more than 74% of the global revenues [28]; therefore, the cost of deploying
such Safe-Nets around drilling operations is rather small and the benefits
would be huge. Laws could be issued by governments in order to enforce
the obligation to oil and gas companies working at sea to use this sensor
networks every time a new site is being surveyed or a new jacket is installed.
This could also apply to existing oil rigs, jack-ups which move between
different places or even for subsea production sites. The ability to have small
devices physically distributed near offshore oil-fields’ operations brings new
opportunities for emergency-cases interoperability up to the higher level of
disaster management systems point of view [29].
Ocean Sampling Networks have been experimented in the Monterey
Bay area, where networks of sensors and AUVs, such as the Odyssey-
class AUVs, performed synoptic, cooperative adaptive sampling of the 3D
coastal ocean environment. Seaweb is an example of a large underwater
sensor network developed for military purposes of detection and monitoring
submarines [30]. Another example is the consortium formed by Massachusetts
Institute of Technology (MIT) and Australia’s Commonwealth Scientific and
Industrial Research Organization which has collected data with fixed and
mobile sensors mounted on autonomous underwater vehicles. The network
was only temporary and lasted only for a few days around the coasts of
Australia [31].
Ocean Observatories Initiative represents one of the largest ongoing under-
water cabled networks, which has eliminated both the acoustic communication
and power supply problems right from the start, by using already existing
underwater cables or new installs. The investment on Neptune project was
huge, approximately $153 billion dollars [32], but the idea seems quite bright
if we look at the most important underwater cables, which are already running
270 Common Framework Model for Multi-Purpose Underwater Data Collection

data under the oceans (Figure 12.9, with courtesy of TeleGeography.com). In

1956, North America was connected to Europe by an undersea cable called
TAT-1. It was the world’s first submarine telephone system, although telegraph
cables had crossed for the ocean for a century. Trans-Atlantic cable capacity
soared over the next 50 years, reaching a huge amount of data flowing back
and forth the continents, nearly 10 Tbps in 2008.

12.2.5 Possible Applications

• Seismic monitoring. Frequent seismic monitoring is of importance in oil
extraction; studies of variation in the reservoir over time are called 4-
D seismic and are useful for judging field performance and motivating
intervention;
• Disaster prevention and environmental monitoring. Sensor networks
for seismic activity mentioned before could also be used for tsunami
warnings to coastal areas. While there is always a potential for sudden
devastation (see Japan 2011), warning systems can be quite effective.
There is also the possibility of pollution monitoring: chemical, biological
and so on and so forth;
• Weather forecast improvement: monitoring of ocean currents and winds
can improve ocean weather forecasts, detecting climate change and also
understanding and predicting the effect of human activities on marine
ecosystems;
• Assisted navigation: sensors can be used to locate dangerous rocks in
shallow waters. The buoys can also signal the presence of submerged
wrecks or potential dangerous areas for navigation;
• Surveillance used for coast-line or border-lines, detecting the presence of
ships in country marine belt. Fixed underwater sensors can monitor areas
for surveillance, reconnaissance or even intrusion detection systems.

12.3 Mathematical Model

We introduce the class of systems which was considered when conducting
the research for the PhD thesis, as well as definitions on configurations of
sensors and remote actuators. This class of distributed parameter systems
which describes important concepts for parameter identification and optimal
experiment design has been adapted from the theoretical and practical research
“Optimal Sensing and Actuation Policies for Networked Mobile Agents in
a Class of Cyber-Physical Systems” [33]. The study presents models for
Figure 12.9 Most important underwater data & voice cables (2008).
12.3 Mathematical Model 271
272 Common Framework Model for Multi-Purpose Underwater Data Collection

aerial drones in the U.S.A., which could take high-resolution pictures of the
agricultural terrain and the algorithm was pointed to data correlation between
meteorological stations on the ground by matching the pictures with the
low-resolution ones taken from satellites. The purpose was to introduce a
new methodology to transform low-resolution remote sensing date about soil
moisture to higher resolution information that contains better knowledge for
use in hydrologic studies or water management decision making. The goal
of the study was aiming to obtain a high-resolution data set with the help
of a combination of ground measurements from instrumentation stations and
low-altitude remote sensing, typically images obtained from a UAV. The study
introduces optimal trajectories and launching points of UAV remote sensors in
order to solve the problem of maximum terrain coverage using least hardware
means, also expensive in their case.
We have taken further this study by matching the agricultural terrain
with our underwater environment and making an analogy between the fixed
instrumentation systems on ground, the meteorological stations and all the
fixed offshore structures already put in place through-out the sea. The mobile
drones are represented by remotely operated vehicles or by autonomous
underwater vehicles which can have data collection sensors on-board and can
be used as mobile network nodes. The optimisation of the best distribution
pattern of the nodes in the underwater environment can be extrapolated only
by neglecting the environment constants, which weren’t taken into account
by the study [33]. This issue is further to be investigated.

12.3.1 System Definition

The class of distributed parameter systems considered can be described by the
state Equation [34]:

ẏ (t) = Ay (t) + Bu (t)
, 0 < t < T, (12.5)
y (0) = y0
where Y = L2 (Ω) is the state space and Ω is a bounded and open subset of Rn
with a sufficiently regular boundary Γ = ∂Ω. The domain Ω is the geometrical
support of the considered system (12.5). A is a linear operator describing the
dynamics of the as the set of linear maps from U to Y is the input operator;
u ∈ L2 (0, T, U ) is the space of integrable functions f : ]0, T [ → U such
that the function t → f (t)p is integrable on ]0, T [ and U is a Hilbert
control space. In addition, the considered system has the following output
equation:
 12.3 Mathematical Model 273

z (t) = Cy (t) , (12.6)

 
Where C ∈ L L2 (Ω) , Z and Z is a Hilbert observation space.
We can adapt the definitions of actuators, sensors, controllability and
observability to system classes that are formulated in the state Equation
form (12.5).
The tradition approach of the analysis in distributed parameter systems is
fairly abstract in its purely mathematical form. Therefore, all the characteristics
of the system related to its spatial variables and geometrical aspects of the
inputs and outputs of the system are considered. To introduce a more practical
approach from an engineering point of view, the study [33] introduces the
concepts of actuators and sensors in the distributed parameter systems point
of view. With these concepts at hand, we can describe more practically the
relationship between a system and its environment, in our case sea/ocean
water. The study can be extended beyond the operators A, B and C, with the
consideration of the spatial distribution, location and number of sensors and
actuators.
The sensors’ measurements are, in fact, the observations on the system,
having a passive role. On the other hand, actuators provide a forcing input
on the system. Sensors and actuators can be of different natures: zone or
point-wise or domain distributed, internal or boundary, stationary or mobile.
An additional important notion is the concept of region of a domain. It is
generally defined as a subdomain of Ω. Instead of considering a problem
on the totality of Ω, the focus can be concentrated only on a subregion
ωε Ω, while the results can still be extended to ω = Ω. Such consider-
ation allows the generalization of different definition and methodologies
developed in previous works on distributed parameter systems analysis and
control.

12.3.2 Actuator Definition

Let Ω be an open and bounded subset of Rn with a sufficiently smooth
boundary Γ = ∂Ω [35]. An actuator is a couple (D, g) where D represents
the geometrical support of the actuator, D = supp (g) ⊂ Ω and g is its spatial
distribution.
An actuator (D, g) is said to be:
• A zone actuator if D is a non-empty sub-region of Ω;
274 Common Framework Model for Multi-Purpose Underwater Data Collection

• A point-wise actuator if D is reduced to a point b ∈ Ω. In this case, we

have g=∂ b where ∂ b is the Dirac function concentrated at b. The actuator
is denoted (b,∂ b ).
An actuator, zone or point-wise, is said to be a boundary actuator if its support
D⊂Γ. An illustration of the actuators supports is given in Figure 12.10:
In the previous definition, we assume that g ∈ L2 (D). For a collection of
p actuators (Di , gi )1≤i≤p , we have U = Rp , B : Rp → L2 (Ω) and:
p

u (t) → Bu (t) = gi ui (t) , (12.7)
i=1

u = (u1 , u2 , . . . , up )T ∈ L2 (0, T, Rp ) and gi ∈ L2 (Di ) with Di =

supp (gi ) ⊂ Ω for i = 1, . . . , p and Di ∩ Dj = ∅ for i = j. So, we have the
following:
B · y = (g1 , y, g2 , y, . . . , gp , y)T , z ∈ L2 (Ω) , (12.8)
where M T is the transpose matrix of M and <·,·>=<·,·>Y is the inner
product in Y and for v ∈ Y , if supp(v)=D, we have:

v, · = v, · L2 (D) . (12.9)
When D does not depend on time, the actuator (D,g) is said to be fixed or
stationary. Otherwise, it is a moving or mobile actuator denoted by (Dt ,gt ),
where D(t) and g(t) are, respectively, the geometrical support and the spatial
distribution of the actuator at time t, as in Figure 12.11:

Figure 12.10 Graphical representation of actuators’ supports.

 12.3 Mathematical Model 275

Figure 12.11 Illustration of the geometrical support and spatial distribution of an actuator.

12.3.3 Sensor Definition

A definition of sensors in the distributed parameters systems point of view is
provided by [35]: a sensor is a couple (D,h) where D is the support of the
sensor, D = supp (h) ⊂ Ω and h its spatial distribution.
A graphical representation of the sensors supports is given in Figure 12.12:
It is usually assumed that h ∈ L2 (D). Similarly, we can define zone or
point-wise, internal or boundary, fixed or moving sensors. If the output of the
system is given by means of q zone sensors (Di , hi )1≤i≤q with hi ∈ L2 (Di ),
Di = supp (hi ) ⊂ Ω for i = 1, . . . , q and Di ∩ Dj = φ if i =j, then in the

Figure 12.12 Graphical representation of the sensor supports.

276 Common Framework Model for Multi-Purpose Underwater Data Collection

zone case, the distributed parameter system’s output operator C is defined by

C : L2 (Ω) → Rp :

y → Cy = (h1 , y, h2 , y, . . . , hp , y)T . (12.10)

And the output is given by:
⎡ ⎤
h1 , yL2 (D1 )
⎢ .. ⎥
z (t) = ⎣ . ⎦. (12.11)
hq , yL2 (Dq )
A sensor (D,h) is a zone sensor if D is a non-empty sub-region of Ω. The
sensor (D,h) is a point-wise sensor if D is limited to a point c ∈ Ω and in this
case h=∂ c is the Dirac function concentrated in c. The sensor is denoted as
(c,∂ c ). If D ⊂ Γ = ∂Ω, the sensor (D,h) is called a boundary sensor. If D is not
dependent on time, the sensor (D,h) is said to be fixed or stationary, otherwise
it is said to be mobile and is denoted as (Dt ,ht ). In the case of q point-wise
fixed sensors located in (ci )1≤i=≤q , the output function is a q-vector given by
the relationship:
⎡ ⎤
y (t, c1 )
⎢ .. ⎥
z (t) = ⎣ . ⎦. (12.12)
y (t, cq )
Where ci is the position of the i-th sensor and y(t,ci ) is the state of the
system in ci at a given time t. [33] based on [36] devines also the notions of
observability and local controllability in the sense of distributed parameters
systems. [33] also shows that due to the nature of the problem of parameter
identification, the abstract operator-theoretic formalism used above to define
the dynamics of a distributed parameter system is not convenient. A formalism
based on n partial differential equations is used instead. According to this
setup, the sensor location and clustering phenomenon problem is ilustrated in
the Fisher Information Matrix (FIM) [37], which is a well-known performance
measure tool when looking for best measurements and is widely used in
optimum experimental design theory for lumped systems [38]. Its inverse
consititues an approximation of the covariance matrix for the estimate of Θ.
However, there is a serious issue in the FIM framework of optimal measure-
ments for parameter estimation of distributed parameters system, which is
the dependence of the solution on the initial guess on parameters [39]. The
dependence of the optimal location on Θ is very problematic; however, some
 12.4 ROV 277

robust design techniques have been developed in order to minimize or elude

the influence and we propose similar methodologies.
By analogy with the study [33], we can try to optimize this solution for
underwater points of interest. But in our case, of course, the problems are
much more complex because of the physical and chemical properties of the
environment.
We can consider two communication architectures for underwater
Safe-Nets. One is a two-dimensional architecture, where sensors are anchored
to the bottom of the ocean, and the other is a three-dimensional architecture,
where sensors float at different ocean depths covering the entire monitored
volume region. While the former is designed for networks whose main
objective is to monitor the ocean bottom, the latter is more suitable to detect and
observe phenomena in the three-dimensional space that cannot be adequately
observed by means of ocean bottom sensor nodes. The mathematical model
above refers only to the two-dimensional architecture case and we are looking
into further researches for the three-dimensional optimization, especially when
talking about the sensor-clustering phenomena.

12.4 ROV
A remotely operated vehicle (ROV) is a non-autonomous underwater robot.
They are commonly used in deep-water industries such as offshore hydro-
carbon extraction. A ROV may sometimes be called a remotely operated
underwater vehicle to distinguish it from remote control vehicles operating
on land or in the air. ROVs are unoccupied, highly manoeuvrable and
operated by a person aboard a vessel by means of commands sent through a
tether.
They are linked to the ship by this tether (sometimes referred to as an umbil-
ical cable), which is a group of cables that carry electrical power, video and
data signals back and forth between the operator and the vehicle. The ROVs
are used in offshore oilfield production sites, underwater pipelines inspection,
welding operations, subsea BOP (Blow-Out Preventer) manipulation as well
as other tasks:
• Seabed Mining – deposits of interest: gas hydrates, manganese nodules,
metals and diamonds;
• Aggregates Industry – used to monitor the action and effectiveness of
suction pipes during extraction;
• Cable and Node placements – 4D or time lapse Seismic investigation of
crowded offshore oilfields;
278 Common Framework Model for Multi-Purpose Underwater Data Collection

• Jack-up & Semi-Submersible surveys – preparation and arrival of a

Jack-up or Semi-Submersible drilling rig;
• Drilling & Wellhead support – monitoring drilling operations, instal-
lation/removal of template & Blow-Out Preventers (BOP), open-hole
drilling (Bubble Watch), regular inspections on BOP, debris removal,
IRM and in-field maintenance support (Well servicing);
• Decommissioning of Platforms / Subsea Structures – Dismantle struc-
tures safely and environment friendly;
• Geotechnical Investigation – Pipeline route surveys, Pipeline Lay –
Startup, Touchdown monitoring, Pipeline Pull-ins, Pipeline crossings,
Pipeline Lay-downs, Pipeline Metrology, Pipeline Connections, Post-lay
Survey, Regular Inspections;
• Submarine Cables – Route Surveys, Cable Lay, Touchdown monitoring,
Cable Post-Lay, Cable Burial;
• Ocean Research – Seabed sampling and surveys;
• Nuclear Industry – Inspections, Intervention and Decommissioning of
Nuclear Power Plants;
• Commercial Salvage – Insurance investigation and assessment surveys,
Salvage of Sunken Vessels, Cargoes, Equipment and Hazardous Cargo
Recovery;
• Vessel and Port Inspections – Investigations, Monitoring of Ports and
homeland security.
We are going to use PerrySlingsby Triton XLS and XLR models of the
remote operated vehicles (ROV), which are currently available in the Black
Sea area. While having the bigger goal in mind - deploying such networks
on a large scale - we can only think now for a small test bed and before
any physical implementation we are creating simulation scenarios on the
VMAX ROV Simulator. Simulation helps preventing any damages to the
ROV itself or any of the subsea structures we encounter. This also prevents any
real-life impossible design-situations to occur, e.g.: the ROV’s robotic arms
have very good dexterity and their movement is described by many degrees
of freedom - however, sometimes we find out the limits of motion and in
some given situations, deploying objects in some positions may prove difficult
or even impossible. We address these hypothetical situations and try to find
the best solutions for securely deploying the sensors by anchors to the sea
floor or by tethering to any metallic or concrete structures: jackets, jack-up
legs, autonomous buoys, subsea well production heads, offshore wind farm
production sites, so on and so forth.
 12.4 ROV 279

In the Black Sea area, operating in Romania’s territorial sea coast line,
we identified 4 working-class ROVs, out of which 2 are manufactured by
PerrySlingsby U.K.: 1 Triton XLX and 1 Triton XLR - first prototype of its
kind, which led to our models used in simulation.

12.4.1 ROV Manipulator Systems

Schilling Robotics’ TITAN 4 manipulator with 7 degrees of freedom
(Figure 12.13) is the industry’s premier system that offers the optimum
combination of dexterity and strength. Hundreds of TITAN manipulators are
in use worldwide every day, and are the predominant manipulator of choice for
work-class ROV systems. Constructed from titanium, the TITAN 4 is uniquely
capable of withstanding the industry’s harsh environment and repetitive
needs.
The movement of the 7-function Master Arm Control (Figure 12.14) is
transmitted through the fibre optics inside the tether and the underwater media
converters situated on the ROV pass the information to the Titan 4 Manipulator
after it is checked for send/receive errors. The exact movement of the joints
of the 7-function above the sea level represent the movement of the Titan-4

Figure 12.13 Titan 4 Manipulator 7-F.

280 Common Framework Model for Multi-Purpose Underwater Data Collection

Figure 12.14 Master Arm Control.

underwater. This provides the dexterity and degrees of freedom needed to

execute most difficult tasks (Figure 12.15):
Schilling’s RigMaster is a five-function, rate-controlled, heavy-lift grabber
arm that can be mounted on a wide range of ROVs (Figure 12.16). The
grabber arm can be used to grasp and lift heavy objects or to anchor the
ROV by clamping the gripper around a structural member at the work site.

Figure 12.15 Titan 4 – Stow dimensions.

 12.4 ROV 281

Figure 12.16 RigMaster 5-F.

Constructed primarily of aluminium and titanium, the RigMaster delivers the

power, performance, and reliability required for such demanding work. A
typical work-class ROV utilizes a combination of the five-function RigMaster
and seven-function TITAN 4.
With these two manipulator systems, any type of sensor can be deployed
or fixed on the ocean bottom. In order for a better understanding of the process
and likely problems which can occur during the installation, we are going to
use the VMAX Tech. – PerrySlingsby ROV Simulator for which we are going
to develop a modelling and simulation scenario concerning the deployment of

Figure 12.17 RigMaster range of motion – Side view.

282 Common Framework Model for Multi-Purpose Underwater Data Collection

underwater sensors Safe-Net surrounding areas of offshore oil and gas drilling
operations.

12.4.2 Types of Offshore Constructions

Offshore constructions represent the installation of structures and facilities
in a marine environment, usually for the production and transmission of
electricity, oil, gas or other resources. We have taken into consideration most
usual encountered offshore types of structures and facilities, focusing on the
shapes which are found underwater:
• Fixed platforms;
• Jack-up oil and gas drilling and/or production rigs;
• Jackets with top sides;
• Spars or floating platforms;
• Semi-submersibles;
• Drilling ships;
• Floating tension legs;
• Floating production storage and offloading (FPSO);
• Subsea well production heads;
• Offshore wind farm production sites.
We have created a simple scenario in which we use a PerrySlingsby Triton XLS
ROV connected to a TMS (Tether Management System) and where we can
use the physics of the robotic arms in order to understand which movements
are going to be needed in order to implant sensors of different sizes into the
ocean floor, as well as nearby the types of subsea structures mentioned above.
We try to create handling tools for the Schilling Robotics 7-F arm in order
to easily deploy and fix or common framework device model and also try to
find best spots for all the offshore types of structures we encountered in our
offshore experience inquiry [40].

12.5 ROV Simulator

The VMAX Simulator is software and hardware package intended to be
used by engineers to help in the design process of procedures, equipment
and methodologies, having a “physics based simulation” for the offshore
environment. The simulator is capable of creating scenarios that are highly
detailed and focused on one area of operation or broad in scope to allow an
inspection of an entire subsea field. For creating a scenario, there are two
 12.5 ROV Simulator 283

skill sets needed: 3D Studio Max modelling and “.lua” script programming
skills.
In order to safely deploy our Safe-Nets’ sensor balls into the water and fix
them to jack-up rigs metallic structures or to any other offshore constructions,
we first try to develop models of those structures and include them into a
standard fly-alone ROV simulation scenario. This is a two-step process as any
object’s model has to be created in 3D Studio Max software and afterwards
it can be programmatically be inserted into the simulation scenario. The
simulation scenarios are initialized by a series of Lua scripts, which is very
similar to C++ programming language and The VMAX Scenario Creation is
open source. The scripts are plain text files that can be edited using many
programs, including Microsoft Windows Notepad. The file names end with
.lua extension and are recommended to be opened with jEdit editor. This is
also an open-source editor which requires the installation of Java Runtime
Environment (JRE).
We have altered the simulation scenarios as it can be seen in Figure 12.18
and Figure 12.19 in order to obtain a better model of the Black Sea floor
through-out Romania’s coast line, which usually contains more sand because
of the Danube sediments coming from The Danube Delta. Geologists working
on-board the Romanian jack-ups considered the sea-floor in the VMAX ROV
Simulator very much alike with the one in the geological and oil-petroleum
interest zones up to 150-160 miles out in the sea. Throughout these zones
the water depth doesn’t exceed 80-90m, which is the limit at which drilling
jack-up rigs can operate (legs have 118m in length).
The simulator which is open-source was the starting base for a scenario
where we translated the needs of the ROV in terms of sensor handling, tether
positioning and pilot techniques combined with the specifications of the sea-
floor where the Safe-Nets will be deployed. The scenarios are initialized by
a series of .Lua scripts and the typical hierarchical file layout is presented in
Figure 12.20.
The resources are divided into two large classes of information: Scenarios-
related data and Assets. The former contains among others: Bathymetry,
Lua, Manipulators, Tooling, TMS (Tether Management System), Vehicles,
Components and IP (Internet Protocol communications between assets).
Bathymetry directory contains terrain information about a specific loca-
tion, where we could alter the sand properties on the sea floor. The terrain
stored here may be used across several scenarios. We could add a graphic
asset by using the template for the bathymetry part. The collision geometry
can be later generated based on the modelled geometry. We remind that the
284 Common Framework Model for Multi-Purpose Underwater Data Collection

Figure 12.18 Triton XLS ROV in simulation scenario.

Figure 12.19 Triton XLS schilling robotics 7-Function arm in scenario.

Courtesy of TelegeoGraphy.com
Figure 12.20 Typical hierarchical file layout.
12.5 ROV Simulator 285
286 Common Framework Model for Multi-Purpose Underwater Data Collection

simulator scenario creation software is open-source and we present in the

following lines some of the parts of the basic scenario provided with the
full-up simulator which we modified in order to accommodate our specific
needs:
graphicAsset {
assetID = "bathymetry",
castShadow = true,
– can be false for very flat terrain
– Our terrain model =
"assets/Bathymetry/TER 500m v2.0/TERBLKSEA 500m v1.0.ive",
receiveShadow = true,
scale = { 2, 2, 2 }
– specific to this particular model
}
–We changed the environment table to look like this:
environment = {
assembly = {
– Various items in the environment starting with bathymetry.
parts = {
– add the bathymetry based on a template
createFromTemplate(templates.bathymetry,
{
collisions = {
– The first item in the array is for the collision
– geometry automatically created from the model.
{
– set the area over which the bathymetry spans
size = { 100, 100, 1 },
– must be specified
}
– collision primitives may be appended to this array
},
– set the depth of the bathymetry
position = { 0, 0, REFERENCE DEPTH - 20 }}),}
constraints = { },
selfCollide = true,},
bathymetryPartName = "bathymetry",
pickFilter = { "bathymetry" },
 12.6 Common Modular Framework 287

currentDirectionTable = { 0 },
currentSpeedTable = { 1 },
depthTable = { 0 }}
The bathymetry template uses triangle mesh collision for the terrain. This will
provide collisions that are contoured to the bathymetry model
The Manipulators directory contains sub-directories for each arm and each
sub-director contains a model file with a .Lua script function used to create
the manipulator and add it to the ROV. We are looking forward to creating a
new manipulator usable for each case of sensor deployment.
The Tooling directory has either categories of tools or some uncategorized
ones, each having a model file name “.ive” or “.flt” and a Lua script file with
the code to create that specific tool [41].
Whereas the typical training scenarios include mainly a fly-around and
getting used to the ROV commands for the pilot and assistant pilot, we have
used the auxiliary commands in order to simulate the planting of the Safe-
Net around a jacket or a buoy for example. As far as the training scenario
is concerned, we covered the basics for a pilot to get around a jacket safely,
carrying some sort of object in the Titan4 manipulator robotic arm, without
dropping it, or without having the ROV’s tether tangling with the jacket
metallic structure. The tether contains hundreds of fibre-optic cables covered
with a Kevlar reinforcement, but it is recommended that no more than 4 total
3600 spins are made in one direction, clockwise or counter-clockwise, even
having this strengthened cover, in order to avoid any loss of communication
between the control console and the ROV itself. Any interaction between the
tether and the metallic structure could represent a potential threat to the ROV’s
integrity.

12.6 Common Modular Framework

An overview of the challenges and application possibilities of deploying under-
water sensor networks nearby oil rigs drilling operations areas and offshore
construction sites surroundings led to the conclusion that a standard device
is needed in order to deploy multi-purpose underwater sensor networks. We
detected the need for a standard, common, easy-to-use device framework for
multi-purpose underwater sensors in marine operations, as we were preparing
the devices for future use. This framework should be used for multiple different
sensors and we consider the modular approach to be best suited for future use,
providing the much-needed versatility.
288 Common Framework Model for Multi-Purpose Underwater Data Collection

We considered the buoyancy capabilities needed for a stand-alone device

launched on the sea surface and we started with an almost spherical-shaped
model Figure 12.21. If tethering should be needed, a small O-ring cap on one
of the sphere’s poles can be mounted:
The device will be able to accommodate a variety of sensors, adapted
within the inside “drawers”, its layers being highly modular. In this manner,
with the same network node, we will be able to empower a lot of types of
applications and this is an essential step in justifying the costs of development.
We believe this characteristic is critical for improving the financial desirability
of any future Safe-Nets offshore project.
Our simulation scenario is still scarce in modelled objects as the process of
creating them quite realistic is taking a long time. However many simulation
scenario variables we may alter, after finding out real types of situations
which occur on offshore structures, we learned that simulating the deployment
and deciding spots of anchoring for our sensors can only help, but not solve
real-life deployment, as parameters decided beforehand on shore can change
dramatically offshore.
However, we believe that our 3D models for underwater multi-purpose
sensors still stand as a good idea for our Safe-Nets real-life development and
implementation. Tethered or untethered, the upper hemisphere can include a
power adapter which can be used also as batteries compartment if the sensor is
wireless. The sensors have enough drawers for electronic modules and Type 03
is designed with built-in cable management system. Also, Type 03 is designed
with a membrane for a sensitive pollution sensor. We have chosen a very
simple closing mechanism for starters, using clamps on both sides, which can

Figure 12.21 Spherical-shaped model designed for common framework; a >= b; c is

Tether/Cable entry point diameter.
 12.6 Common Modular Framework 289

ensure the sealing of the device. The upper and lower hemispheres close on
top of an O-ring seal which can be lubricated additionally with water repellent
grease. Also, we have designed a unidirectional valve which can be used for
a vacuum pump to clear out the air inside. The vacuum strengthens the seal
against water pressure. In Figure 12.22, we present a few prototypes which
we tried to model and simulate:
Within the same common modular framework, we have thought at a multi-
deployment method for 3 or more sensors at the same time. Actually, the
following ideas were issued because of the repeated fail trials with an ROV
to grab and hold a Safe-Net sensor long enough in order to place it in a hook
coming from an autonomous buoy above the sea surface, affected by wave
length and height. Because of the real difficulties encountered, especially when
inserting higher waves into the scenario, we have thought of a way to get the
job done more quickly (Figure 12.23):

Figure 12.22 Underwater multi-purpose devices prototypes 01 – 05.

Figure 12.23 Grouping method for multiple simultaneous sensor deployment.

290 Common Framework Model for Multi-Purpose Underwater Data Collection

Moreover, the spherical model framework of the sensor, the basic node
of the Safe-Net, will prove to be very difficult to handle using the simple
manipulator, as it tends to slip, and the objective is to carry it without drop-
ping. Therefore, we have designed a “cup-holder” shape for grabbing more
easily the sphere and if it contains a cable connection, it should not be tampered
by the grabber, as it can be seen in Figure 12.24:

12.7 Conclusions
Most of the state-of-the-art solutions regarding underwater sensor networks
rely on specific task-oriented sensors, which are developed and launched
with different means and no standardization. The studies we found usually
use power from batteries and all sorts of resilient algorithms in order to
minimize battery draining and use sleep-awake states of the nodes, which
finally are recovered from water in order to retrieve data collections. Our
approach is trying to regulate the ways of deploying and fixing the sensors
towards offshore structures and moreover to offer solutions to more than one
application task. This may seem as a general approach, but this is needed
in order to avoid launching different technology nodes which afterwards
will not be able to communicate with each other. Development of a virtual
environment-based training system for ROV pilots could be the starting point
for deploying underwater sensor networks worldwide, as these are the people
who will actually be in the position to implement it.
This chapter investigates the main challenges for the development of an
efficient common framework for multi-purpose underwater data collection

Figure 12.24 Basic sensor device holder designed for simulation.

 References 291

devices (sensors). Several fundamental key aspects of underwater acoustic

communications are also investigated. We want to deploy around existing
offshore constructions and this research is still a work in progress. The main
challenges for the development of efficient networking solutions posed by
deploying sensors in the underwater environment are detailed at all levels. In
short, this article has analyzed the necessity of considering the physical fun-
damentals of an underwater network development in the planetarium ocean,
starting from the instrumentation needs surrounding offshore oil drilling sites
and early warning systems for disaster prevention worldwide.
We suggest various extension possibilities for applications of these safe-
nets, starting from pollution monitoring around offshore oil drilling sites, early
warning systems for disaster prevention (earthquakes, tsunami) or weather
forecast improvement, up to military surveillance applications, all in order to
overcome the cost of implementation of such underwater networks.

References
[1] K. Eshghi and R. C. Larson, ‘Disasters: lessons from the past 105 years’,
Disaster Prevention and Management, Vol. 17, pp.61–82, 2008.
[2] D. Green, ‘Acoustic modems, navigation aids and networks for undersea
operations’, IEEE Oceans Conference Proceedings, pp.1–6, Sydney,
Australia, May 2010.
[3] J. Heidemann, Y. Li and A. Syed, ‘Underwater Sensor Networking:
Research Challenges and Potential Applications’, USC Information
Sciences Institute, USC/ISI Technical Report ISI-TR-2005–603, 2005.
[4] T. Melodia, Ian F. Akyildiz and D. Pompili, ‘Challenges for Effi-
cient Communication in Underwater Acoustic Sensor Networks’, ACM
Sigbed Review, vol.1, no.2, 2004.
[5] A. Cerpa and et al., ‘Habitat monitoring: Application driver for wireless
communications technology’, ACM SIGCOMM Workshop on Data
Communications in Latin America and the Caribbean, San Jose, Costa
Rica, 2001.
[6] D. Whang, N. Xu and S. Rangwala, ‘Development of an embedded
sensing system for structural health monitoring’, International Workshop
on Smart Materials and Structures Technology, pp. 68–71, 2004.
[7] D. Steere, A. Baptista and D. McNamee, ‘Research challenges in envi-
ronmental observation and forecasting systems’, 6th ACM International
Conference on Mobile Computing and Networking, Boston, MA, USA,
2000.
292 Common Framework Model for Multi-Purpose Underwater Data Collection

[8] L. Dignan, ‘Google’s Data Centers’, [Online] 2011.http://www.zdnet.com

/blog/btl/google-makes-waves-and-may-have-solved-the-data-center-
conundrum/9937http://www.datacenterknowledge.com/archives/2008/
09/06/google-planning-offshore-data-barges/.
[9] A. S. Outlaw, ‘Computerization of an Autonomous Mobile Buoy’,
Florida Institute of Technology, Vol. Master Thesis in Ocean Engineer-
ing, Melbourne, FL, 2007.
[10] C. Garcier, et al., ‘Autonomous Meteorogical Buoy’, Instrumentation
Viewpoint, vol. 7, Winter, 2009.
[11] V. Dache, M.C. Caraivan and V. Sgarciu, ‘Advanced Building Energy
Management Systems - Optimize power consumption’, INCOM 2012,
pp. 426, Bucharest, 2012.
[12] C. Hong-Jun, et al., ‘Challenges: Building Scalable and Distributed
Underwater Wireless Sensor Networks (UWSNs) for Aquatic Appli-
cations’, UCONN CSE Technical Report, UbiNet-TR05–02, 2005.
[13] D. Pompili, T. Melodia and A.F. Ian, ‘A Resilient Routing Algorithm
for Long-term Applications in Underwater Sensor Networks’, Atlanta,
2005.
[14] Aquaret, [Online], 2008. www.aquaret.com
[15] PelamisWaves, [Online], 2012. http://www.pelamiswave.com/pelamis-
technology
[16] WaveBob, [Online], 2009. http://www.wavebob.com
[17] Buoy, OE, [Online], 2009. www.oceanenergy.ie
[18] PowerBuoy, [Online], 2008. www.oceanpowertechnologies.com
[19] Oyster, [Online], 2011. www.aquamarinepower.com
[20] WaveDragon, [Online], 2011. www.wavedragon.net
[21] AWS, [Online], 2010. http://www.awsocean.com
[22] F. Mosca, G. Matte and T. Shimura, ‘Low-frequency source for very
long-range underwater communication’, Journal of Acoustical Society
of America Express Letters, vol. 133, 10.1121/1.4773199, Melville, NY,
U.S.A., 20 December 2012.
[23] D. Pompili and T. Melodia, ‘An Architecture for Ocean Bottom
UnderWater Acoustic Sensor Networks (UWASN)’, Georgia, Atlanta,
2006.
[24] R. Urick, ‘Principles of underwater sound’, McGraw Hill Publishing,
New York, NY, U.S.A., 1983.
[25] S.W. Rienstra and A. Hirschber, ‘An Introduction to Acoustics’,
Eindhoven University of Technology, Eindhoven, The Netherlands,
2013.
 References 293

[26] J. Wills, W. Ye and J. Heidemann, ‘LowPower Acoustic Modem

for Dense Underwater Sensor Networks’, USC Information Sciences
Institute, 2008
[27] M. Stojanovic, ‘On the relationship between capacity and distance in an
underwater acoustic communication channel’, ACM Mobile Computing
Communications, Rev.11, pp.34–43, doi:10.1145/1347364.1347373,
2007.
[28] Wikipedia.org, Wikipedia List of Companies by Revenue, [Online],
2011. http://en.wikipedia.org/wiki/List of companies by revenue
[29] V. Nicolescu, M. Caraivan, ‘On the Interoperability in Marine Pollution’,
IESA’14 7th International Conference on Interoperability for Enterprise
Systems and Applications, Albi, France, 2014.
[30] J. Proakis, J. Rice, et al., ‘Shallow water acoustic networks’, IEEE
Communications Magazine, pp. 114–119, 2001.
[31] I. Vasilescu, et al., ‘Data collection, storage and retrieval with an
underwater sensor network’, 3rd ACM SenSys Conference Proceedings,
pp.154–165, San Diego, CA, U.S.A., November 2005.
[32] P. Fairley, ‘Neptune rising’, IEEE Spectrum Magazine #42, pp. 38–45,
doi:10.1109/MSPEC.2005.1526903, 2005.
[33] C. Tricaud, ‘Optimal Sensing and Actuation Policies for Networked
Mobile Agents in a Class of Cyber-Physical Systems’, Utah State
University, Logan, Utah, 2010.
[34] A. El Jai, ‘Distributed systems analysis via sensors and actuators’,
Sensors and Actuators A’, vol. 29, pp.1–11, 1991.
[35] A. El Jai and A.J. Pritchard, ‘Sensors and actuators in distributed
systems’, International Journal of Control, vol. 46, iss. 4, pp. 1139–1153,
1987.
[36] A. El Jai, et al., ‘Regional controllability of distributed-parameter
systems’, International Journal of Control, vol. 62, iss. 6, pp.1351–1356,
1995.
[37] E. Rafajowics, ‘Optimum choice of moving sensor trajectories for
distributed parameter system identification’, International Journal of
Control, vol. 43. pp.1441–1451, 1986.
[38] N.Z. Sun, ‘Inverse Problems in Groundwater Modeling’, Theory and
Applications of Transport in Porous Media, Kluwer Academic Publish-
ers, Dodrecht, The Netherlands, 1994.
[39] M. Patan, ‘Optimal Observation Strategies for Parameter Estimation of
Distributed Systems’, University of Zielona Gora Press, Zielona Gora,
Poland, 2004.
294 Common Framework Model for Multi-Purpose Underwater Data Collection

[40] M. Caraivan, V. Dache and V. Sgarciu, ‘Simulation Scenarios for

Deploying Underwater Safe-Net Sensor Networks Using Remote Oper-
ated Vehicles’, 19th International Conference on Control Systems and
Computer Science Conference Proceedings, IEEE CSCS’19 BMS#
CFP1372U-CDR, ISBN: 978–0-7695–4980-4, Bucharest, Romania,
2013.
[41] B. Manavi, VMAX Technologies Inc. Help File, Houston, 77041–4014
Texas, TX, U.S.A., 2010.
 13
M2M in Agriculture – Business Models
and Security Issues

S. Gansemer1 , J. Sell1 , U. Grossmann1 , E. Eren1 , B. Horster2 ,

T. Horster-Möller2 and C. Rusch3
1 University of Applied Sciences and Arts Dortmund, Dortmund, Germany
2 VIVAI Software AG, Dortmund, Germany
3 Claas Selbstfahrende Erntemaschinen GmbH, Harsewinkel, Germany

Corresponding author: S. Gansemer <[email protected]>

Abstract
Machine-to-machine communication (M2M) is one of the major innovations
in the ICT sector. Especially in agricultural business with heterogeneous
machinery, diverse process partners and high machine operating costs, M2M
offers large potential in process optimization. Within this paper, a concept for
process optimization in agricultural business using M2M technologies is pre-
sented using three application scenarios. Within that concept, standardization
and communication as well as security aspects are discussed. Furthermore, cor-
responding business models building on the presented scenarios are discussed
and results from economic analysis are presented.

Keywords: M2M, agriculture, communication, standardization, business

case, process transparency, operation data acquisition, business model,
security.

13.1 Introduction
Machine-to-machine communication (M2M) currently is one of the major
innovations in the ICT sector. The agricultural sector is characterized by
heterogeneous machinery, diverse process partners and high operational

Advances in Intelligent Robotics and Collaborative Automation, 295–312.


c 2015 River Publishers. All rights reserved.
296 M2M in Agriculture – Business Models and Security Issues

machinery costs. Many optimization solutions aim to optimize a single

machine but not the whole process. This paper deals with improving the entire
process chain within the agricultural area. In the first part of this paper, a
concept for supporting process optimization in heterogeneous process chains
in agricultural business using M2M communication technologies is discussed.
The second part presents business cases for the proposed system and outcomes
from economic analysis. In the third part last not least security aspects related
to the proposed system are discussed.

13.2 Related Work

The application of M2M technology in agriculture is targeted by several
other research groups. Moummadi et. al. [1] present a model for an agricul-
tural decision support system using both multi-agent-system and constraint
programming. The systems purpose is controlling and optimizing water
exploitation in greenhouses.
Wu et. al. [2] present a number of models for M2M usage in different
sectors such as utilities, security and public safety, tracking and tracing,
telematics, payment, healthcare, remote maintenance and control and con-
sumer devices. They discuss technological market trends and the influence of
different industries on M2M applications.
An insurance system based on telematics technology is demonstrated by
Daesub et. al. [3]. They investigate trends in insurance industry based on
telematics and recommend a supporting framework.
A business model framework for M2M business models based on cloud
computing is shown by Juliandri et. al. [4]. They identify nine basic building
blocks for a business model aiming to increase value while reducing costs.
Gonçalves and Dobbelaere [5] discuss several business scenarios based on
specific technical scenarios. Within the presented scenarios, the stakeholders
assume different levels of control over the customer relationship and the assets
determining the value proposition.
A model for software updates of mobile M2M devices is presented in [6].
They aim on low bandwidth use and avoidance of system reboot.

13.3 Communication and Standardization

The agricultural sector is characterized by heterogeneous machinery and
diverse process partners. Problems arise from idle times in agricultural pro-
cesses, suboptimal machine allocation and improper planning. Other problems
 13.3 Communication and Standardization 297

are generated by incompatibilities of machinery built by different manufactur-

ers. Because of proprietary signals on machine buses not fitting on one another
collaboration between machines may be inhibited [7, 8].
To support collaboration of heterogeneous machinery a standardized
communication language is needed. Communication takes place either direct
via machine to machine or via machine to cloud.
Sensors in machines record different parameters such as position, moving
speed, mass and quality of harvested produce. These operational and machine
logging data from the registered machines are synchronized between machines
and finally sent via telecommunication network to a recording web portal. Data
are stored within the portal´s database and are used for optimizing process
chain or develop and implement business models based on that data. All data
is sent through machine’s ISO- and CAN-bus in proprietary syntax.
Within the concept, each machine uses a “black-box” which trans-
lates manufacturer specific bus signal data to a standardized data format.
The concept is shown in Figure 13.1. Machines may be equipped with
diverse numbers of sensors resulting in different numbers of signals available.
The standard should cover most of those signals. However, due to the
diverse machinery available, not every signal available on the machine can be
supported within the proposed concept.
Within this paper, the concept of a portal (M2M-Teledesk) is presented
suited for dealing with the problems mentioned above. The system’s
framework is shown in Figure 13.2. The black-boxes installed on each
machine are interfaces between the machine’s internal buses and the portal

Figure 13.1 Synchronization of standards.

298 M2M in Agriculture – Business Models and Security Issues

Figure 13.2 M2M teledesk framework [9].

(M2M-Teledesk). Black-boxes are equipped with mobile network communi-

cation interfaces for transferring data between machines among each other
and between machines and the portal as well.
Every machine is set up with a black-box which reads internal buses
and translates signals to the proposed open M2M standard, runs different
applications and communicates data to and from the machine using WIFI
or mobile data communication networks. The system uses a public key
infrastructure for safety and trust reasons (see Section 13.7). Within the portal
collected data is aggregated and provided to other analyzing and evaluating
systems (e.g. farm management). Depending on the machine a full set or a
subset of data specified in the standard can be used. Older machines may be
retrofitted with a black-box providing only a subset of available data as a
smaller number of sensors are available only.
 13.3 Communication and Standardization 299

The data is visualized within the portal and helps the farmer to optimize
business processes to meet documentation requirements or to build data-based
business models. Especially when it comes to complex and detailed records
of many synchronized machines, the system shows its advantages.
Communication between machines takes place either directly from
machine to machine or via a mobile communication network (e.g. GSM or
UMTS). Within agricultural processes operating in rural areas, the availability
of mobile communication networks is not always given. There are two
strategies to increase the availability of network coverage:
• National roaming SIM cards;
• Femtocells.
With national roaming SIM cards being able to roam into all available
networks, the availability of mobile network coverage can be increased, while
with standard SIM cards only one network can be used in the home country
[10]. National roaming SIM cards are operating in a country different from
their home location (e.g. a spanish SIM card operating in Germany). The SIM
card can roam into all available networks as long as issuing provider and
network operator signed a roaming agreement. Although network coverage
can be increased, a communication channel cannot be guaranteed.
With femtocells [2], dedicated base station is placed on the field where
machines are operating. The concept is presented in Figure 13.3. Machines

Figure 13.3 Femtocell communication in agriculture [9].

300 M2M in Agriculture – Business Models and Security Issues

communicate to the base station e.g. via WLAN or GSM/UMTS, while base-
station is connected to the portal by GSM/UMTS or satellite connection.
The location of the femtocell base-station should be chosen in a way that
coverage is given at every location within the corresponding area either via the
femtocell or via direct connection to a mobile network. This strategy enables
communication even without network coverage by the operator. However, the
implementation effort is significantly higher than in case of using national
roaming SIM cards.

13.4 Business Cases

The described system can be used in different manners. Three main business
cases have been identified:
• Process Transparency (PT);
• Operation Data Acquisition (ODA);
• Remote Software Update (RSU).
Process transparency (PT) mainly focuses on in-time optimization of process
chains, while ODA uses downstream analysis of data. Remote software
update (RSU) aims to securely install applications or firmware updates on
machines without the use of a service technician. These three business cases
are described below in more detail.

13.4.1 Process Transparency (PT)

Processes in agricultural business are affected by several process participants.
Furthermore, the used machines in many cases are operating with high costs.
A visualization of an exemplary corn-harvesting process is presented in
Figure 13.4. During the harvesting process, a harvester is e.g. cropping corn.
Synchronously, a transport vehicle needs to drive in parallel to the harvester to
transport the harvested produce. Machines involved in this sub-process need to
be synchronized in real time. In case of the transport vehicle being filled up, it
has to be replaced by another empty transport vehicle. Full transport vehicles
make their way to e.g. a silo or a biogas power plant where the transport
vehicle has to enter via a scale to measure the mass of the harvested produce.
Furthermore, a quality check of the harvested produce is carried out manually.
This process may be optimized by the portal in different ways. Due to the
registration of sensor data, the weighting and quality check part in the process
may be skipped or reduced to spot checks if the customer deems the data
within the system to be trustworthy. Furthermore, the data is visualized by the
Figure 13.4 Information- and controlflow of scenario harvesting process.
13.4 Business Cases 301
302 M2M in Agriculture – Business Models and Security Issues

portal to give the production manager the opportunity to optimize the process
in near real-time. Before starting the process, a production plan is prepared
by the production manager either manually with support by the system or
automatically by the system. Within the plan, machines are allocated with
time and position data. When the system registers a plan deviation, the plan is
updated either manually or automatically. This approach allows reducing idle
times saving costs and resources.

13.4.2 Operations Data Acquisition (ODA)

Within Operations Data Acquisition (ODA) scenario data gathered by
the machine sensors is saved for downstream processing and analysis.
While process transparency aims to synchronize process data in real-time
to support process optimization, ODA data is gathered and sent to the server
after the process is finished. Analysis is done e.g. to generate yield maps or to
analyze machine behavior.

13.4.3 Remote Software Update (RSU)

The remote software update (RSU) process aims to remotely install software
on a machine. Software update includes two sub scenarios, firmware upgrade
and app-installation. App-installation means the installation of an additional
piece of software from a third-party-software provider while firmware updates.
The main aspect of software update is to ensure that the software is installed in
a secure way, meaning that the machine proof to install software which comes
from an authorized source and was not changed during network transport.
Details on the security measures can be found in Section 13.7.

13.5 Business Models

Based on the scenarios and the data described above business and licensing
models are developed. Figure 13.5 shows the value chain of M2M Teledesk
consisting of six partners.
For all partners of the value chain business potential has been analyzed
and is shown in Table 13.1. The table shows the partner’s roles, the expected
revenue and cost development and the resulting business potential.

Figure 13.5 Value chain of M2M-Teledesk [9].

 13.5 Business Models 303

Table 13.1 Revenue, costs and business potential for partners along M2M value chain
Revenue Cost
Development Development Business
Partner Role (Per Unit) (Per Unit) Potential
Module Manufacturer Constant Declining +
manufacturer of black-box
Machine Manufacturer Progressive Declining ++
manufacturer of machines
Mobile Data transport, Constant Declining +
network SIM
operator management
3rd party Software Constant/ Depending on +
software developer, progressive business model
provider application (depending on
provider business
model)
Portal Portal operator Progressive Declining ++
provider

The module manufacturer produces the black-boxes (see Figure 13.2) built
into the machines or used to retrofit older machines. Revenues for module
manufacturers mostly come from black-box sales. Costs per unit are expected
to decline with increasing number of sold units.
The machine manufacturer’s revenues come from machine sales as well
as services delivery and savings due to remote software updates. The cost of
development is expected to be declining with the increasing number of sold
units.
The mobile network operator’s role is to deliver data through a mobile
network. SIM card management may also be done by the network operator
but can also be done by an independent partner. Revenues consist of fees
for data traffic as well as service fees for SIM card supply and management.
Additional costs for extra data volume over an existing network are very low.
Third-party software providers can be part of the value chain; however,
this is not compulsory. They either supply an application bringing additional
functions to the machinery or implement an own business model based on the
data held in the portal.
The software is sold through the portal and is delivered to the machinery by
the remote software update process described above. The revenues develop-
ment per unit depends on the employed business model. When only software is
sold, revenues per unit are constant. With additional business models, revenues
may also develop progressively.
304 M2M in Agriculture – Business Models and Security Issues

Costs are mostly one-time costs for software development as well as

running costs for maintenance. However, additional costs may arise depending
on the business model. The portal provider operates and manages the portal.
Revenues consist of usage fees, revenues from third party app sales, fees for
delivering software updates and other service fees. Costs are mainly for portal
operation, support and data license fees. The end users’ revenues come from
savings due to increased process efficiency, while costs arise for additional
deductions for machines, additional costs for higher skilled workforce, system
usage fees and so on. Business potential is given for all partners involved in
the value chain.
With applications developed by third-party software developers a variety
of new business models can be implemented. One model is given by “pay-per-
use” as well as “pay-how-you-use” insurance or leasing. Within this business
model insurance or leasing companies are able to calculate insurance or leasing
rates more adequate to risk depending on real-usage patterns. The insurance
or leasing company is integrated in the value chain as a third-party software
provider. For running the business model, data showing the usage pattern
is needed. To gain this data, the third-party software provider needs to pay
license fees.

13.6 Economic Analysis

Economic analysis of the system leads to a model consisting of linear
equations. For visualizing the quantitative relations between different services,
sub-services and partners of a so-called swimlane-gozintograph is used. Based
on standard gozintograph methodology as described in [11] the resulting figure
is adapted by including swimlane methodology [12] to show the involved part-
ners. Figure 13.6 shows the corresponding swimlane-gozintograph. Columns
represent the involved partners; transparent circles indicate different services
delivered by the partners. Shaded circles represent business cases, i.e. services
delivered externally.
The figure shows the relations between internal and external services
and the share of each partner in the different business cases. From this
gozintograph, mathematical equations can be derived, enabling the calculation
of the gross margins for each business case.
From Figure 13.6, linear equations are derived, including transfer prices,
amounts of service delivery and external sales prices for cost and gross margin
calculation.
 13.6 Economic Analysis 305

Figure 13.6 Service-links between delivering and receiving services.

Variable costs for the business cases can be calculated using Equation
system (13.1).
c1 = a11 · b1 + a12 · b2 + . . . + a1n · bn
.. (13.1)
.
cm = am1 · b1 + am2 · b2 + amn · bn ,

where aij – amount of service j delivered for service i; bj – transfer prices

of service j; ci – variable costs of finally receiving service i; m – number of
finally receiving services; n – number of delivering services.
306 M2M in Agriculture – Business Models and Security Issues

The system of linear equations yields relation matrix A=(aij ) and transfer
price vector B=(bj ). The vector C=(ci ) of variable costs can be represented
by Equation (13.2).
C = A · B. (13.2)
Using the vector D=(di ) consisting of sales prices of finally receiving services
Equation (13.3) leads to the vector M=(mi ) of gross margin per unit of all
business cases, i.e. finally receiving services.

M = D − A · B. (13.3)

Figure 13.7 exemplifies the input matrix A and vectors B and D with estimated
quantities. In matrix A, the rows indicate the business cases PT (row 1), ODA
(row 2) and RSU (row 3). Columns represent delivering services indicated
as white circles in Figure 13.6. The elements of vector B represent transfer
prices of delivering services. Elements of the vector D represent sales prices
of the three business cases.
The results of economic analysis are shown in Figure 13.8. Elements of
the calculated vector C indicate variable costs of the three business cases. It

Figure 13.7 Relation matrix A, transfer price vector B and sales price vector D.

Figure 13.8 Vector of variable costs C and vector of marginal return per unit M.
 13.7 Communication Security 307

can be seen that the marginal return per unit is positive for all three business
cases with the highest marginal return for business case Process Transparency.

13.7 Communication Security

Securing the communication channel against unauthorized access and manip-
ulation is another important factor which has to be taken into account.
One has to consider the following communication scenarios: communica-
tion between machines and portal using mobile networks such as 3G/4G and
WLAN, secure remote firmware update and covering dead spots.
The whole security concept is based on asymmetric encryption. Every
participant in the communication chain (machines, machine manufacturer,
M2M portal, provider) needs a key pair which should be created on the
machine to keep the private key on the device.
This security concept was developed in the context of a bachelor thesis at
the FH Dortmund [12]. The main target was the use of open and established
standards [13, 14].

13.7.1 CA
The central instance of the security structure is a CA (Certificate Authority)
which provides services like issuing certificates (by providing a CSR (Cer-
tificate Signing Request)), revoking certificates, checking certificates if they
are rejected (through CRLs/OCSP). During the key creation process, a CSR
is being created which will be passed to the PKI. The CSR is signed and the
certificate is sent back to the device (machine).

13.7.2 Communicating On-the-Go

The communication between the machines and the portal is secured by means
of a mutually authenticated HTTPS connection. The portal identifies itself to
the machine by presenting its certificate and vice versa. During the initiation
of the connection, every device has to check the presented certificate by the
other part: 1) is the certificate signed by the M2M CA (this prevents man-in-
the-middle-attacks)? If yes: 2) check the certificate of the counterpart against
the CA if the certificate is revoked or not. This is done by using OCSP or CRLs
(as a fallback in case OCSP is failing).
After the connection has been initiated, both partners can communicate
securely, while the security of the underlying network(s) (like mobile 2G/3G,
WLAN etc.) is no more important.
308 M2M in Agriculture – Business Models and Security Issues

13.7.3 Covering Dead Spots

In case that a mobile communication is not possible due to lacking avail-
ability, the collected data has to be transferred using other methods. Here,
other vehicles (such as transportation vehicles) have to deliver the data
from the machine within the dead spot to areas with mobile network
coverage from where they are sent to the portal. During the transporta-
tion, the data has to be secured against manipulation and unauthorized
access.
Preparing a data packet for delivery involves the following steps: At first
the data is encrypted using the portal’s public key. In order to check if the public
key is still valid, it is checked with the corresponding certificate against the CA
(through OCSP/CRL). This prevents unauthorized access. In the next step, the
signature of the encrypted data is created. Therefore, the checksum of the data
is calculated and encrypted with the private key of the originating machine.
Both the signature (encrypted checksum) and the encrypted data are sent to
the vehicle.
The portal checks the signature by decrypting the checksum using the orig-
inating machine’s public key (key/certificate is checked through OCSP/CRL)
and by creating the checksum itself of the data package. If both checksums
match, the data has not been manipulated and can be decrypted using the
private key of the portal.

13.7.4 Securing WLAN Infrastructures

In the vicinity of a farm, a wireless LAN connection will be used instead of
mobile network connection. The M2M project elaborated a reference WLAN
network which can be installed on the farm premises. This network is designed
and optimized for the M2M system. In order to guarantee that only authorized
machines have access to the network, the authentication scheme is based on
IEEE 802.1X with a RADIUS using AES/CCMP encryption (IEEE 802.11i
RSN). Furthermore, a DHCP/DNS service is provided by the gateway which
interconnects the network to the internet and acts as relay to the M2M portal.
A machine connects to the M2M wireless network by using its X.509v2
certificate. The certificate is presented to the RADIUS server which performs
checks (OCSP/CRL) against the CA whether it is revoked or not and whether
the certificate is signed by the M2M CA. The machine itself has to check the
RADIUS certificate whether it belongs to the M2M CA in order to avoid rogue
access points. If all checks are passed successfully, the RADIUS server grants
access to the network.
 13.8 Resume 309

13.7.5 Firmware Update

It is necessary to periodically apply updates for the software systems on the
machines. The updates are passed from the manufacturer of the machine
through the portal to the destination machine.
Since the update packages may contain critical and confidential data, the
provision of the update package has to be secured accordingly. Because of the
file size (100MB and up), asymmetric encryption is not appropriate. Instead, a
256-bit symmetric AES key is generated and is used to encrypt the update. This
key is secured using the public key encryption (after checking the correspond-
ing certificate through OCSP/CRL). Here, the public key of the destination
machine is used. In the next step, the signature of the encrypted update file
is calculated by generating the hash value which then is encrypted with the
private key of the manufacturer. Now the signature, the encrypted file and
the encrypted AES key are sent to the destination machine.
On the latter, the signature is checked by generating the checksum of the
encrypted file and by comparing it with the decrypted checksum. The check-
sum is decrypted with the public key of the manufacturer. The corresponding
certificate has to be checked, too) checksum. If both checksums match, the
update did not lose integrity. Finally, the AES key can be decrypt using
the private key of the destination machine and the update can be decrypted
(Figure 13.9).

13.8 Resume
This paper presents a concept for the optimization of process information
chain to improve efficiency in agricultural harvesting process. Machine-to-
machine communication plays a central role to synchronize data between
diverse process partners.
The information gathered by sensors at agricultural machines plays
the central role to build new business models. Business model analysis
shows that all parties along the value chain gain good business potential.
It has been shown that the three described business models can be operated
with positive marginal return per unit under the assumptions made in the
project.
However, security issues and business models play an important role
for a successful system operation. With the described security measures,
system operation can be done ensuring confidentiality, integrity as well as
availability.
310 M2M in Agriculture – Business Models and Security Issues

Figure 13.9 Large file encryption.

 References 311

As the system is designed with an open and generic approach, adoption

to other branches such as construction and the opportunity to bring in new
functions and business models via third-party software brings additional
market potential.
The concepts presented in this paper were developed within the project
M2M-Teledesk. The project aims to implement a prototypical system follow-
ing the concept described above.

13.9 Acknowledgement
The presented work was done in the research project “M2M-Teledesk” funded
by the state government of North-Rhine-Westfalia and European Union Fund
for regional development (EUROPÄISCHE UNION - Europäischer Fonds
für regionale Entwicklung - Investition in unsere Zukunft). Project partners
of M2M-Teledesk are University of Applied Sciences and Arts, Dortmund,
VIVAI Software AG, Dortmund and Claas Selbstfahrende Erntemaschinen
GmbH, Harsewinkel.

References
[1] K. Moummadi, R. Abidar and H. Medromi, ‘Generic model based on
constraint programming and multi-agent system for M2M services and
agricultural decision support’, In Multimedia Computing and Systems
(ICMCS), 2011:1–6.
[2] G. Wu, S. Talwar, K. Johnsson, N. Himayat and K. Johnson, ‘M2M: From
Mobile to Embedded Internet’, In IEEE Communications Magazine,
2011:36–42.
[3] Y. Daesub, C. Jongwoo, K. Hyunsuk and K. Juwan, ‘Future Automotive
Insurance System based on Telematics Technology’, In 10th Interna-
tional Conference on Advanced Communication Technology (ICACT),
2008:679–681.
[4] A. Juliandri, M. Musida and Supriyadi, ‘Positioning cloud computing in
machine to machine business models’, In Cloud Computing and Social
Networking (ICCCSN), 2012:1–4.
[5] V. Goncalves and P. Dobbelaere, ‘Business Scenarios for Machine-to-
Machine Mobile Applications’, In Mobile Business and 2010 Ninth
Global Mobility Roundtable (ICMB-GMR), 2010.
312 M2M in Agriculture – Business Models and Security Issues

[6] Y. Chang, T. Chi, W. Wang and S. Kuo, ‘Dynamic software update

model for remote entity management of machine-to-machine service
capability’, In IET Communications Journal, 2012.
[7] S. Blank, G. Kormann and K. Berns, ‘A Modular Sensor Fusion
Approach for Agricultural Machines’, In XXXVI CIOSTA\& CIGR
Section V Conference, Vienna, 2011.
[8] M. Mau, ‘Supply Chain Management in Agriculture - Including Eco-
nomics Aspects like Responsibility and Transparency’, In X EAAE
Congress Exploring Diversity in European Agriculture, 2002.
[9] S. Gansemer, U. Grossmann, B. Horster, T. Horster-Moeller and
C. Rusch, ‘Machine-to-machine communication for optimization of
information chain in agricultural business’, In 7th IEEE International
Conference on Intelligent Data Acquisition and Advances Computing
Systems, Berlin, 2013.
[10] K. Johansson, ‘Cost efficient provisioning of wireless access: Infrastruc-
ture cost modeling and multi-operator ressource sharing’, Thesis KTH
School of Electrical Engineering, Stockholm, 2005.
[11] A. Vazsonyi, ‘The use of mathematics in production and investory
control’, Management Science, 1 (1), Jan. 1955.
[12] R.K. Ko, S. S. G. Lee and E. W. Lee, ‘Business process management
(BPM) standards: a survey’, Business Process Mansgement Journal,
15(5):744–791, 2009.
[13] J. Sell, ‘Konzeption einer Ende-zu-Ende Absicherung für eine M2M-
Telematik Anwendung für die Firma Claas’, Thesis FH Dortmund,
Dortmund, 2013.
[14] E. Eren and K. Detken, ‘Mobile Security. Risiken mobiler Kommunika-
tion und Lösungen zur mobilen Sicherheit’, Wien, Carl Hanser Verlag,
2006.
[15] E. Eren and G. Aljabari, ‘Virtualization of Wireless LAN Infrastruc-
tures’, In 6th IEEE Workshop on Intelligent Data Acquisition and
Advanced Computing Systems, Prague, 2011.
 Index

A common framework 257, 282, 290

activation function 152 communications 3, 14, 113, 236, 258
Agriculture 295, 299 complementary filter 57
alarm 225, 231, 241, 258 complete data 30, 36, 41
algorithm 30, 77, 95 computing 61, 229, 235, 253
ambient 235, 245, 252 conductive rubber 173, 180
ambient monitoring 235 control 3, 15, 58, 96, 104
anchoring 75, 288 control architecture 123, 135
architecture 1, 86, 237, 277 control system 27, 51, 134, 156
artificial curiosity 74, 148 controller 54, 149, 206
artificial intelligence 74, 148 converter 13, 181, 218, 247, 279
Artificial Neural Networks 60 Coordinate transformation 204
assembly 96, 97
automated interpretation 73 D
autonomous 2, 21, 49, 73, 126 damping 59, 62, 149
autonomous learning 73, 75 data acquisition 49, 52, 173, 193, 244
autonomous robots 2 data processing 33, 193, 199, 243
autonomous systems 49 dead spot coverage 308
axiomatic theory 211, 213 Decision-making agency 215
disturbance 27, 39, 47
B driving module 125, 132
beacon 124, 136, 143
belief 76, 81, 87 E
body sensor 235, 238, 254 Edge detection 193, 207
body sensor network 235, 238, 253 effector 6, 17, 127
Business case 295, 300, 306 efficient robot control 123
business model 295, 302, 311 electromagnetic wave 217, 221
embedded 13, 52, 235, 254
C energy 3, 21, 59, 235, 261
calibration 64, 217, 242 environment 1, 83, 104, 194, 223
camera 14, 85, 98, 111, 194 epistemic curiosity 76, 82
CAN 13, 195, 297
capacity sensor 167, 188 F
Certificat Authority (CA) 307 factorization 32, 44
chain based architecture 3 feature 1, 10, 112, 118, 240
clamping force 169, 184 femtocell 235, 240, 300
clamping gripper 184, 187 first variation 32

313
314 Index

fitness 78, 82 infrared sensors 196

forecasting 68 input 27, 46, 87, 127, 148
fraction-rational matrix 32 In-Range filter 199
FSM 96, 99 intelligent 73, 95, 147, 167, 211
functional 7, 31, 168, 245 intelligent robot 74, 167, 184, 212
intelligent sensory system 211
G intelligent system 73, 106, 211, 232
gas 229, 235, 282 interfaces 85, 95, 298
genetic algorithm 77, 81 internal and external services 304
gozintograph 304 interpretation 73, 78, 81, 217
graphical user interface 55
gripper 4, 98, 169, 182, 280 J
JRE 283
H
Hall sensor 174, 178 K
hardware 70, 114, 205, 222, 254 kinematics 3, 27, 46, 130, 153
heading control 51, 58
heterogeneous architecture 4, 16 L
heterogeneous modular architecture 4, lane detection 193, 199
9, 21 language 75, 97, 102, 283
hidden layer 60, 150, 156 LASCOS 221, 225, 227
high-level commands 8 lattice robots 2
high-pass filter 58 learning 73, 95, 123, 134, 140
holon 126, 129 learning algorithm 149, 254
Holonic agents 126, 223 low-pass filter 58
holonic homogenous multi-agent system
123, 130, 142 M
home monitoring 235, 244, 253 M2M 295, 298, 303
homogeneous architectures 3, 5 magnetic SDS 175
Hough transformation 193, 200 manipulator 1, 16, 42, 96, 287
human-machine 95, 118 Manipulator with parallel kinematics 27
humanoid 73, 85, 188 Matlab 38, 45, 206, 253
hybrid 237, 241, 264 metrology 213, 232, 278
hybrid file encryption 310 MFM 98, 104, 119
microcontrollers 14, 193, 197
I mobile platform 27, 123, 135
IEEE 802.15.4 237 model 29, 47, 52, 106, 124
image 18, 33, 73, 83, 101 modelling 62, 211, 257
image processing 109, 193, 207 modular architecture 1, 9, 21
implementation 1, 9, 51, 85, 235 modular manipulators 3, 16
IMU 45, 56, 69 modular mobile robotic systems 2
incomplete data 33, 41 Modular robots 7, 13, 19
indicator 183 monitoring 51, 115, 173, 211, 223
industrial automation 1, 16 mote 247
information processing 74, 167, 181, 243 motion 8, 36, 50, 97, 169, 281
 Index 315

multi-agent decomposition 125, 132 Polymetric signal 211, 219

multi-agent system 125, 142, 296 polymetric system 212, 221, 230
multi-component robotic systems 2 power consumption 12, 123, 236, 249
multilayer 60, 147, 253 process transparency 300, 307
multilevel 98, 235
multi-robot systems 1 Q
multivariable object 35 Q-function 129, 143
Q-learning 123, 143
N
natural speech 75 R
network communication 298 real time 53, 73, 179, 213, 257
network nodes 272 regulator 148, 160
network ambient assisted living 235 reinforcement learning 123, 134
neural network 60, 147, 253 remote software update 300, 302
neural network speed gradient Remotely Operated Vehicles 257, 272
method 147 reproduction 95, 112, 116
neuron 60, 69, 151 resonance 49, 60, 68
nodes 9, 14, 238, 259, 290 roaming 299
robot 1, 73, 149, 167, 182
O robot control 95, 123, 167, 214
Object modelling 257 robot dynamics 147, 159
observation 78, 82, 175, 273 robot model 149, 167
observation point 174, 177 roll 50, 66
obstacle 18, 86, 110, 182 roller 172, 187
oceanographic data collection 257 rotational module 4, 10, 20
OpenCV 199 routing 97, 193, 235, 278
operation data acquisition 295, 300 ROV 257, 278
optimal speed 137, 139
output 27, 44, 183, 275 S
output layer 60, 150, 162 SADCO 212
Safe-Nets 259, 264, 283, 290
P saliency 73, 77
parallel kinematic 27, 41 scale model 49, 119, 194
parametric 49, 61, 160 scenario 80, 127, 148, 249, 259
parametric resonance 49, 61 Scilab 206
perception 76, 98, 125, 212 security 9, 21, 64, 179, 236
perceptual curiosity 76 secure firmware update 300, 309
Perry Slingsby 259 semantic robot-human interaction 73
PID controller 56 sensitized glove 99, 117
poles 30, 47, 288 sensor system 167, 215
pollution 257, 288 sensors 6, 52, 99, 167, 193
polymetric measurement 211, 215 sensory system 168, 211, 230
polymetric perceptive agent 215 separation 33, 96, 124, 174, 267
polymetric sensing 216, 231 servo 11, 54, 196
polymetric sensor 211, 221, 229 service links 305
316 Index

Ship scale model 49, 53, 61 trajectory 100, 124, 153, 184
simulation 46, 83, 123, 153, 177 transfer function 30, 47, 57
Simulink 45, 206 transient process 155, 160
slider module 8, 11, 15, 19 trial motion 168, 173, 187
slip displacement 168, 172 two way HTTPS-conenction 307
slippage registration 168
software 53, 86, 141, 286, 300 U
space robot 95, 104 UART 198
spectral density 33, 44 uncertain dynamics 147
speed control 55, 141 uncertainty 77, 147, 160
speed gradient 147, 165 underwater component 257
stability diagrams 61, 67 Underwater robot 149, 156, 165, 277
Standardization 296 underwater sensors 257, 268, 282
static characteristic 174, 178 USART 198
stroboscopic transformation 218
structural identification 27, 33, 47 V
system architecture 51, 235, 245 value chain 302, 309
variable costs 305
T VB. NET 54
tactile sensor 112, 167 vector 28, 36, 107, 153, 306
task training by demonstration 95, 99, vehicle 124, 133, 193, 207, 257
113, 115 virtual coordinate frame 132, 135
TDR 216 Visual saliency 73
teaching 101, 106, 148 VMAX 259, 278, 283
Telecontrol 95, 104, 113
telescopic module 8, 11, 17 W
testing 22, 49, 59, 88, 194 weight coefficient 148, 156, 164
tests 3, 49, 62, 214 wireless sensor network 235, 242, 253
topology 3, 16, 97, 138, 236 WLAN security 300, 308
towing tank 49, 70
towing tank tests 49, 62 X
training 64, 95, 102, 110, 290 XBEE 198
training by demonstration 95, 102, 119
 Editor’s Biographies

Richard J. Duro received the B.Sc., M.Sc., and Ph.D. degrees in physics
from the University of Santiago de Compostela, Spain, in 1988, 1989, and
1992, respectively.
He is currently a Full Professor in the Department of Computer Science
and head of the Integrated Group for Engineering Research at the University
of A Coruña, Coruña, Spain. His research interests include higher order neural
network structures, signal processing, and autonomous and evolutionary
robotics.

Yuriy P. Kondratenko, Doctor of Science, Professor, Honour Inventor of

Ukraine (2008), Corr. Academician of Royal Academy of Doctors (Barcelona,
Spain), Professor of Intelligent Information Systems at Petro Mohyla Black
Sea State University, Ukraine. He has received a Ph.D. (1983) and a Dr.Sc.
(1994) in Elements and Devices of Computer and Control Systems from
Odessa National Polytechnic University. He received several international
grants and scholarships for conducting research at Institute of Automation
of Chongqing University, P.R.China (1988–1989), Ruhr-University Bochum,
Germany (2000, 2010), Nazareth College and Cleveland State University,
USA (2003). His research interests include robotics, automation, sensors
and control systems, intelligent decision support systems, fuzzy logic, soft
computing, elements and devices of computing systems. He is the principal
researcher of several international research projects with Spain, P.R. of China
et al. and author of more than 120 patents and 12 books (including author’s
chapters in monographs) published by Springer, World Scientific, Pergamon
Press, Academic Verlag etc. He is a member of the GAMM, DAAAM, AMSE
UAPL and PBD-Honor Society of International Scholars and visiting lecture
at the universities in Rochester, Cleveland, Kassel, Vladivostok and Warsaw.

317
 Author’s Biographies

Francisco Bellas received the B.Sc. and M.Sc. degrees in physics from the
University of Santiago de Compostela, Spain, in 1999 and 2001, respectively,
and the Ph.D. degree in computer science from the University of A Coruña,
Coruña, Spain, in 2003.
He is currently a Profesor Contratado Doctor at the University of A
Coruña. He is a member of the Integrated Group for Engineering Research
at the University of A Coruña. His current research interests are related
to evolutionary algorithms applied to artificial neural networks, multiagent
systems, and robotics.

Mitru Corneliu Caraivan has received his B.Sc. degree in Automatic Control
and Computers in 2009, from Politehnica University of Bucharest, Romania.
Following the master thesis in Fakultät für Inginieurwissenshaften, Univer-
sity of Duisburg-Essen, Germany, he earned the Ph.D. degree in Systems
Engineering in 2013 from Politehnica University of Bucharest. Since 2009
he is part-time Assistant Professor in the Faculty of Applied Sciences and
Engineering, Ovidius University of Constanta, Romania. His main research
interests focus on offshore oil and gas automatic control systems on drilling
and exploration rigs, programmable logic controller networks, instrumentation
data sensors and actuators, systems redundancy and reliability. Since 2010,
he has gained experience on offshore jack-up rigs as IT/Electronics Engineer
focusing on specific industry equipment: NOV Amphion Systems, Cameron
(former TTS-Sense) X-COM Cyber-Chairs 5th gen, drilling equipment instru-
mentation, PLCs, SBCs, fire and gas alarm systems, satellite communications
and networking solutions.

Valentin Dache has received his B.Sc. degree in Automatic Control and
Computers in 2008, from the Politehnica University of Bucharest, Romania.
Following the master thesis in the same university, he is currently a Ph.D.
student in Systems Engineering, having an on-going research which focuses
on intelligent buildings management system. Other interests include the field

319
320 Author’s Biographies

of remotely controlled radiation scanners and non-intrusive inspection systems

for customs control and borders security - ROBOSCAN and ROBOSCAN 2M
AERIA systems - two times winner of the Grand Prix of Salon International
des Inventions de Genève, in 2009 and 2013. Working in the team which won
the International Exhibition of Inventions he also owns one invention patent.

Álvaro Deibe Díaz received the M.S. degree in Industrial Engineering in 1994
from the University of Vigo, Spain, and a Ph.D. in Industrial Engineering in
2010 in the University of A Coruña. He is currently Titular de Universidad in
the Department of Mathematical and Representation Methods at the same
University. He is a member of the Integrated Group for Engineering Research
at the University of A Coruña. His research interests include Automation and
Embedded Systems.

Alexander A. Dyda was graduated from Electrical Engineering Department

of Far-Eastern Polytechnic Institute (Vladivostok) in 1977. He received degree
of Candidate of Science (PhD) from Leningrad Electrical Engineering Institute
in 1986. From 1979 to 1999 he was Assistant Professor, principal lecturer and
Associate Professor of Department of Technical Cybernetics & Informatics.
In 1990–1991 and 1995 he was researcher in the Department of System
Sciences at “La Sapienza” University (Rome, Italy) and the Department
of System Science & Mathematics of Washington University (Saint-Louis,
USA), respectively. In 1998 he received degree of Doctor of Technical
Sciences (DSc) from Institute of Automation & Control Processes, Russian
Academy of Sciences. From 1999 to 2003 he was Chairman of Department
of Information Systems. Since 2003 he is Full Professor of Department of
Automatic & Information Systems and head of the laboratory of Nonlinear &
Intelligent Control Systems at Maritime State University.

Uladzimir Dziomin is a PhD student at Intelligent Information Technology

Department of Brest State Technical University. He received MSc and
Diploma in Information Technology at BrSTU. The area of his research is
Intelligent Control, Autonomous Learning Robots and Distributed Systems.

Chernakova Svetlana Eduardovna was born on August 3, 1967 in the

town of Kalinin in the Kalinin Region. She graduated from the Leningrad
Polytechnic Institute in 1990, has written more than 80 scientific papers,
and possesses considerable knowledge related to development and cre-
ation of television systems for different purposes. Area of interest: pattern
 Author’s Biographies 321

identification, intelligent systems “human-to-machine”, intelligent technol-

ogy of training by demonstration, mixed (virtual) reality. Place of work:
St. Petersburg Institute of Informatics of the Russian Academy of Sci-
ences (SPIIRAS), St. Petersburg E-mail: S [email protected] Address:
199226, Saint-Petersburg, Nalichnaya Street, H# 36, Building 6, Apt 160

Andrés Faíña received the M.Sc. and Ph.D. degrees in industrial engineering
from the University of A Coruña, Spain, in 2006 and 2011, respectively.
He is currently a Postdoctoral Researcher at the IT University of
Copenhagen, Denmark. His interests include modular and self-reconfigurable
robotics, evolutionary robotics, mobile robotics, and electronic and mechani-
cal design.

Vladimir Golovko, Professor, is a Head of Intelligent Information Technolo-

gies Department and Laboratory of Artificial Neural Networks since 2003.
He received his PhD degree from the Belarus State University and Doctor of
Sciences degree in Computer Science from United Institutes of Informatics
Problem of National Academy of Sciences (Belarus). His research inter-
ests include Artificial Intelligence, Neural Networks, Autonomous Learning
Robots, Intelligent signal processing.

Marcos Miguez González received his master degree in Naval Architecture

and Marine Engineering in 2004, from the University of A Coruña, Spain,
and his PhD related with the analysis of parametric roll resonance for the
same university in 2012. He worked for three years in shipbuilding industry.
Since 2007, he is a researcher in the GII (Integrated Group for Engineering
Research) of the University of A Coruña, formerly under a scholarship from
the Spanish Ministry of Education and nowadays as an assistant professor. His
research topics are ship behaviour modelling, parametric rolling and stability
guidance systems.

Boris Gordeev received the PhD degree from Leningrad Institute of High-
precision Mechanics and Optics, Leningrad, USSR, in 1987 and Doctor of
Science degree from Institute of Electrodynamics of National Academy of
Sciences of Ukraine in Elements and Devices of Computer and Control
Systems in 2011.
He is currently a Professor of Marine Instrumentation Department at
National University of Shipbuilding, Ukraine. His current research interest is
related to polymetric signals generation and processing for measuring systems,
322 Author’s Biographies

including control systems for continuous measurement of the quantitative

and qualitative parameters of the fuels and liquefied petroleum gas. He is
the author of more than 150 scientific publications in the above mentioned
field.

Anton Kabysh, assistant of Intelligent Information Technologies Department

of Brest State Technical University. He is working on Ph.D thesis in Multi-
Agent Control and Reinforcement Learning at Robotics Laboratory in BrSTU.
Research interest also includes distributed systems, swarm intelligence and
optimal control.

Volodymyr Y. Kondratenko received a PhD degree in Applied Mathematics

from Colorado University Denver in 2015. He has received a B.S. summa
cum laude in Computer Science and Mathematics at Taras Shevchenko Kyiv
National University (2008), and a M.S. at University of Colorado (2010) as a
Winner of Fulbright Scholarship. He is the author of numerous publications
in such fields as Wildfire Modelling and Fuzzy logic, Medallist of the
Small Academy of Science of Ukraine (2004, 2nd place) and winner of All-
Ukrainian Olympiad in mathematics (2004, 3rd place). His research interests
include Computational Mathematics, Probability Theory, Data Assimilation,
Computer Modelling, and Fuzzy Logic.

Sauro Longhi received the Doctor degree in Electronic Engineering in 1979

from the University of Ancona, Italy, and post-graduate Diploma in Automatic
Control in 1984 from the University of Rome “La Sapienza”, Italy. From 1980
to 1981 he held a fellowship at the University of Ancona. From 1981 to 1983
he was with the R&D Laboratory of the Telettra S.p.A., Chieti, Italy, mainly
involved in activities of research and electronic design of modulation and
demodulation systems for spread spectrum numerical transmission systems.
Since 1983 he has been at the Dipartimento di Elettronica e Automatica of
the University of Ancona, now Information Engineering Department, of the
Università Politecnica delle State Marche-Ancona. From July 2011 to October
2013 he was the Director of this Department. Since November 1, 2013 he is
the Rector of the Università Politecnica delle Marche.

Andrei Maciuca obtained his PhD in 2014 from the Department of Automatic
Control and Industrial Informatics, University “Politehnica” of Bucharest with
a thesis on “Sensor Networks for Home Monitoring of Elderly and Chronic
Patients”.
 Author’s Biographies 323

Kurosh Madani received his PhD degree in Electrical Engineering and

Computer Sciences from University Paris-Sud, Orsay (France) in 1990.
In 1995, he received the Doctor Habilitate degree (Dr Hab.) from UPEC.
Since 1998 he has worked as Chair Professor in Electrical Engineering at
University Paris-Est Creteil (UPEC). Co-initiator of the Images, Signals and
Intelligent Systems Laboratory (LISSI/EA 3956), he is head of one of the four
research groups of LISSI. His current research interests include bio-inspired
perception, artificial awareness, cognitive robotics, human-like robots and
intelligent machines. Since 1996 he is electedAcademician of the International
Informatization Academy. In 1997 he was elected as Academician of the
International Academy of Technological Cybernetics.
Kulakov Felix Mikhailovich was born on July 4, 1931 in the town of
Nadezhdinsk in Ural Region. He graduated from the Leningrad Polytechnic
Institute in 1955, and has written more than 230 scientific papers. He works in
the sphere of robot supervision and research of automatic performance of the
mechatronic systems. He is an Honored Worker of Science of the Russian
Federation, Professor, Chief Research scientist, St. Petersburg Institute of
Informatics of the Russian Academy of Sciences (SPII RAS), head of the
branch of the department of the “Mechanics of the controlled movement” of
the St. Petersburg State University at SPIIRAS. E-mail: [email protected]
Address: 198064, Saint-Petersburg, Tikhoretsky Avenue, H# 11, building 4,
Apt 50
Andrea Monteriù received the Laurea Degree (joint BSc/MSc equivalent)
summa cum laude in Electronic Engineering and the Ph.D. degree in Artificial
Intelligence Systems from the Università Politecnica delle Marche, Ancona,
Italy, in 2003 and 2006, respectively. His MSc thesis has been developed to
the Automation Department of the Technical University of Denmark, Lyngby,
Denmark. In 2005 he was a visiting researcher at the Center for Robot
Assisted Search & Rescue of the University of South Florida, Tampa, Florida.
Since 2005, he is Teaching Assistant of Automatic Control, Automation
Systems, Industrial Automation, Modelling and Identification of Dynamic
Processes. Since 2007, he has a PostDoc and Research Fellowship at the
Dipartimento di Ingegneria dell’Informazione of the Università Politecnica
delle Marche, where currently he is a Contract Professor.

Alexander Nakonechniy received the B.Sc. and M.Sc. degrees in High

Voltages Electro-physics from National University of Shipbuilding, Ukraine,
in 2007 and 2009 respectively.
324 Author’s Biographies

He is currently a post-graduate student at Marine Instrumentation Depart-

ment of National University of Shipbuilding, Ukraine. His current research
interests are polymetric measurements, sensors and sensors networks, gener-
ation of short pulses. He is the author of several scientific publications in the
above fields.

Sergey I. Osadchy Doctor of Science, Professor, Chief of Production Pro-

cess Automatic Performance Department at Kirovograd National Technical
University, Ukraine. He has received a Ph.D. (1987) and a Dr.Sc. (2013) in
Automatic Performance of Control Processesfrom National Technic Univer-
sity of Ukraine “Kiev Polytechnic Institute”. His research interests include
robotics, automation, sensors and optimal control systems synthesis, analysis
and identification, underwater supercavitating vehicle stabilization and con-
trol. He is the author of more than 140 scientific articles, 25 patents and 4 books.

Dmitry A. Oskin was graduated from Electrical Engineering Department of

Far-Eastern State Technical University (Vladivostok) in 1997. He received
degree of Candidate of Science (PhD) in 2004. From 2000 to 2005 he was
an Assistant Professor of Department of Information Systems. Since 2005, he
is Associate Professor of Department of Automatic & Information Systems
and Senior researcher of the laboratory of Nonlinear & Intelligent Control
Systems at Maritime State University.

Fernando López Peña received the Master degree in Aeronautical Engineer-

ing from the Polytechnic University of Madrid, Spain, in 1981, a Research
Master from the von Karman Institute of Fluid Dynamics, Belgium, in 1987,
and the Ph.D. degree from the University of Louvain, Belgium, in 1992. He
is currently a professor at the University of A Coruña, Spain. He authored
about 100 papers on peer reviewed journals and international conferences,
holds five patents, and has leaded more than 80 research projects. His current
research activities are related to intelligent hydrodynamics and aerodynamics
design and optimization, flow measurement and diagnosis, and signal and
image processing.

Dan Popescu obtained the MSc degree in Automatic Control in 1974, MS

in mathematics in 1980, PhD degree in Electrical Engineering in 1987. He
is currently Full Professor, in the Department of Automation and Industrial
Informatics, University “Politehnica” of Bucharest, head of the labora-
tory Artificial Vision and responsible for master program Complex Signal
 Author’s Biographies 325

Processing in Multimedia Applications. He is PhD adviser in the field of

Automatic Control. Current scientific areas are: equipments for complex
measurements, data acquisition and remote control, wireless sensor networks,
alerting systems, pattern recognition and complex image processing, interdis-
ciplinary approaches. He is author of 15 books, more than 120 papers, and
director of five national research grants. He is IEEE member, SRAIT member,
and he received the 2008 IBM Faculty Award. His research interests lay in
the field of wireless sensor networks, image processing and reconfigurable
computing.

Dominik Maximilián Ramík received his PhD in 2012 in signal and image
processing from University of Paris-Est, France. Member of LISSI laboratory
of University Paris-Est Creteil (UPEC), his current research topic con-
cerns processing of complex images using bio-inspired artificial intelligence
approaches and consequent extraction of semantic information with use in
mobile robotics control and industrial processes supervision.

Christian Rusch studied Mechatronics between 1999 and 2005 and grad-
uated from the Technical University of Braunschweig. Then he started
his scientific career as Research Scientist at the Technical University of
Berlin. The Title of his PhD Thesis is “Analysis of data security of self-
configuring radio networks on mobile working machines illustrated by the
process documentation”. He finished the PhD (Dr. –Ing.) in 2012. In 2011
he moved in to the industry and worked for CLAAS as project manager with
focus on wireless communication for mobile working machines. Since 2014 he
is system engineer in CLAAS Electronic Systems. Furthermore, he is project
leader of the standardization group “Wireless ISOBUS Communication”.

Félix Orjales Saavedra received the master degree in Industrial Engineering

in 2010 from the University of León, Spain. He is currently working on his
Ph.D. degree in the Department of Industrial Engineering at the University of
A Coruña, Spain, related with collaborative unmanned aerial vehicles. He is
a researcher at the Integrated Group for Engineering Research, and his main
research topics include autonomous systems, robotics, and electronic design.

Christophe Sabourin received his PhD in Robotics and Control from Uni-
versity of Orleans (France) in November 2004. Since 2005, he has been
a researcher and a staff member of Images, Signals and Intelligent Sys-
tems Laboratory (LISSI/EA 3956) of University Paris-Est Creteil (UPEC).
326 Author’s Biographies

His current interests relate to areas of complex and bio-inspired intelligent

artificial systems, cognitive robotics, humanoid robots, collective and social
robotics.

Valentin Sgârciu is full professor in the Faculty of Automatic Control and

Computers, Politehnica University of Bucharest, Romania since 1995 and PhD
coordinator in the same university since 2008. His teaching activity includes
undergraduate and postgraduate courses: “Sensors and Transducers”, “Data
Processing”, “Transducers and Measuring systems”, “Reliability and Diagno-
sis”, “Security of Informatics Systems” and “Product and System Diagnosis”.
His research activity include the fields of intelligent buildings, intelligent
measurements, wireless sensor networks, software engineering, operating
systems, control system theory, e-learning, distributed heterogeneous systems
and middleware technologies. He has over 100 scientific papers presented
in national and international symposiums and congresses, published in the
associated preprints or in specialized Romanian and international journals.
He is first author of 16 books and owns one invention patent.

Daniel Souto received the M.Sc. degree in industrial engineering from the
University of A Coruña, Coruña, Spain, in 2007. He is working towards
the Ph.D. degree in the Department of Industrial Engineering at the same
university.
He is currently a Researcher at the Integrated Group for Engineering
Research. His research activities are related to automatic design and mechan-
ical design of robots.

Grigore Stamatescu is currently Assistant Professor within the Department

of Automatic Control and Industrial Informatics, University “Politehnica”
of Bucharest – Intelligent Measurement Technologies and Transducers lab-
oratory. He obtained his PhD in 2012 from the same institution with a
thesis on “Improving Life and Work with Reliable Wireless Sensor Net-
works”. He is a member of the IEEE Instrumentation and Measurement
and Industrial Electronics societies and has published over 50 papers in
international journals and indexed conference proceedings. His current
research interests lay in the fields of intelligent sensors, embedded networked
sensing and information processing.

Ralf Stetter is a Professor in the Department of Mechanical Engineering

at the Hochschule Ravensburg-Weingarten since 2004. His Teaching Area
 Author’s Biographies 327

is “Design and Development in Automotive Technology”. He received his

PhD and Diploma in Mechanical Engineering from the Technische Universität
München (TUM). He is currently Vice-Dean of the Department of Mechanical
Engineering. He works as Project Manager in the Steinbeis-Transfer-Centre
“Automotive Systems”. He was Team Coordinator at Audi AG, Ingolstadt in
the Automotive Interior Product Development.

Mircea Strutu obtained his PhD in 2014 from the Department of Automatic
Control and Industrial Informatics, University “Politehnica” of Bucharest with
a thesis on “Wireless Networks for Environmental Monitoring and Alerting
based on Mobile Agents”.

Anna S. Timoshenko, Lecturer in Department of Information Technology

at Kirovograd Flight Academy of National Aviation University. She grad-
uated from the State Flight Academy with honors in 2002. She graduated
from the Kirovograd National Technical University in 2006. Then she took
post-graduate studies in Kirovograd Flight Academy of National Aviation
University (2009–2012). She is the author of numerous publications in such
fields as Navigation and Motion Control, Automation and Robotics.

Blanca María Priego Torres received the title of Telecommunications Engi-

neer in 2009 from the University of Granada, Spain. In 2011, she obtained the
Master’s Degree in Information and Communications Technologies in Mobile
Networks from the University of A Coruña, Spain. She is currently working
in her PhD as a member of the Integrated Group for Engineering Research
at the University of A Coruña. Her research interests include the analysis,
processing and interpretation of signals and multi-dimensional images in
relation to the Industrial and Naval Field, and through the application of
techniques based on new neural-network structures, soft computing and
machine learning.

Yuriy Zhukov received the PhD and Doctor of Science degrees from
Nikolayev Shipbuilding Institute (now NUOS), Ukraine, in 1981 and 1994,
respectively.
He is currently a Professor, Chief of Marine Instrumentation Department
and Neoteric Naval Engineering Institute at National University of Shipbuild-
ing, Ukraine. He is the author of more than 250 scientific publications in fields
of ship design and its operational safety, dynamic systems behavior monitoring
328 Author’s Biographies

and control, decision making support systems and artificial intelligence, etc.
His current research interest is related to intelligent multi-agent sensory
systems application in the above mentioned fields.

Alexey Zivenko received the PhD degree in Computer Systems and Compo-
nents from Petro Mohyla Black Sea State University, Ukraine, in 2013.
He is currently an Associate Professor of Marine Instrumentation Depart-
ment at National University of Shipbuilding, Ukraine. His current research
interests are polymetric measurements, non-destructive assay of liquids, intel-
lectual measurement systems and robotics, data acquisition systems anddata
mining. He is the author of more than 20 scientific publications in the above
mentioned fields.

Valerii A. Zozulya, Associate Professor in Department of automation of

production processes at Kirovograd National Technical University, Ukraine.
He has received a Ph.D. (1999) and received the academic rank of associate
professor (2003) at Kirovograd National Technical University. He is the author
of numerous publications in such fields as motion control, robotics and optimal
control systems synthesis, analysis and identification.