Human–Robot Interaction Based on

Motion and Force Control

Martin Karlsson

Department of Automatic Control

PhD Thesis TFRT-1124
ISBN 978-91-7895-016-4 (print)
ISBN 978-91-7895-017-1 (web)
ISSN 0280–5316

Department of Automatic Control

Lund University
Box 118
SE-221 00 LUND
Sweden

© 2019 by Martin Karlsson. All rights reserved.

Printed in Sweden by Media-Tryck.
Lund 2019
Abstract

Industrial robots typically require detailed programming and carefully con-

figured work cells to perform well. The large engineering effort implicates
high cost and long preparation time, and this is the major obstruction when
mediating tasks to robots. The research in this thesis therefore aims to
make robot programming faster and more accessible. Methods that allow
for programmers to mediate and modify tasks by means of demonstration
are presented. Further, robots’ abilities to replan with respect to unfore-
seen changes in their surroundings are enhanced, thus lowering the effort
needed for work-cell configuration.
We first consider adjustment of robot movements generated by dynamical
movement primitives (DMPs). DMPs are motion-control laws with emphasis
on easy modification. For instance, goal configuration and time scale for a
certain movement can be updated through one parameter each, commonly
without further consideration. In this research, these capabilities are ex-
tended to support modifications based on demonstrations through physical
human–robot interaction. Further, the motion-control laws are extended to
support online replanning for overcoming unforeseen movement deviations.
Subsequently, a method that enables robots to recognize contact
force/torque transients acting on the end-effector, without using a
force/torque sensor, is proposed. This is achieved using machine learn-
ing. The robot is first exposed to examples of force/torque transients. Based
on these data, a recurrent neural network (RNN) is trained to recognize
such transients. The functionality is used to automatically determine when
a robotic subtask is finished, to proceed to the next subtask at the right
time. Finally, a control algorithm for teleoperation with force feedback is
developed. It allows for an operator to demonstrate movement and forces
remotely. One robot arm is moved directly through physical contact with
the operator, and a distant robot arm moves accordingly. Interaction forces
are reflected to each side of the interface.
Each of the methods presented in this thesis is implemented in a real-
time application and verified experimentally on an industrial robot.

3
Acknowledgments

I would like to thank my supervisors Prof. Rolf Johansson and Prof. Anders
Robertsson, for invaluable advice and support throughout this work. You
have given me insights in all aspects of robot research, including theory,
implementation, and presentation of results. Thanks to your ability to give
me good advice in any situation, while also giving me the freedom to define
my work, I have looked forward to every next day of this journey.
Prof. Bo Bernhardsson, I am happy that I took your advice to work at
the department, both times. First as a teaching assistant, and then as a
PhD student. I would also like to thank you and Prof. Fredrik Tufvesson
again for the guidance as my Master’s thesis supervisors many years ago.
Dr. Fredrik Bagge Carlson, it has been great to have you as a close
colleague. It has been fun to improve robots together with you. We have a
common interest not only in our work, but also in solid-state physics and
especially defected semiconductors, and I look forward to review this area
regularly together with you in the future.
Dr. Björn Olofsson, thank you for being a great role model and for
generously sharing insights about robotics in both theory and practice. Your
morale, good mood, and unquestionable competence set a great example.
Prof. Charlotta Johnsson, thank you for bringing me into the robotic
surgery project Surgeon’s Perspective. I really enjoy the work with you and
our project colleagues, and I highly appreciate your guidance in where to
direct our work. I would like to acknowledge M.D. Kiet Phan Tran for
introducing me to cardiac surgery, for sharing your knowledge extremely
generously, and for the interesting discussions we have about the intersec-
tion of robotics and surgery.
I would like to acknowledge my colleagues and former colleagues in the
Robotics Lab, Dr. Mahdi Ghazaei Ardakani, Dr. Anders Nilsson, Anders
Blomdell, Pontus Andersson, Asst. Prof. Mathias Haage, Prof. Jacek Malec,
Assoc. Prof. Elin Anna Topp, Assoc. Prof. Klas Nilsson, Dr. Maj Stenmark,
Dr. Magnus Linderoth, Dr. Olof Sörnmo, Martin Holmstrand, Matthias

5
Mayr, Prof. Volker Krüger, Julian Salt, Alexander Dürr, Dr. Karl Berntorp,
and Dr. Andreas Stolt. Thank you for great cooperation.
Dr. Mårten Wadenbäck, thank you for introducing me to the exciting
field of visual odometry, and for being a great collaborator and friend. It
was worth all the difficulties with network connections, camera settings,
and even two minor electric shocks.
Throughout the years I have had the honor and pleasure of sharing office
with many great colleagues. Thank you Olof Troeng, Dr. Björn Olofsson, Nils
Vreman, Martin Morin, Victor Millnert, Dr. Ola Johnsson, Lic. Tech. Jonas
Dürango, Irene Zorzan, Dr. Meike Rönn, Dr. Magnus Linderoth, Dr. Gabriel
Turesson, and Lic. Tech. Alina Andersson, for creating such a nice working
climate!
Prof. Karl Johan Åström, you are gratefully acknowledged for fruitful
discussions about this thesis, and about automatic control in general. You
are a great source of knowledge and inspiration at our department.
I am grateful to the administrative staff at the department, Ingrid Nils-
son, Mika Nishimura, Monika Rasmusson, Cecilia Edelborg Christensen,
and Dr. Eva Westin, for your help with various things related to work. You
are always so positive, well organized, and helpful. A special thanks to In-
grid and Monika, for always being supportive and understanding when I
did my best with project-time reporting.
Leif Andersson, thank you for helping me using LATEX to its full potential.
Without your aid, this thesis, as well as my other publications, would have
looked much worse. You have helped me to position figures correctly, number
equations and subequations, and to remove unnecessary LATEX code that I
naively copied from the Internet. On a related note, I thank Dr. Björn
Olofsson, Axel Karlsson, Dr. Fredrik Bagge Carlson, Julian Salt, Matthias
Mayr, and Dr. Mahdi Ghazaei Ardakani for proofreading of this thesis.
I also thank all colleagues at the department. Each one creates a posi-
tive, inspiring, and fun working environment, and contributes to interesting
discussions about control theory and related topics.
Last but not least, Daniel, Ann-Eli, Axel, and Johan Karlsson, have
always been a fantastic family. Heartfelt thanks for all the love and support!

6
Financial Support
Financial support is gratefully acknowledged from the VINNOVA project
Kirurgens Perspektiv (Surgeon’s Perspective in English). Financial sup-
port is also gratefully acknowledged from the European Commission, un-
der the Framework Programme Horizon 2020 within grant agreement No
644938–SARAFun, and under the 7th Framework Programme within grant
agreement No 606156–FlexiFab, as well as from the Swedish Foundation
for Strategic Research through the SSF project ENGROSS. The author is a
member of the LCCC Linnaeus Center, supported by the Swedish Research
Council, and the ELLIIT Excellence Center, supported by the Swedish Gov-
ernment.

7
Contents

1. Introduction 13
1.1 Background and Motivation . . . . . . . . . . . . . . . . . . 13
1.2 Robot Programming and Control . . . . . . . . . . . . . . . 15
1.3 Introduction to Machine Learning . . . . . . . . . . . . . . 16
1.4 Problem Formulation . . . . . . . . . . . . . . . . . . . . . 18
1.5 Contributions . . . . . . . . . . . . . . . . . . . . . . . . . . 19
1.6 Publications . . . . . . . . . . . . . . . . . . . . . . . . . . . 20
1.7 Thesis Outline . . . . . . . . . . . . . . . . . . . . . . . . . 22
1.8 Videos . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 24
2. Robots and Interfaces for Experimental Validation 25
2.1 YuMi Prototype Version . . . . . . . . . . . . . . . . . . . . 26
2.2 YuMi Product Version . . . . . . . . . . . . . . . . . . . . . 26
3. Dynamical Movement Primitives—An Overview 29
3.1 Related Research . . . . . . . . . . . . . . . . . . . . . . . . 32
3.2 Parameter Choices . . . . . . . . . . . . . . . . . . . . . . . 33
3.3 Comparison with Alternative Movement Representations . 34
4. Modification of Robot Movement Using Lead-Through Pro-
gramming 35
4.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 35
4.2 Motivating Examples . . . . . . . . . . . . . . . . . . . . . . 37
4.3 Problem Formulation . . . . . . . . . . . . . . . . . . . . . 39
4.4 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 41
4.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . 44
4.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 45
4.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 46
4.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 48
5. Temporally Coupled Dynamical Movement Primitives in Rn 49
5.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 49
5.2 Previous Research on Temporally Coupled DMPs for Pertur-
bation Recovery . . . . . . . . . . . . . . . . . . . . . . . . . 51

9
Contents

5.3 Problem Formulation . . . . . . . . . . . . . . . . . . . . . 53

5.4 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 55
5.5 Simulations . . . . . . . . . . . . . . . . . . . . . . . . . . . 57
5.6 The Proposed Temporally Coupled DMPs are Globally Expo-
nentially Stable . . . . . . . . . . . . . . . . . . . . . . . . . 59
5.7 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . 65
5.8 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 69
5.9 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 70
5.10 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
5.A Passivity of Temporally Coupled Dynamical Movement Prim-
itives . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 79
6. Temporally Coupled Dynamical Movement Primitives in
3D Orientation Space 83
6.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 83
6.2 Previous Research . . . . . . . . . . . . . . . . . . . . . . . 83
6.3 Problem Formulation . . . . . . . . . . . . . . . . . . . . . 84
6.4 The Unit Quaternion Set Minus One Single Point is Con-
tractible . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 84
6.5 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 89
6.6 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . 92
6.7 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 94
6.8 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 95
6.9 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 100
6.A Stability Analysis . . . . . . . . . . . . . . . . . . . . . . . . 100
7. Detection and Control of Contact-Force Transients 105
7.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 105
7.2 Problem Formulation . . . . . . . . . . . . . . . . . . . . . 107
7.3 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 109
7.4 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . 114
7.5 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 114
7.6 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 116
7.7 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 122
8. A Dual-Arm Haptic Interface for Task Demonstration 123
8.1 Introduction . . . . . . . . . . . . . . . . . . . . . . . . . . . 123
8.2 Problem Formulation . . . . . . . . . . . . . . . . . . . . . 125
8.3 Previous Research . . . . . . . . . . . . . . . . . . . . . . . 126
8.4 Method . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 127
8.5 Experiments . . . . . . . . . . . . . . . . . . . . . . . . . . 130
8.6 Results . . . . . . . . . . . . . . . . . . . . . . . . . . . . . . 131
8.7 Discussion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 133
8.8 Conclusion . . . . . . . . . . . . . . . . . . . . . . . . . . . . 136
8.A Proof of Kinematic Constraint . . . . . . . . . . . . . . . . 136

10
 Contents

9. Conclusion and Future Research 138

Bibliography 140

11
1
Introduction

This thesis considers human–robot interaction, both for robot programming

and for collaboration during program execution. The overall aim is to make
robot programming more accessible. The research is based on motion con-
trol, force control, and machine learning. Motion-control algorithms that
generate robot movement capable of adapting to unforeseen events are de-
veloped. Force control is used to allow for a programmer to demonstrate
desired movement and expected interaction forces. Further, machine learn-
ing is used to recognize interaction forces between robot and work object, in
order to detect key events during robotic manipulation. Such detection can,
for instance, be used to determine when a subtask has been accomplished,
in order to switch to the next subtask in the program.

1.1 Background and Motivation

The long-term ambition in the field of robotics is broad: Robots should do
work to enrich humanity. The word robot comes from robota, the Czech word
for work. In some cases, we would like robots to replace human workers.
There might be several reasons for certain tasks to be undesirable for
humans. Such tasks might be too time consuming, monotonous, unhealthy,
or even hazardous. In other cases, we would like robots to do work that can
not be done by humans, for instance due to physical limitations. Already
today, robots are used in a wide range of tasks. The automotive industry
is extensively robotized, and robots for vacuum cleaning and lawn mowing
are common in households. Robots are also used in surgery and planet
exploration, but in contrast to the previous examples, these are controlled
remotely by humans instead of operating completely autonomously; see,
e.g., [Spong et al., 2006; Siciliano et al., 2010] for an introduction to robot
modeling and control.
Industrial robots typically operate by performing sequences of predefined
movements, with high speed, high position accuracy, and without rest. This

13
Chapter 1. Introduction

has already enabled automation of repetitive tasks where position control

suffices, such as spray painting and welding. For such tasks, robots can be
programmed to outperform humans.
However, predefined robotic movements are meaningful only in care-
fully structured environments, specifically designed to fit the robot and
the given task. The task must therefore be highly repetitive, and possible
deviations from the original plan must have been foreseen by the robot pro-
grammer. Even under favorable conditions, traditional robot programming
is time consuming and requires expert knowledge. As a result, human labor
is still more cost effective than automation for many seemingly repetitive
tasks, such as assembly tasks. Electronic products such as smartphones
and robot vacuum cleaners are put together manually in large factories.
Unfortunately, these assembly tasks are too repetitive to be desirable for
humans, and at the same time insufficiently repetitive to be robotized.
Robotic assembly has been addressed in, e.g., [Thomas and Wahl, 2001;
Thomas et al., 2007; Björkelund et al., 2011; Stolt et al., 2012a; Linderoth,
2013; Stolt, 2015]. For human workers, these assembly tasks cause fatigue
and leave little room for personal development. Even though these tasks
appear monotonous for humans, small tolerances and tiny variations be-
tween similar assembly parts make it inadequate to just perform a series
of accurate movements. There is also a trend toward manufacturing a given
product in a smaller volume and for a shorter time, and then changing to a
new one. This shortens the acceptable preparation time even further.
Industrial robots are commonly strong and heavy, and can not safely
operate close to humans. There is nowadays a trend to build inherently safe
robots, with less weight and less powerful motors. YuMi [ABB Robotics,
2018], Panda [Franka Emika, 2018], Baxter [Rethink Robotics, 2018], and
LBR iiwa [KUKA, 2018], are examples of such collaborative robots, also
referred to as light-weight robots. This type of safe hardware is appropriate
for the physical human–robot interaction considered in this thesis.
Many tasks that humans do in society, such as buying groceries, cooking,
and cleaning, are not yet possible to fully automate. Even though dedicated
machines such as dish washers and vacuum cleaners exist, significant hu-
man engagement is still required to make these machines useful. Predefined
movement sequences would by no means work in an everyday environment,
because it is not sufficiently predictable. Instead, a general ability to adapt
to the surroundings is necessary. This requires a more general type of
intelligence than robots currently possess.
These circumstances have motivated the following research objectives:
1. Enable faster and more intuitive robot programming;
2. Enable robots to take suitable action with respect to their tasks and
surroundings.

14
 1.2 Robot Programming and Control

In this thesis, research toward these objectives is presented. The objec-

tives are partly overlapping. For example, if a robot can replan with respect
to its work space, the programmer does not have to consider all possible
scenarios in advance. This reduces the required programming competence
and effort. Vice versa, intuitive means of robot programming could allow
for an operator to mediate suitable behavior, given certain states or events,
to the robot, and thereby enhance the adaptability.
The research in this thesis has been done within the projects SARA-
Fun [SARAFun, 2019] and Surgeon’s Perspective. SARAFun is short for
Smart Assembly Robot with Advanced Functionalities, and the aim of the
project was to enable non-expert operators to program collaborative robots
for assembly tasks. SARAFun was concluded at a final review in April 2018,
where some of the research in this thesis was presented and demonstrated.
The aim of Surgeon’s Perspective is to develop robot-assisted surgery, and
the thesis author has been responsible for developing the algorithms for
robot control described in this thesis. The two projects overlap, as both aim
to facilitate robot programming and enhance online replanning capabilities.

1.2 Robot Programming and Control

Whereas humans would prefer to program on a high level of abstraction,
for instance through natural language, robots require very detailed instruc-
tions, for instance time series of desired joint torques. Robot programming
is therefore difficult in general. High-level instructions can work well be-
tween humans, but leave too many details unspecified to be suitable for
robot programming.
For a robot program to be realizable, it is necessary that appropriate low-
level quantities such as motor currents can be determined automatically and
unambiguously based on the program and on measurements or estimates.
Thanks to automatic control [Wen and Murphy, 1990; Wen and Murphy,
1991; Åström and Hägglund, 1995; Khalil, 2002; Åström and Wittenmark,
2013], of which basic knowledge is assumed in the following, robot programs
can be specified on more intuitive levels than that of motor currents. Some
applications require only that the robot reaches desired positions. It is then
sufficient to specify these positions. Motion control is used to determine the
corresponding joint torques in real time. In turn, inner control loops are
used to command motor currents that realize the joint torques. Similar to
motion control, desired interaction forces at the robot end-effector can be
achieved using force control, though this introduces additional challenges
as addressed in [Johansson et al., 2015].
Robot programming may consist of ordinary computer programming,
disregarding hardware considerations. Source code is written on a computer

15
Chapter 1. Introduction

and automatically transformed to executable machine code. To facilitate the

programming, graphical user interfaces such as ABB RobotStudio [ABB
Robotics, 2019b] have been introduced. Reference coordinate systems and
sequences of movement targets can be defined graphically and converted to
source code. It is also common to move the physical robot with a joystick
to desired configurations, and save these as part of the code. Still, many
operations are more suitable to define by direct source-code editing. Further,
programs can be tested with visual simulation tools such as RobotStudio,
rviz [ROS, 2018], and Gazebo [Gazebo, 2018]. These established means of
robot programming and simulation require special knowledge.
An easier way to mediate behavior is by means of demonstration. Robot
programming by demonstration is also referred to as robot learning from
demonstration, apprenticeship learning, and imitation learning. A survey
is given in [Argall et al., 2009]. Finding robust and general frameworks for
programming by demonstration is a current research challenge. One pos-
sibility is to let the user do the task, document it by, e.g., video recording,
and automatically generate a robot program that achieves the same thing.
However, in that kind of procedure it is difficult for the user to take all phys-
ical limitations of the robot, such as gripper design, joint limits, and reach
limits, into account. A natural way to incorporate these limitations is to
instead let the user guide the robot through the intended task, for instance
by means of lead-through programming where the robot is moved through
physical contact with the user [Pan et al., 2010; Stolt et al., 2015a; Capurso
et al., 2017]. The intended configurations of the robot will then be unam-
biguously demonstrated, at least given the work-space configuration. This
is the kind of robot programming considered in this thesis. Demonstrated
movement can be represented and replayed by, e.g., dynamical movement
primitives (DMPs), see [Ijspeert et al., 2013] and Chapter 3. Further, longer
demonstrations can be segmented automatically into key phases using, e.g.,
probabilistic approaches [Fox et al., 2009; Lee et al., 2015].
Not only configuration data can be mediated through demonstration. For
instance, image data, force/torque data, etc., at different stages of a task can
be mediated to the robot in the form of examples. Based on these examples,
model parameters can be determined automatically using machine learning
[Bishop, 2007; Murphy, 2012] or deep learning [Goodfellow et al., 2016].
Such models could potentially be used by the robot to autonomously monitor
the task.

1.3 Introduction to Machine Learning

Computers and robots have traditionally been programmed by writing ex-
plicit code, specifying sets of rules and behaviors in detail. In predictable

16
 1.3 Introduction to Machine Learning

scenarios this strategy works well, but most of the tasks that humans
accomplish in their everyday life are far too complex to mediate in such
fashion. Whereas distinguishing, for instance, whether a certain image rep-
resents a car or a bicycle is typically easy for humans, hand crafting the
classification rules from raw image data is not feasible. The introduction
in [Goodfellow et al., 2016] provides interesting examples of tasks that are
easy for humans to perform, but difficult to describe formally.
Machine learning is the most promising framework to address such prob-
lems. This framework consists of three major parts; supervised learning,
unsupervised learning, and reinforcement learning [Bishop, 2007; Murphy,
2012; Goodfellow et al., 2016], all of which are relevant in robotics. The
main idea is to provide computers with example data, and use these to
define models, instead of specifying the models manually. In this thesis, we
mainly consider supervised learning. Whereas traditional automatic control
is useful for achieving specific robot movements, machine learning can po-
tentially allow for programming on higher levels of abstraction. For instance,
if natural language or image data could be interpreted in a meaningful way,
these could potentially be used as a decision basis by a robot. In machine
learning, such data are typically analyzed using artificial neural networks
[Kröse and Smagt, 1993; Goodfellow et al., 2016], which were inspired by
the connected neurons in biological brains.
Iterative learning control is closely related to reinforcement learning,
and has been used for motion control [Norrlöf, 2002; Norrlöf and Gunnars-
son, 2002] and force control [Marchal et al., 2014] of industrial robots. The
main idea is to iteratively update the control signal based on previous trials,
to compensate for repetitive sources of error. System identification [Ljung,
1987; Johansson, 1993; Lindsten et al., 2013; Schön et al., 2015] is closely
related to machine learning, and is specialized on finding dynamical models
from time series of input and output signals.
Supervised machine learning is used to approximate a given function,
y( x), with a parameterized function, ŷ( xpθ ), which takes some input data x,
and maps it to an output ŷ. Here, θ denotes the model parameters. In the
example of image recognition, x could be pixel values, y would be the true
image category, and ŷ( x) could be interpreted as the probability distribution
over the two categories, i.e., car and bicycle, given x.
In order to learn ŷ( x), the model is exposed to a large data set of
examples, called training data. In supervised learning, the training data
consist of both input data and the corresponding known outputs, usually
manually labeled. In the training phase, the elements of θ are adjusted
to fit the training data by means of optimization. A loss function, L, in
which some measure of the error of ŷ( x) compared to y( x) is included, is
minimized with respect to the model parameters.
Since the training data can only include a small subset of all possible

17
Chapter 1. Introduction

data points, an important aspect of machine learning is generalization, i.e.,

to predict the output given input not used during training. In order to
achieve this, the complexity of the model is typically restricted, by keeping
the number of parameters low, or by penalizing the complexity by including
some measure of it in L. Further, test data, not directly used to optimize
the model parameters, are used to estimate how well the model generalizes.
It is, however, common to determine some model hyperparameters based
on the performance on test data. Therefore, it is good practice to use yet
another data set to investigate the generalizability, without affecting the
model in any way. Such data are called validation data.
This general approach is adopted in Chapter 7, where the aim is to
take a step toward more accessible robot programming. Ideally, a non-
expert operator should be able to provide a robot with data, enabling it
to learn from experience. Similar to the image-classification example, a
model is trained to determine a class given input data. In particular, the
input consists of robot joint torques, and the task is to determine whether
a certain force/torque transient, acting on the robot end-effector, is present
or not. One important difference from the image-recognition example is
that the input consists of a time series rather than a static representation,
which should be taken into account when choosing the structure of the
model. Related approaches have been presented in [Rodriguez et al., 2010;
Rojas et al., 2012; Stolt et al., 2015b].

1.4 Problem Formulation

This section describes the problems addressed on a conceptual level. More
technical problem formulations are given gradually throughout this the-
sis. In summary, the main problem addressed in this thesis is that robot
programming is expensive and time consuming, and that robots, once pro-
grammed, exhibit low adaptability to the environment.
The first aim of this thesis is to investigate whether it is possible to
automatically interpret demonstrated corrections of robotic movement. In
the scenarios addressed, the original movement has a faulty ending, for
instance as a result of new work-space configuration. It would be desir-
able to specify the correction through demonstration, and the result of the
correction should be sufficiently predictable for an operator.
A common problem is robots’ inability to handle unforeseen events.
Demonstrated movements can be saved and replayed exactly, but such func-
tionality is meaningful only in predictable environments. Unforeseen events
may force a robot to deviate from the intended movement, for instance to
avoid collision. It should therefore be investigated whether unforeseen de-
viations of robotic movement could be recovered from, so that the intended

18
 1.5 Contributions

task could still be accomplished. A motion controller to achieve this should

be formulated. Such a control system should have the necessary features
known from automatic control: asymptotic stability, safe stability margins,
and control signals of moderate magnitude. It should support both real co-
ordinate spaces such as position in joint or Cartesian space, and orientation
in Cartesian space which is more challenging due to the topology of the 3D
rotation group SO(3).
Robotic manipulation requires in each step a validation procedure, to
ensure that the intended subtask has been accomplished. Without such
validation, successive subtasks and eventually the entire manipulation task
would be jeopardized. This could leave the task unfinished, and even damage
the robot or the work objects. We therefore address the question of whether
validation models could be mediated to a robot through demonstration,
rather than explicit programming which is nowadays the most common
approach.
In ordinary lead-through programming, the operator demonstrates
movements by moving the robot arm through physical contact in such a way
that the robot arm will complete an intended task. This strategy suffices
to teach behavior in the position domain. Contact forces induced between
the robot and the work objects are also of importance, but more difficult to
demonstrate. Unfortunately, these contact forces can not be distinguished
from contact forces from the operator in ordinary lead-through program-
ming. Finally, we therefore address the question of whether the contacts
could be separated using two robot arms, where one is in contact with the
operator allowing for direct lead-through programming, and the other one
is controlled remotely to manipulate the work object.

1.5 Contributions
The main contributions of this thesis are:

• A control architecture for adjustment of robot movement through phys-

ical human–robot interaction;

• Improved replanning capabilities of robot movements;

• Sensorless detection of contact-force transients;

• Implementation of a dual-arm haptic interface for task demonstration.

Each contribution includes experimental validation on an industrial robot.

19
Chapter 1. Introduction

1.6 Publications
In this section, publications authored or co-authored by the thesis author are
presented. The published manuscripts have been subjected to peer review.
For each manuscript, A. Robertsson and R. Johansson have contributed with
supervision, including insights regarding previous work and our ongoing
research, as well as suggestions for improvement. First, the publications on
which this thesis is based are described.

Karlsson, M. (2017). On Motion Control and Machine Learning for Robotic

Assembly. Licentiate Thesis, TFRT–3274–SE, Dept. Automatic Control,
Lund University, Lund, Sweden.

Parts of the research presented in this PhD thesis have previously been
published in the Licentiate thesis above.

Karlsson, M., A. Robertsson, and R. Johansson (2017). “Autonomous in-

terpretation of demonstrations for modification of dynamical movement
primitives”. In: IEEE International Conference on Robotics and Automa-
tion (ICRA). May 29–June 3, Singapore, pp. 316–321.

In the publication above, M. Karlsson formulated the method for updat-

ing a partly faulty trajectory representation, based on a corrective demon-
stration. Further, M. Karlsson implemented the method and verified it
experimentally. Chapter 4 is based on this publication. By this we mean
that the chapter contains an updated version of the publication, where the
author has made changes and extensions as found suitable.

Karlsson, M., F. Bagge Carlson, A. Robertsson, and R. Johansson (2017).

“Two-degree-of-freedom control for trajectory tracking and perturbation
recovery during execution of dynamical movement primitives”. In: 20th
IFAC World Congress. July 9–14, Toulouse, France, pp. 1959–1966.

M. Karlsson and F. Bagge Carlson identified the necessity of augment-

ing the existing DMP framework, to make related research approaches on
DMP perturbation recovery practically realizable. M. Karlsson formulated
the augmentation, and verified it in simulations and experimentally, while
frequently discussing the work with F. Bagge Carlson. Further, F. Bagge
Carlson implemented the method presented as an open-source Julia pack-
age, which can be found online [Bagge Carlson, 2016]. Example code in
Matlab, written by M. Karlsson, can be found online [Karlsson, 2017c].
Chapter 5 is partly based on this publication.

20
 1.6 Publications

Karlsson, M., A. Robertsson, and R. Johansson (2018). “Convergence of

dynamical movement primitives with temporal coupling”. In: European
Control Conference (ECC). June 12–15, Limassol, Cyprus, pp. 32–39.

M. Karlsson formulated the convergence proof and verified it experimen-

tally. Chapter 5 is partly based on this publication.

Karlsson, M., A. Robertsson, and R. Johansson (2019). “Temporally cou-

pled dynamical movement primitives in Cartesian space”. Manuscript
prepared for submission to review for publication.

M. Karlsson formulated the control algorithm and verified it experimen-

tally. Chapter 6 is based on this manuscript.

Karlsson, M., A. Robertsson, and R. Johansson (2018). “Detection and con-

trol of contact force transients in robotic manipulation without a force
sensor”. In: IEEE International Conference on Robotics and Automation
(ICRA). May 21–25, Brisbane, Australia, pp. 21–25.

M. Karlsson formulated the transient-detection method and verified it

experimentally. Chapter 7 is based on this publication.

Ghazaei Ardakani, M. M., M. Karlsson, K. Nilsson, A. Robertsson, and R.

Johansson (2018). “Master-slave coordination using virtual constraints
for a redundant dual-arm haptic interface”. In: IEEE/RSJ Interna-
tional Conference on Intelligent Robots and Systems (IROS). October
1–5, Madrid, Spain, pp. 8751–8757.

The idea of the dual-arm haptic interface was conceived by K. Nilsson

and M. Ghazaei Ardakani. The latter developed and formulated the control
algorithm for the haptic interface. M. Karlsson implemented the interface
and evaluated it experimentally. Chapter 8 is based on this publication.
Andreas Stolt, Cognibotics AB, is gratefully acknowledged for identifying
the dynamics of the robot used in the experiments. Mathias Haage at Dept.
Computer Science, Lund University, is gratefully acknowledged for guidance
in compiling separate parts of the software implementation with different
compilation flags.

The following publications cover topics in robotics, nonlinear state esti-

mation, and positioning. They are, however, outside the scope of this thesis.

21
Chapter 1. Introduction

Bagge Carlson, F., M. Karlsson, A. Robertsson, and R. Johansson (2016).

“Particle filter framework for 6D seam tracking under large external
forces using 2D laser sensors”. In: IEEE/RSJ International Conference
on Intelligent Robots and Systems (IROS). October 9–14, Daejeon, South
Korea, pp. 3728–3734.
Haage, M., S. Profanter, I. Kessler, A. Perzylo, N. Somani, O. Sörnmo, M.
Karlsson, S. G. Robertz, K. Nilsson, L. Resch, and M. Marti (2016). “On
cognitive robot woodworking in SMErobotics”. In: ISR 2016: 47th Inter-
national Symposium on Robotics. VDE. June 21–22, Munich, Germany,
pp. 1–7.
Karlsson, F., M. Karlsson, B. Bernhardsson, F. Tufvesson, and M. Persson
(2015). “Sensor fused indoor positioning using dual band WiFi signal
measurements”. In: European Control Conference (ECC). July 15–17,
Linz, Austria, pp. 1669–1672.
Karlsson, M., F. Bagge Carlson, J. De Backer, M. Holmstrand, A. Roberts-
son, and R. Johansson (2016). “Robotic seam tracking for friction stir
welding under large contact forces”. In: 7th Swedish Production Sympo-
sium (SPS). October 25–27, Lund, Sweden.
Karlsson, M. and F. Karlsson (2016). “Cooperative indoor positioning by
exchange of bluetooth signals and state estimates between users”. In:
European Control Conference (ECC). June 29–July 1, Aalborg, Denmark,
pp. 1440–1444.
Persson, M., M. Karlsson, and F. Karlsson (2014). Method for improved in-
door positioning and crowd sourcing using PDR. International patent,
Publication number: WO 2015/118369 (A1), Applicant: Sony Corpora-
tion, Examined and granted under the patent cooperation treaty (PCT).
Sony. url: https://patents.google.com/patent/WO2015118369A1/en (visited on
2019-03-02).
Wadenbäck, M., M. Karlsson, A. Heyden, A. Robertsson, and R. Johansson
(2017). “Visual odometry from two point correspondences and initial au-
tomatic camera tilt calibration”. In: 12th International Joint Conference
on Computer Vision, Imaging and Computer Graphics Theory and Appli-
cations, Volume 6. VISIGRAPP. February 27–March 1, Porto, Portugal,
pp. 340–346.

1.7 Thesis Outline

This thesis is a monograph. In terms of scientific contributions, Chap-
ters 4 to 6 focus on movement adjustment using motion control, whereas
Chapters 7 and 8 focus on programming by demonstration based on ma-
chine learning and force control. Despite this division, both programming

22
 1.7 Thesis Outline

by demonstration, motion control, force control, and machine learning are

used throughout the thesis to put the contributions into context. Finally, a
conclusion is presented in Chapter 9. References in the digital version of
the thesis are clickable and marked in blue.

Chapter 2: Robots and Interfaces for Experimental Validation

The robots and research interfaces used in the experiments in this thesis
are presented.

Chapter 3: Dynamical Movement Primitives—An Overview

The fundamentals of DMPs, used to model and generate robot movement,
have been developed in previous research. In this chapter, the main results
from previous research regarding DMPs are presented. Further, the DMP
framework is compared with alternative movement models.

Chapter 4: Modification of Robot Movement Using

Lead-Through Programming
A framework for modification of DMPs by means of corrective demonstra-
tions is presented. It allows for an operator to adjust a faulty part of a
trajectory using lead-through programming, while retaining satisfactory
parts.

Chapter 5: Temporally Coupled Dynamical Movement Primitives

in Rn
The DMP framework is augmented to enable recovery from perturbations
during DMP execution. The control algorithm for temporal coupling pro-
posed in [Ijspeert et al., 2013] is used as a foundation, and evaluated in the
beginning of this chapter. It is first shown that it is necessary to lower the
feedback control gains in the previously proposed algorithm. It is further
shown that it is necessary to add a feedforward signal to the controller. A
new control algorithm for temporally coupled DMPs is proposed based on
these insights. It is shown mathematically that exponential convergence to
the goal state can be guaranteed. The configuration space is assumed to be
a real coordinate space, such as joint space or position in operational space.

Chapter 6: Temporally Coupled Dynamical Movement Primitives

in 3D Orientation Space
The temporally coupled DMPs are extended to support orientation in Carte-
sian space. Orientations are represented by unit quaternions, and the spe-
cial topology of the unit quaternion set is taken into account in the design
and stability analysis of the control system.

23
Chapter 1. Introduction

Chapter 7: Detection and Control of Contact-Force Transients

A machine-learning procedure for detecting force/torque transients acting
on a robot end-effector, by measuring robot joint torques, is developed. The
procedure does not require any external force/torque sensor, the optimiza-
tion time of the detection model is moderate, and the detection model can
generalize to other tasks not used for model training. These are the main
benefits as compared to the state of the art.

Chapter 8: A Dual-Arm Haptic Interface for Task Demonstration

An immersive haptic interface for task teaching is proposed. By retaining
a fixed pose offset between two robot end-effectors in task space, while en-
abling lead-through programming on both arms, any arm can be controlled
directly by an operator while the other one moves accordingly. Interaction
forces are reflected to each side as haptic feedback. Hence, the operator can
act and sense through the system. The interface supports mapping between
two not-necessarily-similar robot arms, mounted on any surfaces and with
any initial configurations. Further, the arms can be interacted with on any
point on their structures.

1.8 Videos
Videos that show each of the functionalities developed are available in the
entry for this thesis in the Lund University Research Portal:
http://portal.research.lu.se/portal/

These videos are also available on the author’s Youtube channel:

https://www.youtube.com/channel/UCIYU9qpJwkBbsFWscI2vV-g

Each video will be referred to and explained throughout this thesis.

24
2
Robots and Interfaces for
Experimental Validation

Most of the experiments presented in this thesis have been done in the
Robotics Lab shared by the departments of Automatic Control and Com-
puter Science at Lund University, and some have been done at ABB Cor-
porate Research in Västerås, Sweden. Different versions of the ABB YuMi
robot have been used. YuMi is a dual-arm collaborative robot [Kock et al.,
2011; ABB Robotics, 2018]. Each arm has seven joints and is therefore re-
dundant. It is designed for safe human–robot interaction [Matthias et al.,
2011], with low mass, weak motors, smooth surfaces, and speed limitations.
Each joint is controlled by the ABB IRC5 control system [ABB Robotics,
2019a], running at 2 kHz. The control system consists of a current-control
loop to achieve desired joint torque, and outer control loops to achieve de-
sired joint position and velocity; see, e.g., Fig. 2.4 in [Stolt, 2015] for a block
diagram of the IRC5 system. In this thesis, we will refer to the IRC5 sys-
tem as the internal robot controller. The algorithms developed in this thesis
were implemented on ordinary PCs, and communicated with the internal
robot controller through the research interfaces described in the following
sections.
YuMi is backdrivable, which means that it is possible to move it through
physical contact without being prevented by high friction, large mass, and
high gear ratio between the arm and motor sides. This is a common prop-
erty for light-weight robots, whereas larger traditional industrial robots are
typically not backdrivable. Sensorless lead-through programming is easier
to achieve for robots with such backdrivability. In [Stolt et al., 2015a], it was
implemented on the YuMi prototype described in Sec. 2.1. A version of that
implementation is included in the YuMi product (see Sec. 2.2) by default.

25
Chapter 2. Robots and Interfaces for Experimental Validation

Figure 2.1 The prototype version of YuMi [ABB Robotics, 2018] used in
the experiments. It was located in the Robotics Lab at Lund University.

2.1 YuMi Prototype Version

The prototype version of YuMi used in this thesis is shown in Fig. 2.1. The
research interface ExtCtrl [Blomdell et al., 2005; Blomdell et al., 2010],
running at 250 Hz, was used for communication between the PC and the
internal robot controller. Each wrist of the robot was equipped with a six-
degree-of-freedom ATI [ATI, 2019] Mini40 force/torque sensor. In this thesis,
these external sensors were used for validation only.

2.2 YuMi Product Version

The product version of YuMi mainly used in this thesis was located in
the Robotics Lab at Lund University, and is shown in Fig. 2.2. Some of
the experiments in Chapter 7 were done at ABB Corporate Research in
Västerås, on the robot shown in Fig. 2.3. This robot was also used for some
separate demonstrations in the SARAFun project. The main demonstration
platform used in SARAFun is shown in Fig. 2.4. The product version of YuMi
was controlled through a research-interface version [Bagge Carlson and
Haage, 2017] of Externally Guided Motion (EGM) [ABB Robotics, 2019c],
running at 250 Hz. The product version is not equipped with force/torque
sensors by default.

26
 2.2 YuMi Product Version

Figure 2.2 The product version of YuMi mainly used in the experiments.
It was located in the Robotics Lab at Lund University.

Figure 2.3 The product version of YuMi located at ABB Corporate Re-
search in Västerås, used for some of the experiments in Chapter 7 and for
some separate demonstrations in SARAFun.

27
Chapter 2. Robots and Interfaces for Experimental Validation

Figure 2.4 The product version of YuMi located at ABB Corporate Re-
search in Västerås, used as main demonstration platform in SARAFun.

28
3
Dynamical Movement
Primitives—An Overview

Since robot programs consist of sequences of movements, representation and

execution of these movements is an important area in robotics. In indus-
trial settings, it is necessary that the movements can be realized according
to plan. To enhance robots’ replanning capabilities, dynamical movement
primitives (DMPs) have been developed. These are used to represent and
execute robot movement, and have been developed with programming by
demonstration in mind. We consider the work in [Ijspeert et al., 2013] as
the main source from previous research, because it presents a complete
description of the framework together with some possible extensions and
applications. This chapter presents an overview of DMPs, but no extension
of the framework as compared to [Ijspeert et al., 2013]. Earlier versions
have been introduced in [Schaal et al., 2000; Ijspeert et al., 2002; Schaal
et al., 2003; Ijspeert et al., 2003]. The framework has been inspired by the
biological movement models presented in [Giszter et al., 1993; Mussa-Ivaldi,
1999]. Table 3.1 lists some of the notation used in this chapter.
Although there are other alternatives, the most ubiquitous model for
DMPs is based on the following dynamical system.

τ 2 ÿr = α(β (ˆ − y) − τ ẏ) + f ( x) (3.1)

Here, y denotes robot configuration, α and β are positive constants, and

f ( x) is a so-called forcing term where x is a phase variable. There are
three aspects that support movement modification: The speed of the system
can be set through the positive time parameter τ , the goal configuration
can be set through ˆ , and movement can be generated from any state.
Because of f ( x), the system is nonlinear. It is commonly pointed out that
(3.1) represents a damped-spring system, where f ( x) can be seen as an
external force that directly affects the acceleration [Ijspeert et al., 2013]. In
this thesis, we will view DMPs more from an automatic-control perspective,

29
Chapter 3. Dynamical Movement Primitives—An Overview

Table 3.1 Notation used in this chapter.

Notation Description
+
n ∈Z Dimension of robot configuration
y ∈ Rn Robot configuration
ÿr ∈ Rn Robot acceleration reference
τ ∈ R+ Time parameter
ˆ ∈ Rn Goal state
x ∈ R+ Phase variable
f ( x) ∈ Rn Learnable virtual forcing term
α, β , α x ∈ R+ Positive constants
Nb ∈ Z+ Number of basis functions
Ψ j ( x) ∈ Rn The j:th basis function vector
wj ∈ Rn The j:th weight vector

while retaining the notation from [Ijspeert et al., 2013] where applicable.
The system consists of a proportional-derivative (PD) controller, together
with feedforward by f ( x). The dynamics are achieved by sending ÿr as a
reference acceleration to internal motion-control loops. It is assumed that
the reference acceleration can be realized during free-space motion, so that
ÿ = ÿr . This is reasonable for an acceleration of moderate magnitude and
time derivative. Depending on the level of control, ÿr can also be seen as a
control signal. The control structure is visualized in Fig. 3.1.
Consider for now the simplified case where f ( x) = 0, so that

τ 2 ÿr = α(β (ˆ − y) − τ ẏ) (3.2)

This is a PD controller with ˆ as reference configuration. In the context of

robot position control, overshoot is undesirable due to the risk of collision.

ˆ ÿr y
DMP Robot

-1

Figure 3.1 Schematic overview of the DMP control structure described by

(3.1). The block denoted ’Robot’ includes the internal controller of the robot.
The block denoted ’DMP’ can be seen as a controller, that generates ÿr as
control signal. It can also be seen as a reference generator, or planner, that
generates ÿr as reference for the internal robot controller.

30
 Chapter 3. Dynamical Movement Primitives—An Overview

This is especially important in operations that involve contact with the

environment. This fact motivates the absence of an integral part, which is
otherwise common in control applications [Åström and Hägglund, 1995] but
may cause overshoot. Further, it is advisable to choose the feedback control
parameters α and β so that the system is critically damped, i.e., so that y
converges asymptotically to ˆ without overshoot. In [Ijspeert et al., 2013], it
was stated that this is achieved for β = α/4. In the following, we will verify
this by analyzing the poles of the system. This can be done separately for
each dimension of the configuration, and to keep the notation uncluttered
we consider a one-dimensional case. Note that the dynamical system (3.2) is
linear. The transfer function G ˆ→ y (s) for the closed-loop system in Fig. 3.1
with f ( x) = 0 is given by
αβ
Y (s) τ2
= G ˆ→ y (s) = (3.3)
G(s) α αβ
s2 + s + 2
τ τ
Here, Y (s) represents the Laplace transform of y, and G(s) is the Laplace
transform of ˆ . Inserting β = α/4 yields
α2
Y (s) 2
= G ˆ→ y (s) = 1 4τ 22 (3.4)
G(s) α
s+
2τ
Both poles are located in −α/(2τ ). Because these are strictly negative,
global asymptotic stability can be concluded. Because they are real, no
overshoot will occur. We also see the effect of τ . A larger value yields poles
closer to the origin, and hence slower system dynamics.
The controller in (3.2) hence drives the configuration to the desired state
ˆ , and the speed of the system can be adjusted. While this is satisfactory
for a wide range of control applications, it is in robot motion control very
important which trajectory is traversed to reach a desired end point. For
instance, reach limits, joint angle limits, and external obstacles must be
avoided. The purpose of f ( x) in (3.1) is to enable specification of the trajec-
tory traversed. Each element of f ( x), denoted fi ( x), is given by a weighted
sum of radial basis functions
P Nb
j =1 Ψi, j ( x)wi, j
fi ( x) = P Nb x · (ˆ i − y0,i ) (3.5)
j =1 Ψi, j ( x)

where each basis function, Ψi, j ( x), is determined as

3 4
1
Ψi, j ( x) = exp − 2 ( x − c i, j )2 (3.6)
2σi, j

31
Chapter 3. Dynamical Movement Primitives—An Overview

Here, Nb denotes the number of basis functions, w denotes weights, y0 is

the initial configuration, and c and σ are the center and width of each basis
function, respectively. Further, the phase variable evolves according to
τ ẋ = −α x x (3.7)
where α x is a positive constant. It can be seen that x converges exponentially
to 0. Because x is a factor of f ( x), and the remaining factors are bounded,
f ( x) has significant magnitude only in a finite time window. Because f ( x)
in practice vanishes after some time, convergence to ˆ can be concluded
intuitively based on the stability of (3.2). Global exponential stability for
DMPs has been shown in [Perk and Slotine, 2006; Wensing and Slotine,
2017], based on contraction theory [Lohmiller and Slotine, 1998].
Next, we will see how f ( x) can be determined based on a demonstrated
movement. Again, this can be done separately for each dimension of the
configuration, and to keep the notation uncluttered we consider a one-
dimensional case. Given a discrete-time trajectory, ydemo , a corresponding
DMP can be determined. Then, ˆ is given by the end position of ydemo ,
whereas τ can be set to get a desired time scale. Using (3.1), the corre-
sponding virtual force can be determined as
 
ftarget
1

 ftarget 
 2 
ftarget = τ ÿdemo − α(β (ˆ − ydemo ) − τ ẏdemo ) = 
2
 .. 
 (3.8)
 . 
N
ftarget
s

where Ns is the number of samples. The weights can be found by locally

weighted linear regression [Atkeson et al., 1997; Schaal and Atkeson, 1998]
with the solution
vT Γj ftarget
wj = (3.9)
vT Γj v
where
x1 (ˆ − ydemo
1
)
 
 x2 (ˆ − y1 ) 
demo 
v=

..  (3.10a)
 . 
x Ns (ˆ − ydemo
1
)
N
Γj = diag( Ψ 1j , Ψ 2j · · · Ψ j s ) (3.10b)

3.1 Related Research

Thanks to its support for movement adjustment, the DMP framework has
been widely used and developed within robot research, though it has yet to

32
 3.2 Parameter Choices

be established in industry. It was used for robot learning from unstructured

demonstrations in [Niekum et al., 2015], where pick-and-place tasks and as-
sembly tasks were considered. Demonstrated movements were generalized
to new but static goals. It was used to reach moving goals in the context of
object handover between human and robot in [Prada et al., 2014]. In [Kulvi-
cius et al., 2012], the DMP framework was modified and a method for joining
movements for tasks such as handwriting was developed. Trajectory-based
reinforcement learning has been used to tune DMP parameters automati-
cally in, e.g., [Kober et al., 2008; Kroemer et al., 2010; Stulp and Schaal,
2011; Pastor et al., 2013]. In the context of robotic grasping, reinforcement
learning has been used to learn goal parameters in [Stulp et al., 2011] and
goals and weights simultaneously in [Stulp et al., 2012b]. Impedance was
learned from experience in [Stulp et al., 2012a]. Online movement adapta-
tion to reference force profiles was considered in [Abu-Dakka et al., 2015].
In [Chen et al., 2016], the framework for unsupervised learning of state-
space models in [Karl et al., 2016] was applied to DMPs, to generate DMP
representations in low-dimensional state spaces. It has also been explored
how to use multiple demonstrations to refine a given movement [Matsub-
ara et al., 2011], and how to segment movements using a library of DMPs
[Meier et al., 2011].
These are just some examples where DMPs have been applied and ex-
tended. More results from previous research will be described throughout
this thesis.

3.2 Parameter Choices

It has already been recommended to let β = α/4. There is one degree of
freedom left to be chosen, and it corresponds to the location of the double
pole −α/(2τ ). A larger value of α corresponds to a double pole further
away from the origin, and hence a faster control system. A larger value
of τ has the opposite effect. Given a demonstrated trajectory, τ should be
chosen before determining the weights w j . If the same value of τ would
be used later during DMP execution, the movement would be as fast as
during the demonstration. One reasonable choice is to let τ = α x Tdemo /3
while determining w j , where Tdemo is the duration of the demonstrated
movement. This corresponds to 95 % convergence of x in (3.7) at time Tdemo .
The number of basis functions Nb corresponds to movement resolution.
In the implementations in this thesis, Nb = 50 was chosen. It is reasonable
to choose c i, j so that the basis functions have equal spacing in time, which
corresponds to exponential spacing in x [Ijspeert et al., 2013]. Further, it
is recommended to choose σi, j so that at each time of the movement, some
basis function is activated.

33
Chapter 3. Dynamical Movement Primitives—An Overview

The constant α x occurs in previous publications, but it does not affect

the functionality as long as it is strictly positive, since it must be compen-
sated for when determining DMP parameters based on desired movements.
Therefore, α x was set to 1 in the simulations and experiments in this thesis.
Similarly, the initial value of x could be chosen freely and was set to 1.

3.3 Comparison with Alternative Movement

Representations
A straightforward but naive approach would be to represent robot move-
ments as time series of reference positions, to be sent to the internal robot
controller during execution. The time scale of such a representation would
be easy to adjust. However, it would be unclear how to adjust goal configu-
ration and how to replan online, for instance to avoid obstacles.
Potential fields and splines are often brought up as two alternatives
to DMPs for movement representation. Similar to DMPs, potential fields
represent attractor landscapes, with convergence to a goal position and
without explicit time dependence. Potential fields have been considered for
robot control by many researchers, see, e.g., [Khatib, 1986; Koditschek, 1987;
Li and Horowitz, 1999]. Stability aspects of potential fields for force and
motion control have been addressed in [Johansson et al., 2009; Johansson
and Robertsson, 2009]. Design of potential fields for some obstacle avoidance
scenarios has been done in, e.g., [Koditschek, 1987]. Vector fields typically
define the movement based on given positions, but determining the vector
field given a desired behavior is not straightforward. Further, while DMPs
allow for different control signals from the same position, this can not be
achieved with conventional potential fields.
For imitation learning, splines have been widely used. Splines are func-
tions that are defined piecewise by polynomials, and retain smoothness
where the polynomials connect. It has been shown in, e.g., [Miyamoto et al.,
1996; Wada and Kawato, 2004] that demonstrated trajectories can be rep-
resented and successfully reproduced by means of splines. However, online
replanning is not supported, and temporal and spatial scaling can be done
only by recomputation of the spline polynomials.

34
4
Modification of Robot
Movement Using
Lead-Through Programming

4.1 Introduction
Similar to most software products, robot programs need to be updated from
time to time. A change of product to be manufactured, a desire to shorten
cycle time, previous programming mistakes, or a changed work-cell config-
uration are typical reasons for this. It might also be desirable to migrate
functionality between different robot stations, or to be able to download
preliminary programs or general templates put together elsewhere and
subsequently make final adjustments for the intended robot cell. To make
it worthwhile to reuse existing but imperfect robot programs rather than
programming from the beginning, easy means of modification are necessary.
This chapter presents a framework that allows for a robot operator to
adjust robot movements in an intuitive way. Given a faulty trajectory, the
operator can use lead-through programming [Stolt et al., 2015a] to demon-
strate a corrective one. Based on this, a modified trajectory representation
is determined automatically. DMPs, described in [Ijspeert et al., 2013] and
Chapter 3, are used to represent and generate the movement, though this
presented framework does not require that the faulty trajectory is generated
by a DMP.
In the process of program updating, it is natural that previously de-
fined movements need modification. In the case where a starting point has
been updated without further consideration, which in principle is supported
by the DMP framework [Ijspeert et al., 2013], a DMP-generated trajectory
would still converge to the goal configuration, given that no physical limi-
tations prevent this, but it would necessarily go through a different path.
The path modification would not be specified by the operator by default,

35
Chapter 4. Modification of Robot Movement. . .

and it would be larger for a larger difference between the original and the
updated starting points. Similarly, if the complete trajectory is of interest
it is not enough to modify the goal state. One way to solve the problem
would be to record an entirely new trajectory, and then construct a cor-
responding DMP. However, this would be unnecessarily time consuming
for the operator, if only some part of the trajectory had to be modified.
In this chapter, we consider the unfavorable case where the last part of a
robot trajectory is unsatisfactory. This is a realistic scenario, because many
manipulation movements start by approaching a work object with large
position tolerances, and end with physical contact and small position tol-
erances. Section 4.2 presents two such scenarios, in the context of robotic
assembly.
A standard approach has been to use demonstrated configuration tra-
jectories to determine preliminary DMPs. However, simply repeating these
trajectories does not guarantee success. Further refinements and ability to
adapt based on sensor data are typically necessary. Adaptation based on
force measurements to follow force profiles in robotic assembly was consid-
ered in [Abu-Dakka et al., 2015]. Further, initial demonstrations have been
followed by trajectory-based reinforcement learning for many different tasks.
In [Pastor et al., 2013], a simulated legged robot was considered. A DMP
that enabled the robot to make a small jump was first manually defined.
After applying reinforcement learning, the robot could make a longer jump
over a gap. The same principle was applied in [Kober et al., 2008], where
a simulated robot arm was taught to catch a ball in a cup. In [Kroemer
et al., 2010], reinforcement learning and imitation learning were combined
to achieve object grasping in an unstructured environment. Compared to
such refinements, the modification presented here is less time consuming
and does not require engineering work. On the other hand, the previous
work offers modulation based on sensor data and finer movement adjust-
ment. Therefore, the framework presented in this chapter can be used as an
intermediate step where, if necessary, a DMP can be modified to prepare
for further improvement, see Fig. 4.1. This modification can be used within
a wide range of tasks. In this chapter, we exemplify with peg-in-hole tasks.
In [Pastor et al., 2009; Ijspeert et al., 2013; Pastor et al., 2013], avoid-
ance of spherical obstacles was incorporated in the DMP framework by
modeling repulsive forces from the obstacles. In [Stavridis et al., 2017], this
research was extended for avoidance of obstacles contained in ellipsoids.
These approaches have been verified for several scenarios, but require an
infrastructure for obstacle detection, as well as some coupling-term pa-
rameters to be defined. In practice they preserve convergence to the goal
configuration in uncomplicated settings, but since the path to get there is
modified by the obstacle avoidance, it is not guaranteed to follow any spe-
cific trajectory to the goal. This is often significant, for instance in assembly

36
 4.2 Motivating Examples

Further
improvement

Successful
Initial
Create DMP Evaluation
demonstration
Unsuccessful

Modify Corrective
DMP demonstration

Figure 4.1 Schematic visualization of the work flow, from an operator’s

perspective. A DMP was created based on a demonstration. Subsequently,
the DMP was executed while evaluated by the operator. If unsuccessful,
the operator demonstrated a correction, which yielded a modified DMP to
be evaluated. Once successful, further improvement could be done by, e.g.,
trajectory-based reinforcement learning, though that was outside the scope of
this work. Steps that required direct, continuous interaction by the operator
are marked with red color. Steps that required some attention, such as
supervision and initialization, are marked with blue. The operations in the
white boxes were done by the software in negligible computation time, and
required no human involvement. The research in this chapter focuses on the
steps within the dashed rectangle.

tasks. Further, for more complex environments that are realistic in robotic
manipulation, modeling repulsive forces from obstacles would imply a risk
of getting stuck, thereby jeopardizing task completion. Instead, the method
described here allows for the operator to lead the manipulator backwards,
approximately along the part of the deficient path that should be adjusted,
followed by a desired trajectory, as visualized in Fig. 4.2.

4.2 Motivating Examples

We here describe two scenarios where the framework would be useful. These
also form the experimental setups considered in Secs. 4.5 and 4.6, where
more details are given.

Setup 4.1 Consider the setup shown in Fig. 4.3, where a button should
be placed into a yellow case. A DMP was run for this purpose, but due to
any of the reasons described above the movement was not precise enough,
and the robot got stuck on its way to the target. Hitherto, such a severe
shortcoming would have motivated the operator to teach a completely new

37
Chapter 4. Modification of Robot Movement. . .

150 Deficient trajectory

Lead-through transportation
Corrective trajectory
Modified trajectory
y3 [mm]
100
Projection on horizontal plane

50

100

50 100
y2 [mm] 00
y1 [mm]
Figure 4.2 Trajectories of the robot’s end-effector from one of the experi-
ments. The arrow indicates the motion direction. A deficient trajectory was
generated from an original DMP. After that, the operator demonstrated a
corrective trajectory. Merging of these resulted in a modified trajectory. The
projection on the horizontal plane is only to facilitate the visualization.

DMP, and erase the old one. With the method proposed in this chapter, the
operator had the opportunity to approve the first part of the trajectory, and
only modify the last part. This was done by leading the robot arm backwards,
approximately along the faulty path, until it reached the acceptable part.
Then, the operator continued to lead the arm along the desired path to the
goal. When this was done, the acceptable part of the first trajectory was
merged with the last part of the corrective trajectory. After that, a DMP
was fitted to the resulting trajectory. Compared to just updating the target
point, this approach also enabled the operator to determine the trajectory
leading there.

Setup 4.2 For the setup in Fig. 4.4, there existed a DMP for moving the
robot arm from the right, above the button that was already inserted, to a
position just above the hole in the leftmost yellow case. However, under the
evaluation the operator realized that there would have been a collision if a
button were already placed in the case in the middle. A likely reason for
this to happen would be that the DMP was created in a slightly different
scene, where the potential obstacle was not taken into account. Further, the
operator desired to extend the movement to complete the peg-in-hole task,
rather than stopping above the hole. With the method described herein, the
action of the operator would be similar to that in Setup 4.1, again saving
work compared to previous methods.

38
 4.3 Problem Formulation

(a) (b)

(c) (d)
Figure 4.3 Visualization of Setup 4.1. The evaluation started in (a), and
in (b) the robot failed to place the button in the hole because of inadequate
accuracy. Between (a) and (b), the deficient trajectory was recorded. The
operator led the robot arm backwards (c), approximately along a proportion of
the deficient trajectory, and subsequently led it to place the button properly,
while the corrective trajectory was recorded. The robot then made the entire
motion, starting in a configuration similar to that in (a), and ending as
displayed in (d). Results from this setup are shown in Figs. 4.7 and 4.8.

4.3 Problem Formulation

In this chapter, we address the question of whether it is possible to auto-
matically interpret a correction, made by an operator, of the last part of a
DMP trajectory, while still taking advantage of the first part. The human–
robot interaction must be intuitive, and the result of a correction predictable
enough for its purpose. The correction should result in a new DMP, of which
the first part behaves qualitatively as the first part of the original DMP,
whereas the last part resembles the corrective trajectory. Any discontinuity
between the original and corrective trajectories must be mitigated.

39
Chapter 4. Modification of Robot Movement. . .

(a) (b)

(c) (d)

(e) (f)

Figure 4.4 Visualization of Setup 4.2. The initial goal was to move the
button to the leftmost yellow case, above the hole, to prepare for placement.
The evaluation started in (a), and in (b) the trajectory was satisfactory as the
placed button was avoided. In (c), however, there would have been a collision
if there were a button placed in the middle case. Further, it was desired to
complete the peg-in-hole task, rather than stopping above the hole. Hence,
the evaluated trajectory was considered deficient. In (d), the operator led
the robot arm back, and then in a motion above the potential obstacle, and
into the hole, forming the corrective trajectory. Based on the modified DMP,
the robot started in a position similar to that in (a), avoided the potential
obstacle in (e) and reached the new target in (f). The trajectories from one
attempt are shown in Fig. 4.9.

40
 4.4 Method

Table 4.1 Notation used in this chapter.

Notation Description
+
n ∈Z Dimension of robot configuration
y ∈ Rn Robot configuration
yd ∈ Rn Deficient trajectory
yc ∈ Rn Corrective trajectory
ydr ∈ Rn Retained part of deficient trajectory
ym ∈ Rn Modified version of ydr
yk ∈ Rn Robot configuration at time sample k
τ ∈ R+ Time parameter
ˆ ∈ Rn Goal state
Nb ∈ Z+ Number of basis functions
wj ∈ Rn The j:th weight vector

4.4 Method
In this section a method to determine which part of a deficient trajectory to
retain is presented, followed by a description of how it should be merged with
a corrective trajectory to avoid discontinuities. Finally, some implementation
aspects are addressed. Figure 4.1 displays a schematic overview of the work
flow of the application, from the user’s perspective. Table 4.1 lists some of
the notation used in this chapter.

Interpretation of Corrective Demonstration

Consider the case where a deficient trajectory, denoted yd , has been gen-
erated. The operator would then use lead-through [Stolt et al., 2015a] to
move the robot arm backwards, approximately along the deficient path, un-
til it reached an acceptable part. We refer to this movement as lead-through
transportation, as opposed to lead-through programming, because the pur-
pose of this movement is not to demonstrate directly but rather to prepare
for demonstration. Hence, it is not necessary to log the resulting trajectory,
though it was done in the experiments for the purpose of visualization.
Thereafter, lead-through programming was used to demonstrate a desired
path to the goal. This movement is referred to as the corrective trajec-
tory, denoted yc . Examples of these trajectories are shown in Figs. 4.2, 4.5
and 4.6. A trajectory formed by simply appending yc to yd would likely take
an unnecessary detour. Thus, only the first part of yd was retained. This is
illustrated in Fig. 4.6. The retained part, denoted ydr , was determined as
the part previous to the sample of yd that was closest to the initial point of
yc , denoted yc1 , i.e.,

41
Chapter 4. Modification of Robot Movement. . .

50
Deficient trajectory
Lead-through transportation
40
Corrective trajectory

30
y3 [mm]

d∗
20

10

0
100 105 110 115 120 125 130 135 140
y1 [mm]

Figure 4.5 Visualization of shortest distance, here denoted d ∗ , used to

determine the left separation marker in Fig. 4.6. The trajectories are the
same as in Figs. 4.2 and 4.6, except that the modified trajectory is omitted
here.

m
ydr = ydm , ∀m ∈ {1, 2 . . . M } (4.1)

where

M = argmin d ( ydk , yc1 ) (4.2)

k=1,2... K

Here, d (·, ·) denotes distance between two configurations, and K is the

number of samples in yd ; see Fig. 4.5 for an illustration. The approach of
using the shortest distance as a criterion, was motivated by the assumption
that the operator led the robot arm back, approximately along the deficient
trajectory, until the part that was satisfactory. At this point, the operator
started the corrective demonstration, thus defining yc1 (see right marker in
Fig. 4.6). By removing parts of the demonstrated trajectories, a significant
discontinuity between the remaining parts was introduced. In order to
counteract this, ydr was modified into ym , of which the following three
features were desired:
1. ym should follow ydr approximately;
2. The curvature of ym should be moderate;
3. ym should end where yc began, with the same movement direction in
this point.

42
 4.4 Method

Deficient trajectory
50
Lead-through transportation
Corrective trajectory
40 Modified trajectory
Separation point set by user
Separation point set automatically
y3 [mm]

30

20

10

0
100 105 110 115 120 125 130 135 140
y1 [mm]

Figure 4.6 Same trajectories as in Fig. 4.2, but zoomed in on the corrective
trajectory. Arrows indicate directions. The parts of the trajectories between
the separation markers were not retained. The right, blue, separation point
was determined explicitly by the operator during the corrective demonstra-
tion. The left, green, separation point was determined according to (4.2).
Further, what was left of the deficient trajectory was modified for a smooth
transition. However, the part of the corrective trajectory retained was not
modified, since it was desired to closely follow this part of the demonstration.
Note that the trajectories retained were not intended for direct play-back ex-
ecution. Instead, they were used to form a modified DMP, which in turn
generated a resulting trajectory, as shown in Figs. 4.7 to 4.9.

To find a suitable trade-off between these objectives, the following convex

optimization problem was formulated and subsequently solved, for each
dimension of the configuration:

minimize q ydr − ym q2 + λqT(∆2 ) ym q2 (4.3a)

ym
M
subject to ym = yc1 (4.3b)
M M −1
ym − ym = yc2 − yc1 (4.3c)

Here, T(∆2 ) is a second-order finite-difference operator, and λ denotes a con-

stant scalar. A larger value of λ corresponds to higher priority of Feature 2
as compared to Feature 1. Thereafter, yc was appended to ym , and one cor-
responding DMP was created with the method described in [Ijspeert et al.,

43
Chapter 4. Modification of Robot Movement. . .

2013] and Chapter 3. The next step in the work flow was to evaluate the
resulting DMP, as shown in Fig. 4.1.

Software Implementation
Most of the programming was done in C++, where DMPs were stored as
objects of a class. Among the data members of this class were the parame-
ters τ , ˆ , and w1... Nb , as well as some description of the context of the DMP
and when it had been created. It contained member functions for display-
ing the parameters, and for modifying ˆ and τ . The C++ linear algebra
library Armadillo [Sanderson and Curtin, 2016] was used in a major part
of the implementation. Further, the code generator CVXGEN [Mattingley
and Boyd, 2012] was used to generate C code for solving the optimization
problem in (4.3). By default, the solver code was optimized with respect to
computation time. This resulted in a real-time application, in which the
computation times were negligible in teaching scenarios. The optimization
problem was typically solved well below one millisecond on an ordinary PC.
The implementation developed in [Stolt et al., 2015a] was used to enable
lead-through programming. The operator could switch between different
high-level modes such as DMP execution, lead-through transportation, and
corrective movement by pressing buttons in a terminal user interface. Dur-
ing lead-through programming, trajectory logging started when a velocity
significantly different from zero was achieved.

4.5 Experiments
The robot used in the experimental setup was a prototype of the dual-arm
ABB YuMi [ABB Robotics, 2018], described in Chapter 2. The computations
took place in joint space, and the robot’s forward kinematics were used for
visualization in Cartesian space in the figures presented in this chapter.
The setups in Sec. 4.2 were used to evaluate the proposed method. For each
trial, the following steps were taken:

• An initial trajectory was taught, deliberately failing to meet the re-

quirements, as explained in Sec. 4.2;

• Based on this, a DMP was created;

• The DMP was used to generate a trajectory similar to the initial one.
This formed the deficient trajectory;

• The arm was retracted by means of lead-through transportation;

• A corrective trajectory was recorded;

44
 4.6 Results

200

150
y3 [mm]

100
Deficient trajectory
Lead-through transportation
Corrective trajectory
50 Resulting trajectory

50 100 150 200 250

y1 [mm]

Figure 4.7 Trajectories from the experimental evaluation of Setup 4.1. The
deficient trajectory went past the goal in the negative y1 -direction, preventing
the robot from lowering the button into the hole. After correction, the robot
was able to reach the target as the modified DMP generated the resulting
trajectory.

• Based on the correction, a resulting DMP was formed automatically;

• The resulting DMP was executed for experimental evaluation.

First, Setup 4.1 was set up for evaluation, see Fig. 4.3. The scenario
started with execution of a deficient trajectory. For each attempt, a new
deficient trajectory was created and modified. A total of 50 attempts were
made.
Similarly, Setup 4.2 (see Fig. 4.4) was set up, and again, a total of 50
attempts were made.
A video is available as a publication attachment to [Karlsson et al.,
2017b], to facilitate understanding of the experimental setup and results.
A version with higher resolution is available on [Karlsson, 2017d].

4.6 Results
For each attempt of Setup 4.1, the robot was able to place the button properly
in the yellow case after the modification. Results from two of these attempts
are shown in Figs. 4.7 and 4.8. In the first case, the deficient trajectory went
past the goal, whereas in the second case, it did not reach far enough.
Each of the attempts of Setup 4.2 was also successful. After modifica-
tion, the DMPs generated trajectories that moved the grasped stop button

45
Chapter 4. Modification of Robot Movement. . .

200

150
y3 [mm]

100
Deficient trajectory
Lead-through transportation
Corrective trajectory
50 Resulting trajectory

50 100 150 200 250

y1 [mm]

Figure 4.8 Similar to Fig. 4.7, except that in this case, the deficient tra-
jectory did not reach far enough in the negative y1 -direction.

above the height of potential obstacles, in this case other stop buttons, and
subsequently inserted it into the case. The result from one attempt is shown
in Fig. 4.9.

4.7 Discussion
Future work includes integration of the presented framework with rein-
forcement learning [Kober et al., 2008; Kroemer et al., 2010; Pastor et al.,
2013], in order to optimize the motion locally with respect to criteria such as
execution time. The program should also be augmented to take the purpose
of, and relation between, different DMPs into consideration. This extension
will emphasize the necessity of keeping track of different states within the
work flow. To this purpose, a state machine implemented in, e.g., JGrafchart
[Theorin, 2014], or the framework of behavior trees, applied on an aerial
vehicle in [Ögren, 2012] and robot control in [Marzinotto et al., 2014], would
be suitable.
Similar to movement adjustments based on sensor data, adjustments
by the operator during DMP execution could also be considered. Such a
step was taken in [Talignani Landi et al., 2018a], where [Karlsson et al.,
2017b] was extended significantly to enable adjustment by physical contact
during movement. The software in Sec. 4.4, first implemented in [Karlsson
et al., 2017b], was also used as a starting point for the implementation in
[Talignani Landi et al., 2018a]. A presentation of [Talignani Landi et al.,
2018a] in video format, including a demonstration of the functionality, is

46
 4.7 Discussion

120

y3 [mm] 100

80

60 Deficient trajectory
Lead-through transportation
Corrective trajectory
40 Resulting trajectory

−200 −100 0 100 200 300 400

y1 [mm]

Figure 4.9 Trajectories from experimental evaluation of Setup 4.2. For the
deficient trajectory, the end-effector was lowered too early, causing a potential
collision. After the correction, the robot was able to reach the target while
avoiding the obstacles. The movement was also extended to perform the
entire peg-in-hole task, rather than stopping above the hole.

available in [Talignani Landi et al., 2018b].

When robotic movement is adjusted in real time, whether it is due to
exploration in reinforcement learning, physical interaction with a human,
or something else, it is problematic that the phase variable x evolves in-
dependently of this for the ordinary DMP formulation, see [Ijspeert et al.,
2013]. This issue will be addressed in Chapter 5.
Performing the computations in joint space instead of Cartesian space
allowed for the operator to determine the entire configuration of the seven-
degree-of-freedom robot arm, rather than the pose of the tool only. However,
one could think of situations where the operator is not concerned about the
configuration, and the pose of the tool would be more intuitive to consider. It
would therefore be valuable if it could be determined whether the operator
aimed to adjust the configuration or just the pose of the tool. For example,
a large configuration change yielding a small movement of the tool, should
promote the hypothesis that the operator aimed to adjust the configuration.
It should be stated that the scenarios evaluated here are not covering the
whole range of plausible scenarios related to this method, and it remains as
future work to investigate the generalizability, and user experience, more
thoroughly. The last part of the resulting movement is guaranteed to follow
the retained part of the corrective demonstration accurately, given enough
DMP basis functions. Hence, the only source of error on that part is a faulty

47
Chapter 4. Modification of Robot Movement. . .

demonstration. For instance, the movement might require higher accuracy

than what is possible to demonstrate using lead-through programming.
Another limitation with this method is that it is difficult for the operator
to very accurately determine which part of the faulty trajectory to retain,
since this is done autonomously. However, for the experiments performed
here, the estimation by the operator was sufficient to demonstrate the
desired behavior. The benefit with this approach is that it saves time as the
operator does not have to specify all details explicitly.

4.8 Conclusion
In this chapter, an approach for modification of DMPs, using lead-through
programming, was presented. It allowed for a robot operator to modify faulty
trajectories, instead of demonstrating new ones from the beginning. Based
on corrective demonstrations, modified DMPs were formed automatically. A
real-time application, that did not require any additional engineering work
by the user, was developed and verified experimentally. A video showing the
functionality is available online [Karlsson, 2017d].

48
5
Temporally Coupled
Dynamical Movement
Primitives in Rn

5.1 Introduction
In this chapter, we continue to extend the DMP framework to support move-
ment adjustment. Whereas Chapter 4 focused on corrective demonstrations
after faulty movements, this chapter aims to allow for replanning during
movement.
Consider the case where a robot has to deviate from its intended move-
ment. One could think of several motivations for such deviation, for instance
avoiding an obstacle, avoiding blocking camera vision, waiting for a work
object to be handed over from another robot, etc. In this thesis, we use
physical contact with a human to exemplify. A human grasps or pushes the
robot, thereby stopping its movement or pushing it away from its path.
Typical industrial robots would react to this in two possible ways. A large
robot could possibly continue its movement despite the physical contact,
because of its large mass and strength compared to the human. A light-
weight collaborative robot, which is mainly considered in this thesis, has
weaker motors and its movement would be significantly affected by the
human. The motors would not be counteracted by the human for long,
however, because the robot would typically stop if reference states were too
far from measured states, or if joint torques were unexpectedly high. The
robot motors would then deactivate, the brakes would activate, and the
robot would possibly provide the operator with an error message. This is
a safety feature, sometimes called motion supervision, with the purpose of
avoiding damage of the robot and its surroundings. Motion supervision runs
on larger robots as well, though these are less sensitive to external forces.
Triggering such an error in an industrial setting implicates production stop

49
Chapter 5. Temporally Coupled Dynamical Movement Primitives in Rn

and significant engineering work to reset and adjust hardware and software.
Hence, this strategy is not feasible in less predictable environments.
Already the original DMP formulation implicates some replanning ca-
pabilities, see [Ijspeert et al., 2013] and Chapter 3. It is possible to update
the goal configuration during movement as utilized in [Prada et al., 2014],
and to avoid uncomplicated obstacles [Pastor et al., 2009; Ijspeert et al.,
2013; Pastor et al., 2013]. DMPs with compliant behavior were developed in
[Denisa et al., 2016; Batinica et al., 2017]. Further, movement can be gen-
erated from any configuration with guaranteed exponential convergence to
the goal [Perk and Slotine, 2006; Ijspeert et al., 2013; Wensing and Slotine,
2017]. Therefore, it was possible to physically push a robot to correct its
movement in [Talignani Landi et al., 2018a], knowing that it would reach
the goal after the perturbation.
However, in the original formulation, a DMP would continue its time
evolution despite any significant perturbations. The behavior of the robot
would then likely be undesirable and not intuitive, as discussed in [Ijspeert
et al., 2013]. A solution was also proposed in [Ijspeert et al., 2013]. Intu-
itively, the time evolution of the DMP was slowed down in case of deviation
from the intended trajectory. This is called temporal coupling, of which
phase stopping is one component. In this chapter, it is discovered that the
temporal coupling proposed in [Ijspeert et al., 2013] needs to be extended to
be realizable in practice, though the core concept in [Ijspeert et al., 2013] is
very promising and therefore forms a foundation for the method developed
in this chapter.
This chapter provides a practically realizable extension of the temporal
coupling in [Ijspeert et al., 2013], in the form of a two-degree-of-freedom
motion control system; see [Åström and Wittenmark, 2013] for a description
of the general architecture for two-degree-of-freedom control. The feedfor-
ward part of the controller promotes tracking of the DMP trajectory in the
absence of significant perturbations, thus mitigating unnecessarily slow tra-
jectory evolution due to temporal coupling acting on small tracking errors.
The feedback part suppresses significant errors. It is further shown theo-
retically that the control system is globally exponentially stable, which in
practice means that the goal configuration is guaranteed to be reached.
This chapter is organized as follows. Previous research is described
and evaluated in Sec. 5.2. The research problem addressed in this chap-
ter is specified in Sec. 5.3. The proposed control algorithm is presented
in Sec. 5.4. Thereafter, the proposed control algorithm is compared with
previous research through simulations in Sec. 5.5. These simulations also
support understanding of the key concepts of temporally coupled DMPs, and
this is why they are presented before the mathematical stability analysis
in Sec. 5.6. Further, experiments and results are presented in Secs. 5.7
and 5.8, a discussion is presented in Sec. 5.9, and concluding remarks are

50
 5.2 Previous Research on Temporally Coupled DMPs. . .

Table 5.1 Notation used in this chapter.

Notation Description
+
n ∈Z Dimension of robot configuration
y ∈ Rn Physical robot configuration
ÿr ∈ Rn Robot acceleration reference
τ ∈ R+ Time parameter
ˆ ∈ Rn Goal state
x ∈ R+ Phase variable
f ( x) ∈ Rn Learnable virtual forcing term
α, β , α x ∈ R+ Positive constants
e ∈ Rn Low-pass filtered configuration error
yc ∈ Rn Coupled robot configuration
τa ∈ R+ Adaptive time parameter
αe , k t , kc ∈ R+ Positive constants
z ∈ Rn DMP state
Kp, Kv ∈ R+ Large control gains
kp , kv ∈ R+ Moderate control gains
yu ∈ Rn Unperturbed, uncoupled trajectory
Nb ∈ Z+ Number of basis functions
ξ ∈ R5n+1 DMP state vector
qξ i q ∈ R+ 2-norm of DMP state i

finally presented in Sec. 5.10. Table 5.1 lists some of the notation used in
this chapter.
Successful perturbation recovery using the proposed method is visual-
ized in a video available online [Karlsson, 2016]. The video will be explained
in more detail in this chapter, but could already at this point support un-
derstanding of the functionality.
A code example is available in [Karlsson, 2017c], to allow for exploration
of the system proposed. The system was also integrated in the Julia DMP
package in [Bagge Carlson, 2016], originally based on [Ijspeert et al., 2013].

5.2 Previous Research on Temporally Coupled DMPs for

Perturbation Recovery
We here consider the case where a perturbation is introduced, so that the
reference acceleration ÿr in

τ 2 ÿr = α(β (ˆ − y) − τ ẏ) + f ( x) (5.1)

51
Chapter 5. Temporally Coupled Dynamical Movement Primitives in Rn

can not be realized, i.e., the robot configuration y can not evolve according
to plan. In the original DMP formulation, the phase variable would evolve
independently of any perturbation, according to

τ ẋ = −α x x (5.2)

This causes undesired behavior, since it is then likely that the actual tra-
jectory y deviates significantly from the intended trajectory even after the
cause of the perturbation has vanished. This is more thoroughly described
in [Ijspeert et al., 2002; Ijspeert et al., 2013]. Consider, for instance, the
case where the robot is stopped during its movement until x has converged
to 0. This implies that f ( x) = 0, and once released the robot will move to
its goal without taking the parameters of f ( x) into account. In such a case,
the value of the phase variable does not correspond to the progress of the
movement. To mitigate this problem, the solution described in the following
paragraph was suggested in [Ijspeert et al., 2013].
A coupling term C t and an adaptive time parameter τa were introduced.

ė = α e ( y − yc − e) (5.3a)
Ct = kt e (5.3b)
τa = 1 + k c e e
T
(5.3c)

Here, α e , k t , and k c are constant parameters, and e is low-pass filtered

position error. The parameter τa is used to determine the evolution rate
of the entire dynamical system. Further, yc denotes a coupled version of y,
and it is a simulated trajectory that can adjust its evolution rate based on
τa . In contrast, y continues to represent the physical trajectory of the robot.
Further, the evolution of x was determined by

τa ẋ = −α x x (5.4)

A larger value of e yields a larger value of τa and hence slower evolution

rate. One important aspect is that τa and x are scalars, that are shared
across all degrees of freedom of the movement. This supports coordination
between the different degrees of freedom. Note that (5.1) can be written
as a first-order dynamical system, given that the reference acceleration is
realized.

τ ż = α(β (ˆ − y) − z) + f ( x) (5.5a)
τ ẏ = z (5.5b)

Based on this, the dynamics of yc were defined as follows.

τa ż = α(β (ˆ − yc ) − z) + f ( x) + C t (5.6a)
τa ẏc = z (5.6b)

52
 5.3 Problem Formulation

A PD controller given by

ÿr = K p ( yc − y) + K v ( ẏc − ẏ) (5.7)

was used to drive y to yc . Here, ÿr denotes reference acceleration, while K p

and K v are control gains.
The ability to slow down the evolution of x in (5.4) is sometimes called
phase stopping [Abu-Dakka et al., 2015], whereas the entire functionality
of affecting the DMP evolution rate through τa and retaining coordination
between the degrees of freedom is called temporal coupling [Ijspeert et al.,
2013].

Evaluation of Previous Research

The approach from previous research has taken several important parts
of perturbation recovery into account, and it should be emphasized that
it forms the foundation of the research presented in this chapter. In this
section, however, some aspects are considered that motivate extension of
the state-of-the-art framework.
Denote by yu an unperturbed trajectory generated by an uncoupled DMP,
according to (5.1). In the absence of significant perturbations, it is desirable
that the trajectory generated by a temporally coupled DMP, y, evolves in
the same way as yu . To achieve this, y should follow yc closely. If this would
not be achieved, in addition to the deviation itself, y and yc would be slowed
down compared to yu , due to the temporal coupling according to (5.3c). This
phenomenon is shown in Sec. 5.5. In [Ijspeert et al., 2013], very high control
gains for (5.7) were suggested, which would have mitigated this issue under
ideal conditions and unlimited magnitude of the control signals. Specifically,
K p = 1000 and K v = 125 were chosen. However, even for moderate pertur-
bations, this would imply control signals too large to be realized practically.
For instance, a position error in Cartesian space of 0.1 m would yield ÿr =
100 m/s2 , whereas the YuMi robot has a maximum end-effector acceleration
in the order of 10 m/s2 . In Figs. 5.1 and 5.2, two simulated examples are
displayed: one where the actual movement was stopped, and one where it
was moved away from the intended path. The method described in [Ijspeert
et al., 2013] was used for recovery, and prohibitively large values of ÿr were
generated. This control system is also sensitive to noise and has a dan-
gerously low delay margin of 12 ms. The simulations are more thoroughly
described in Sec. 5.5.

5.3 Problem Formulation

In this chapter, we address the question of whether perturbations of DMPs
could be recovered from, while fulfilling the following requirements. Only

53
Chapter 5. Temporally Coupled Dynamical Movement Primitives in Rn

1
y

Position [m]
yc
0.5 yu

0 1 2 3 4 5 6 7 8
50
Acceleration [m/s2 ]

ÿr
0 ÿc
ÿu
−50

−100
0 1 2 3 4 5 6 7 8
Time [s]

Figure 5.1 Simulated trajectories, where y was subjected to a stopping

perturbation from 2 s to 3 s, using the approach in [Ijspeert et al., 2013].
When y was stopped, the evolution of yc slowed down, and when y was
released, it was driven to yc and then behaved like a delayed version of
yu . This behavior was desired. However, a prohibitively large acceleration
reference ÿr was generated.

moderate control signals must be used. The benefits of the DMP framework
described in [Ijspeert et al., 2013], i.e., scalability in time and space as well
as guaranteed convergence to the goal ˆ , must be preserved. Further, in
the absence of significant perturbations, the behavior of y should resemble
that of the original DMP framework described in [Ijspeert et al., 2013] and
Chapter 3.
The stability of the proposed control algorithm requires special attention.
Unstable robot motion control could damage the robot and its surroundings,
such as tooling and work pieces. Further, in robotic manipulation it is crucial
that the robot reaches its goal configuration in each of its movements. If
this would not be achieved, subtasks would likely be left incomplete, yielding
unforeseen hardware configurations, which in turn could result in collision
and broken hardware. For temporally coupled DMPs to be used on a larger
scale in the future, it is therefore necessary to prove that these result in
stable behavior. Stability for the original DMP framework, i.e., not including
temporal coupling, was concluded in [Perk and Slotine, 2006; Wensing and
Slotine, 2017].
Specifically, it will in this chapter be investigated whether the proposed
temporally coupled DMPs are globally exponentially stable. A mathematical

54
 5.4 Method

1.5
y

Position [m]
yc
1
yu
0.5

0
0 1 2 3 4 5 6 7 8
200
Acceleration [m/s2 ]

ÿr
ÿc
0 ÿu

−200

0 1 2 3 4 5 6 7 8
Time [s]

Figure 5.2 Similar to Fig. 5.1, except that y was moved away from the
intended path between 2 s and 3 s. Again, a prohibitively large acceleration
reference ÿr was generated.

definition of exponential stability can be found in, e.g., [Slotine and Li,
1991]. In words, a system is globally exponentially stable if the state vector
converges to the origin faster than an exponentially decaying function. In
the context of DMPs, global exponential stability implicates that the robot
eventually reaches the goal configuration.

5.4 Method
The proposed method extends the approach in [Ijspeert et al., 2013] as
follows. The PD controller in (5.7) is augmented with feedforward control,
as shown in (5.8). Further, the PD controller gains are moderate, to get
a practically realizable control signal. Additionally, the time constant τ is
introduced as a factor in the expression for the adaptive time parameter τa ,
see (5.3c) and (5.9). Our method is detailed in the following.
To enforce y to follow yc , we apply the following control law.

ÿr = k p ( yc − y) + k v ( ẏc − ẏ) + ÿc (5.8)

Here, ÿc is obtained from feedforward of the acceleration of yc . This allows

for the controller to act also for zero position and velocity errors. Therefore,
the trajectory tracking works also for moderate controller gains: k p = 25
and k v = 10 are used throughout this chapter. With these gains, the closed

55
Chapter 5. Temporally Coupled Dynamical Movement Primitives in Rn

ÿc

ˆ yc P d P ÿr y
DMP kp + kv (·) Robot
dt

-1

Figure 5.3 Schematic overview of the control structure described in

Sec. 5.4. The block denoted ’Robot’ includes the internal controller of the
robot.

control loop has a double pole given by −5 rad/s. Since the real parts are
negative, the system (5.8) is asymptotically stable, and since the imaginary
parts are 0, it is critically damped. Such a critically damped system is
obtained for k p = k2v /4. This relationship is advisable to fulfill in the imple-
mentation, since overshoot is then avoided, which is important for avoiding
collision. This is especially important in manipulation that involves contact
between robot and environment; see also Chapter 3, where α and β are con-
strained based on the same reasoning. The delay margin is 130 ms, which
is an improvement compared to 12 ms for the previous method, described in
Sec. 5.2. A schematic overview of the control system is shown in Fig. 5.3.
Further, the relation (5.3c) is modified in order to include the original
time constant τ , as follows.

τa = τ (1 + k c eT e) (5.9)

The coupling term C t is omitted in this present method. This choice is

elaborated on in Sec. 5.9.
Since τa is not constant over time, determining ÿc is more complicated
than determining ÿ by differentiating (5.5b). One option would be to ap-
proximate ÿc by discrete-time differentiation of ẏc . However, instead we
determine the instantaneous acceleration analytically as follows.

3 4
d d z τa ż − τ̇a z żτa − 2τ k c ( eT ė) z
ÿc = ( ẏc ) = = = (5.10)
dt dt τa τa2 τa2

where ė and ż are given by (5.3a) and (5.6a) with C t = 0, respectively. It is

noteworthy that the computation of ÿc does not require any first or second-
order time derivatives of any measured signal, which would have required

56
 5.5 Simulations

prior filtering to mitigate amplification of high-frequency noise. Similarly,

ẏc is determined by (5.6). In contrast, the computation of ẏ should be com-
plemented with a low-pass filter, to mitigate amplification of measurement
noise.
We model the ’Robot’ block in Fig. 5.3 as a double integrator, so that
ÿ = ÿr , see Sec. 5.9. In summary, the proposed control system is given by

ÿr = k p ( yc − y) + k v ( ẏc − ẏ) + ÿc (5.11a)

ė = α e ( y − yc − e) (5.11b)
τa = τ (1 + k c e e)T
(5.11c)
τa ẋ = −α x x (5.11d)
τa ẏc = z (5.11e)
τa ż = α(β (ˆ − yc ) − z) + f ( x) (5.11f )

Feedforward control has been used in the DMP context previously, but then
only for low-level joint control, with motor-torque commands as control sig-
nals, see [Park et al., 2008; Pastor et al., 2009]. This control structure was
also applied in the internal controller used in the implementation in this
chapter. This inner control design should not be confused with the feedfor-
ward control described in this section, which operated outside the internal
robot controller, and was used to determine the reference acceleration for
the robot.

5.5 Simulations
In the simulations, the robot configuration consisted of position in Cartesian
space. First, a demonstrated trajectory, ydemo , was simulated by creating a
time series of positions. Thereafter, the corresponding DMP parameters
were determined. Simulating the resulting DMP without perturbations and
without temporal coupling yielded an unperturbed, uncoupled, trajectory yu .
Position in one dimension was simulated in the following, and two different
perturbations were considered: one where y was stopped, and one where it
was moved. The perturbations took place from time 2 s to 3 s. The systems
were sampled at 250 Hz, using the forward Euler method. The same DMP,
yielding the same yu , was used in each trial. The adaptive time parameter τa
was determined according to (5.9) in all simulations, to get comparable time
scales. First, the controller detailed in [Ijspeert et al., 2013] was applied.
Except for the perturbations themselves, the conditions were assumed to
be ideal, i.e., no delay and no noise were present. The results are shown in
Figs. 5.1 and 5.2. Despite ideal conditions, prohibitively large accelerations
were generated by the controller in both cases.

57
Chapter 5. Temporally Coupled Dynamical Movement Primitives in Rn

1
y

Position [m]
yc
0.5 yu

−0.5
0 2 4 6 8 10 12 14
5
Acceleration [m/s2 ]

ÿr
ÿc
0
ÿu

−5

−10
0 2 4 6 8 10 12 14
Time [s]

Figure 5.4 Simulation with the control structure in [Ijspeert et al., 2013],
except that the gains were lower ( K p = 25 and K v = 10). The reference
acceleration was of realizable magnitude. For instance, the YuMi robot has
a maximum end-effector acceleration of approximately 10 m/s2 . However, the
coupled and real systems were slowed down due to small tracking errors
combined with temporal coupling.

Figure 5.4 shows the result from a simulation where the controller
detailed in [Ijspeert et al., 2013] was used, except that the gains were
lowered to moderate values. The conditions were ideal, and no perturbation
was present. This resulted in realizable control signals. However, small
control errors in combination with the temporal coupling slowed down the
evolution of the coupled system as well as the actual movement.
Thereafter, the controller proposed in this chapter, described in Sec. 5.4,
was used. In order to verify robustness under realistic conditions, noise
and time delay were introduced. Position measurement noise, and velocity
process noise, were modeled as zero-mean Gaussian white noise, with stan-
dard deviations of 1 mm and 1 mm/s, respectively. The time delay between
the process and the controller was 12 ms. This delay was suitable to simu-
late since it corresponds both to the delay margin of the method suggested
in [Ijspeert et al., 2013], and to the actual delay in the implementation
presented in this chapter, see Sec. 5.7. (It is, however, a coincidence that
these two have the same value. Nevertheless, this shows that a 12 ms de-
lay margin is not necessarily enough.) The results are shown in Figs. 5.5
and 5.6. For comparison, the method in [Ijspeert et al., 2013], with the
large gains, was also evaluated under these conditions, although without

58
 5.6 The Proposed Temporally Coupled DMPs are GES

1.5
y

Position [m]
yc
1
yu
0.5

0
0 1 2 3 4 5 6 7 8
20
Acceleration [m/s2 ]

ÿr
10 ÿc
ÿu
0

−10
−20
0 1 2 3 4 5 6 7 8
Time [s]

Figure 5.5 Similar to Fig. 5.1, but with modeled noise and delay, and using
the controller proposed in this chapter. The behavior was satisfactory both
regarding position and acceleration.

any perturbation except for the noise. Because of the time delay, this system
was unstable, as shown in Fig. 5.7.
Finally, two-dimensional movement using the proposed control system
was simulated. The purpose of these simulations was to investigate the
convergence of the proposed control system. At the start of each simulation,
yc was set to the beginning of the demonstrated trajectory ydemo , and y
was deliberately initialized with a randomly chosen error with respect to
yc . The system was simulated 100 times. The simulations gave mutually
similar results. The results from some of the simulations are visualized in
Figs. 5.8 and 5.9. For each simulation, the position converged to the goal.
Further, the DMP states in Fig. 5.9 converged to 0. This is consistent with
the theory that will be presented in Sec. 5.6.

5.6 The Proposed Temporally Coupled DMPs are

Globally Exponentially Stable
In this section, we analyze the stability properties of the proposed tem-
porally coupled DMPs. This analysis requires that the state space is con-
tractible, which is true for Rn (see Chapter 6). See also Sec. 5.A, where
passivity is shown. The entire control system described in Sec. 5.4 is given

59
Chapter 5. Temporally Coupled Dynamical Movement Primitives in Rn

1.5
y

Position [m]
yc
1
yu
0.5

0
0 1 2 3 4 5 6 7 8
20
Acceleration [m/s2 ]

ÿr
10 ÿc
ÿu
0

−10
−20
0 1 2 3 4 5 6 7 8
Time [s]

Figure 5.6 Similar to Fig. 5.5, except that y was moved away from the
intended path between 2 s and 3 s. Again, the behavior was satisfactory both
regarding position and acceleration.

·1022
0.5
Position [m]

y
−0.5 yc
yu
−1
0 1 2 3 4 5 6 7
·1026
1
Acceleration [m/s2 ]

0.5

0 ÿr
−0.5 ÿc
ÿu
−1
0 1 2 3 4 5 6 7
Time [s]

Figure 5.7 Using the control system in [Ijspeert et al., 2013], subject to
the simulated noise and time delay, resulted in unstable behavior.

60
 5.6 The Proposed Temporally Coupled DMPs are GES

0.6
y
yc ˆ
0.4 ydemo

y2 [m] 0.2

−0.2

−0.4

−0.6
−0.4 −0.2 0 0.2 0.4 0.6 0.8 1
y1 [m]

Figure 5.8 Positions of simulated trajectories. The movement direction is

indicated by the arrows. For an uncluttered view, only 10 of the 100 simulated
trajectories are displayed. For each simulation, y started to the left in the
plot, with some distance to yc . Then, y moved toward yc . Subsequently, both
y and yc converged to the goal position ˆ , which is marked with a circle at
the upper right of the plot. Since yc was initialized to the same point for
each simulation, it always followed the same path. Further, it followed the
demonstrated path closely, which is expected given a sufficient number of
basis functions ( Nb = 50 were used).

by

ÿr = k p ( yc − y) + k v ( ẏc − ẏ) + ÿc (5.12a)

ė = α e ( y − yc − e) (5.12b)
τa = τ (1 + k c eT e) (5.12c)
τa ẋ = −α x x (5.12d)
τa ẏc = z (5.12e)
τa ż = α(β (ˆ − yc ) − z) + f ( x) (5.12f )

61
Chapter 5. Temporally Coupled Dynamical Movement Primitives in Rn

0.3 q y − yc q [m]
q ẏ − ẏc q [m/s]
0.2
qξ i q
q eq [m]
0.1

0
0 2 4 6 8 10 12 14 16 18 20
1 x [1]
q yc − ˆq [m]
qξ i q

q zq [m/s]
0.5

0
0 2 4 6 8 10 12 14 16 18 20
Time [s]

Figure 5.9 The magnitudes of DMP states ξ i plotted over time, for one
of the simulations displayed in Fig. 5.8. The notation q · q represents the
2-norm, and the unit symbol [1] indicates dimensionless quantity. It can be
seen that y converged to yc , and that e converged to 0. As will be shown in
Sec. 5.6, this is predicted by Theorem 5.1. Further, each state converged to
0, which is predicted by Theorem 5.2.

We introduce the state vector ξ as

y − yc
 
 ẏ − ẏc 
 
 e 
 
ξ =
 x 
 (5.13)
 
 
 yc − ˆ 
z
and write the system on state-space form,
ẏ − ẏc
   
y − yc
 − kp ( y − yc ) − kv ( ẏ − ẏc ) 
 ẏ − ẏ   
 c  
α e ( y − yc − e)
 
 e   


d 
   αx 
− x
 
 x = 
(5.14)
dt    τa 

   
  1
 y − ˆ  z

 c  
  τa 
α
  
1 
z (β (ˆ − yc ) − z) + f ( x)
τa τa

62
 5.6 The Proposed Temporally Coupled DMPs are GES

The states y − yc , ẏ − ẏc , and e have the same dimension, which we

denote by n. Let In be the identity matrix, and 0n the zero matrix, each of
size n. The upper part of (5.14) is linear and can be written as
    
y − yc 0n In 0n y − yc
d 
 ẏ − ẏc  = − kp In − k v In   ẏ − ẏc 
   
0n (5.15)
dt
e α e In 0n −α e In e

Denote by ξ¯ and Ā the state vector and the system matrix in (5.15), respec-
tively.

Theorem 5.1
The dynamical system defined by (5.15) for DMP operation has ξ¯ = 0 as a
globally exponentially stable equilibrium. 2

Proof. For ξ¯ = 0 we have dξ¯/d t = 0 in (5.15), so ξ¯ = 0 is an equilibrium.

The system is globally asymptotically stable if the real part of each eigen-
value of Ā is strictly negative [Glad and Ljung, 2000]. With k p = k2v /4 (see
Sec. 5.4) the eigenvalues are given by

kv
λ1,...,2n = − (5.16a)
2
λ2n+1,...,3n = −α e (5.16b)

These are strictly negative, since k v , α e > 0. The linear system is hence
globally asymptotically stable. For linear systems, asymptotic stability is
equivalent with exponential stability [Slotine and Li, 1991; Rugh, 1996].
The system is therefore globally exponentially stable. 2

Next, we will show that the system given by

 
y − yc ẏ − ẏc
 

− kp ( y − yc ) − kv ( ẏ − ẏc )
 ẏ − ẏ   
d  c
= 

 e   α e ( y − yc − e) (5.17)
dt

 
αx
   
x − x
τa

is contracting, of which a definition can be found in [Lohmiller and Slotine,

1998]. We note that this is a hierarchical system, consisting of (5.15), which
does not depend on x, and of ẋ = −α x x/τa . To show contraction, we will
use the following proposition.

63
Chapter 5. Temporally Coupled Dynamical Movement Primitives in Rn

Proposition 5.1
If ẋ1 = ˆ 1 ( x1 ) is contracting, and ẋ2 = ˆ 2 ( x1 , x2 ) is contracting for each
fixed x1 , then the hierarchy
3 4 3 4
d x1 ˆ 1 ( x1 )
= (5.18)
dt x2 ˆ 2 ( x1 , x2 )

is contracting. 2

This follows directly from Proposition 2 in [Wensing and Slotine, 2017], by

applying it to autonomous systems.
Proposition 5.2
The system given by (5.17) is contracting. 2

Proof. We know from Theorem 5.1 that (5.15) is globally exponentially

stable. It is therefore contracting [Lohmiller and Slotine, 1998]. For the fixed
point ξ¯ = 0, we have ẋ = −α x x/τ , which is contracting since −α x /τ < 0.
It now follows from Proposition 5.1 that the hierarchical combination (5.17)
is contracting. 2

We now address the stability of the entire control system in (5.14) as follows.
Theorem 5.2
The system given by (5.14) for DMP operation has ξ = 0 as a globally
exponentially stable equilibrium point. 2

Proof. Since x is a factor of f ( x) and the remaining part of f ( x) is bounded,

f (0) = 0, see Chapter 3 and [Ijspeert et al., 2013]. It can therefore be seen
that ξ = 0 yields dξ /d t = 0, so ξ = 0 is an equilibrium point. It remains to
show the stability. We know from Proposition 5.2 that (5.17) is contracting,
and it can be seen that it has the origin as a fixed point. Consider now the
remaining part of (5.14), i.e.,
 
A B 1
z
d yc − ˆ  τa 
=  (5.19)
dt z  α 1 
(β (ˆ − yc ) − z) + f ( x)
τa τa

For the fixed point of (5.17) we have τa = τ and f ( x) = f (0) = 0, and the
system (5.19) then simplifies to
 
A B 0n
1
In  A y − ˆ B
d yc − ˆ τ  c

= (5.20)
dt z  αβ α  z
− In − In
τ τ

64
 5.7 Experiments

This system is linear, and with β = α/4 (see Chapter 3) the eigenvalues of
the system matrix are given by
α
µ1,...,2n = − (5.21)
2τ
Since the eigenvalues are strictly negative, the system (5.20) is contracting.
Further, we note that (5.14) is hierarchical from (5.17) to (5.19). Therefore,
it follows from Proposition 5.1 that (5.14) is contracting. This is true for
the whole state space. Hence, all solutions of (5.14) converge exponentially
to the same trajectory. Since one solution is given by ξ = 0, they must
all converge exponentially to this equilibrium point. Thus, (5.14) is globally
exponentially stable. 2
Since y = ˆ for ξ = 0, Theorem 5.2 implies that the system (5.14)
converges exponentially to a state where the goal configuration is reached
by the actual robot position.

5.7 Experiments
The algorithm in Sec. 5.4 was implemented on the ABB YuMi prototype
robot, see Chapter 2. Similar to the simulations, the control system ran at
250 Hz, and the delay between robot and controller was 3 sample periods,
corresponding to 12 ms. The computations took place in joint space, and the
robot’s forward kinematics were used for visualization in Cartesian space
in some of the figures presented.
Six different setups were used to evaluate the control algorithm. For
each setup, 50 trials were made, yielding a total of 300 trials. Prior to each
trial, a temporally coupled DMP had been determined from demonstration
by means of lead-through programming [Stolt et al., 2015a]. In each trial,
the temporally coupled DMP was executed while data were logged and
saved. Perturbations were introduced by physical contact with a human.
A wrist-mounted ATI Mini40 force/torque sensor was used to measure the
contact force, and a corresponding acceleration was added to ÿr as a load
disturbance. The purpose of Setups 5.1 and 5.2 was to demonstrate the
proposed functionality in general and examine the reference acceleration,
whereas the main purpose of Setups 5.3 to 5.6 was to verify the exponential
stability shown in Sec. 5.6. The six setups were as follows.
Setup 5.1 This setup is visualized in Fig. 5.10. The task of the robot was
to place a stop button in its corresponding case. The assembly parts are
shown in Fig. 5.11. The human introduced two perturbations during the
DMP execution. The first perturbation was formed by moving the end-
effector away from its path, and then releasing it. The second perturbation
consisted of a longer, unstructured, movement later along the trajectory.

65
Chapter 5. Temporally Coupled Dynamical Movement Primitives in Rn

(a) (b)

(c) (d)

Figure 5.10 Setup 5.1. In (a), the robot started to execute a DMP for placing
the stop button in the rightmost yellow case. A human perturbed the motion
twice. The first perturbation (b) was formed by moving the end-effector away
from its path, and then releasing it. The second perturbation (c) lasted for a
longer time, and consisted of unstructured movement. The robot recovered
from both perturbations, and managed to place the stop button in the case
(d). Data from one trial are shown in Fig. 5.15.

Figure 5.11 Yellow case (left), stop button (upper right) and gasket (lower
right) used in Setups 5.1, 5.2 and 5.6.

66
 5.7 Experiments

Setup 5.2 This setup is visualized in Fig. 5.12. A human co-worker realized
that the stop buttons in the current batch were missing rubber gaskets, and
acted to modify the robot trajectory, allowing for the co-worker to attach
the gasket on the stop button manually. During execution of the DMP,
the end-effector was stopped and lifted to a comfortable height by the co-
worker. Thereafter, the gasket was attached, and finally the end-effector
was released. For the sake of completeness, the modified trajectory was
used to form yet another DMP, which allowed for the co-worker to attach
the gaskets without perturbing the trajectory of the robot, for the remaining
buttons in the batch. To verify this functionality, one such modified DMP
was executed at the end of each trial.

Setup 5.3 The robot moved one of its arms without manipulating any ob-
jects. The arm was subjected to two perturbations.

Setup 5.4 The robot moved both of its arms simultaneously, again without
any manipulation. Two perturbations were introduced: first one on the left
arm, and subsequently one on the right arm.

Setup 5.5 This setup is visualized in Fig. 5.13. The robot used both arms
to pick up a ball. One perturbation was introduced on the left arm.

Setup 5.6 This setup is visualized in Fig. 5.14. The objective of the robot
was to mate a stop button, grasped with the right arm, with its correspond-
ing case, grasped with the left arm. The left arm was subjected to one
perturbation.

Videos
A video of Setups 5.1 and 5.2 is available in [Karlsson, 2016]. Further, a
video of Setups 5.3 to 5.6 is available in [Karlsson, 2017a]. In Part I of the
latter video, the experimental setups are shown, except that the temporally
coupled DMPs were executed without any perturbations present. This is to
visualize what should be achieved by the robot, in each setup. In Part II,
DMPs without temporal coupling were used, and the robot was subject to
perturbations. The purpose of this part is to demonstrate that the robot
risks to fail in such a scenario, and hence motivate why temporal coupling
is necessary. This risk of failure was also indicated in [Ijspeert et al., 2013],
using one-dimensional simulation examples. Part III of the video shows the
experiments, where temporally coupled DMPs were executed in the presence
of perturbations.

67
Chapter 5. Temporally Coupled Dynamical Movement Primitives in Rn

(a) (b)

(c) (d)

(e) (f)

Figure 5.12 Setup 5.2. The robot started its motion toward the rightmost
yellow case in (a). The end-effector was stopped and lifted, and the gasket
was mounted in (b). The robot was then released, and continued its motion
to the case, (c) and (d). The actual trajectory was saved and used to form a
modified DMP, and the robot was reset to a configuration similar to that in
(a). When executing the modified DMP, the human co-worker could attach
the gasket without perturbing the motion of the robot (e). The robot finished
the modified DMP in (f). Data from one trial are shown in Fig. 5.16.

68
 5.8 Results

(a) (b)

(c) (d)
Figure 5.13 Setup 5.5. The robot task was to pick up the blue ball us-
ing both arms. The movement started in (a), and a human perturbed the
movement in (b). Thereafter, the robot was released and recovered from the
perturbation. The arms reached the ball simultaneously in (c). The goal state
was then reached in (d). Data from one trial are shown in Fig. 5.19.

5.8 Results
In this section, data from one trial of each of the setups in Sec. 5.7 are
displayed. For each setup, all trials gave qualitatively similar results. In
each of the 300 trials, the perturbations were first recovered from and
thereafter the goal state was reached.
Data from a trial of Setup 5.1 are displayed in Fig. 5.15. As intended,
the two disturbances were successfully recovered from. The reference accel-
eration was of realizable magnitude.
Data from a trial of Setup 5.2 are displayed in Fig. 5.16. First, the
perturbation was successfully recovered from as intended. The reference
acceleration was of realizable magnitude. When the modified DMP was exe-
cuted, it behaved like a smooth version of the perturbed original trajectory.
Data from Setups 5.3 to 5.6 are shown in Figs. 5.17 to 5.20. The figures
show the magnitudes of the states in (5.14) represented in joint space, plot-
ted over time. The perturbations are clearly visible in the data. It can be

69
Chapter 5. Temporally Coupled Dynamical Movement Primitives in Rn

(a) (b)

(c) (d)
Figure 5.14 Setup 5.6. The task was to insert the red stop button into
the corresponding hole in the yellow case. The initial configuration is shown
in (a). In (b), a human perturbed the movement, and in (c) the robot had
recovered from the perturbation. The goal configuration was reached in (d).
Data from one trial are shown in Fig. 5.20.

seen that y converged to yc , and that e converged to 0, after each perturba-

tion. This is expected from Theorem 5.1. The data also show that each state
converged to 0, as predicted by Theorem 5.2.

5.9 Discussion
Compared to the previous research in [Ijspeert et al., 2013], the method
in this chapter contains the following extensions. Feedforward control was
added to the PD controller, thus forming a two-degree-of-freedom controller.
Further, the PD controller gains were reduced to moderate magnitudes. The
expression for τa was also modified, to include the original time constant τ
as a factor. These changes resulted in the following benefits, compared to
the previous method. The feedforward part allowed for the controller to act
also for insignificant position and velocity error, thus improving the trajec-
tory tracking. Because of this, the large controller gains used in previous

70
 5.9 Discussion

y
−0.05 yu
y2 [m]
−0.1

−0.15

0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

y1 [m]
0.15
q y − yc q [m]

0.1

0.05

0
0 5 10 15 20 25
0
y
−0.05 yc
y2 [m]

−0.1 yu

−0.15
−0.2
0 5 10 15 20 25
3
q ÿr q [m/s2 ]

2
1
0
0 5 10 15 20 25
Time [s]

Figure 5.15 Experimental data from a trial of Setup 5.1. The first (from
above) plot shows two position coordinates, y1 and y2 , of the end-effector in
Cartesian space. The arrow indicates the direction of the movement, which
started in the upper right and finished in the lower left of the plot. The two
perturbations are clearly visible. The second plot shows the distance between
y and yc over time. In the third plot, it can be seen that the evolution of yc
slowed down during each perturbation. Subsequently, y recovered, and when
it was close to yc , the movement continued as a delayed version of yu . The
reference acceleration was of realizable magnitude, as shown in the fourth
plot.

71
Chapter 5. Temporally Coupled Dynamical Movement Primitives in Rn

−0.05 y
yu
ym

−0.1
y2 [m]

−0.15

0 0.05 0.1 0.15 0.2 0.25 0.3 0.35 0.4

y1 [m]
0.1
q y − yc q [m]

0.08
0.06
0.04
0.02
0
0 5 10 15 20 25
0
y
−0.05 yc
y2 [m]

−0.1 yu

−0.15
−0.2
0 5 10 15 20 25
3
q ÿr q [m/s2 ]

0
0 5 10 15 20 25
Time [s]

Figure 5.16 Experimental data from a trial of Setup 5.2. The organization
of this figure is similar to that of Fig. 5.15. The perturbation for stopping
and lifting the end-effector took place from time 10 s to 17.5 s, and is clearly
visible in each plot. This perturbation was recovered from as intended, and
the reference acceleration was of realizable magnitude. The uppermost plot
also displays the measured trajectory obtained by executing the modified
DMP, denoted ym . It behaved like a smooth version of the perturbed original
trajectory y.

72
 5.9 Discussion

0.6 q y − yc q [rad]
q ẏ − ẏc q [rad/s]
0.4
qξ i q q eq [rad]
0.2

0
0 5 10 15 20 25
1.5 x [1]
q yc − ˆq [rad]
1
qξ i q

q zq [rad/s]
0.5

0
0 5 10 15 20 25
Time [s]

Figure 5.17 Data from a trial of Setup 5.3, where the left arm was per-
turbed twice during its movement.

1
q y − yc q [rad]
0.8
q ẏ − ẏc q [rad/s]
0.6
qξ i q

q eq [rad]
0.4
0.2
0
0 5 10 15 20 25
x [1]
1 q yc − ˆq [rad]
qξ i q

q zq [rad/s]
0.5

0
0 5 10 15 20 25
Time [s]

Figure 5.18 Data from a trial of Setup 5.4, where both arms moved simul-
taneously and were perturbed once each.

73
Chapter 5. Temporally Coupled Dynamical Movement Primitives in Rn

0.8
q y − yc q [rad]
0.6 q ẏ − ẏc q [rad/s]
qξ i q q eq [rad]
0.4
0.2
0
0 5 10 15 20 25
1.5 x [1]
q yc − ˆq [rad]
qξ i q

1 q zq [rad/s]

0.5

0
0 5 10 15 20 25
Time [s]

Figure 5.19 Data from a trial of Setup 5.5. The setup is shown in Fig. 5.13.

0.6
q y − yc q [rad]
0.4 q ẏ − ẏc q [rad/s]
qξ i q

q eq [rad]
0.2

0
0 5 10 15 20 25
2
x [1]
1.5 q yc − ˆq [rad]
qξ i q

1 q zq [rad/s]

0.5
0
0 5 10 15 20 25
Time [s]

Figure 5.20 Data from a trial of Setup 5.6. The setup is shown in Fig. 5.14.
In this particular trial, the robot was grasped by the human for a significantly
longer time than in the trials shown in Figs. 5.17 to 5.19, though this was
not the case for all trials of Setup 5.6.

74
 5.9 Discussion

research, that were used to mitigate significant tracking errors, could be

reduced to moderate magnitudes. In turn, using moderate gains instead of
very large ones, resulted in practically realizable control signals, instead
of prohibitively large ones. It also improved the delay margin significantly.
The modification of the expression for τa was not the main focus of this
research, but it was necessary since it allowed for the actual trajectory to
evolve according to the DMP, with time constant τ . Without this modifica-
tion, the time parameter τ would not have affected the trajectory generated
by the DMP. Instead, τa would have converged to 1, regardless of τ , which
would not have been desirable.
The work presented here focused on the control structure for trajectory
tracking and perturbation recovery, rather than on the perturbations them-
selves. Even though the perturbations in the experiments considered here
emerged from physical contact with a human, the control structure would
work similarly for any type of perturbation. There are many other possible
perturbations, e.g., a pause of the movement until a certain condition is ful-
filled, superpositioned motion-control signals to explore the surroundings
with a force/torque sensor, a detour to enable line of sight between a camera
and a part of the work space, exploration in reinforcement learning, or any
other unforeseen deviation from the reference trajectory defined by a DMP.
Already in the original DMP versions, without temporal coupling, a
phase variable x is used instead of time [Ijspeert et al., 2013; Pastor et al.,
2013]. This has been motivated by easier resetting, and by results that
suggest that some biological systems do not have access to time; see, e.g.,
[Keating and Thach, 1997]. However, without temporal coupling, x evolves
deterministically according to (5.2). Solving the differential equation yields
1 α 2
x
x = x0 exp − t (5.22)
τ
where t represents time and the initial value x0 could be set to 1; see
Chapter 2. Hence, an expression with explicit time could be used instead of
x; see also [Chen et al., 2016] where explicit time was used to determine the
virtual forcing term. However, this is not as straightforward when temporal
coupling is included. The evolution rate of real time can not be controlled,
and some virtual time with variable time derivative would have to be used.
The phase variable x takes the role of such virtual time. Hence, one could
argue that the phase variable is truly justified only for temporally coupled
DMPs.
The coupling term C t has been introduced in previous research to drive
yc toward y when these were different, see Sec. 5.2. However, whether this
effect is desired, and to what extent, is context dependent. Further, the
effect would be mitigated by the temporal coupling, which would slow down
the evolution of yc in (5.6). Which of these effects that would be dominant

75
Chapter 5. Temporally Coupled Dynamical Movement Primitives in Rn

in different cases would be difficult to predict intuitively. For these two

reasons, the coupling term was not included in the method proposed here,
though it would be straightforward to implement. It was, however, included
in the simulations where the previous method, described in Sec. 5.2, was
evaluated. During the perturbations in Figs. 5.1 and 5.2, the effect of the
temporal coupling was dominant, as yc did not approach y significantly.
It is important to note that the ’Robot’ block in Fig. 5.3 includes both the
robot itself and the internal robot controller. It is possible for the internal
control loop to achieve a reference acceleration, ÿr , with small error, as long
as ÿr is reasonably smooth and limited in magnitude. Therefore, modeling
the robot as a double integrator can be justified by feedback linearization
[Slotine and Li, 1991] even though the dynamics of the robot itself (i.e.,
excluding internal controller, e.g., from joint torques to joint angles) are
more complicated than a double integrator.
A common approach to prove stability of a nonlinear system is to find
a Lyapunov function of the state vector. In Sec. 5.A, it is shown that this
is not straightforward for the control system considered here. Therefore,
instead the contraction theory in [Lohmiller and Slotine, 1998] was used in
this chapter. Some care must be taken when analyzing stability based on
contraction theory, because contraction alone does not imply stability. For
instance, the following system has been given in [Lohmiller and Slotine,
1998] as an example of a contraction.

ẋ1 = − x1 + exp( t) (5.23)

Here, x1 represents a scalar state and t is time. The system is indeed

a contraction because the Jacobian is −1 and hence uniformly negative
definite [Lohmiller and Slotine, 1998]. However, one solution is given by

exp( t)
x1 = (5.24)
2
which is not stable. It can be seen that x1 → ∞ as t → ∞. The contraction
guarantees exponential convergence to a single trajectory [Lohmiller and
Slotine, 1998], in this case to (5.24), but it does not guarantee exponential
convergence to a constant. Nevertheless, for autonomous systems such as
DMPs, global contraction is equivalent with global exponential stability
[Lohmiller and Slotine, 1998].
Even though Theorem 5.2 states that the state vector ξ converges to
the origin, in practice noise and model error prevent the state vector from
staying arbitrarily close to the origin. In the experiments there was an
unmodeled time delay of 12 ms between the controller in (5.8) and the
internal robot controller, which is an example of a model error. Further,
the robot end-effectors reached their targets with an accuracy of ±1 mm.

76
 5.9 Discussion

Because the origin could not be reached by ξ with exactly zero error, the
movement could for instance be considered finalized once qξ q < ρ for some
small constant ρ , or upon force-based detection of task completion as in
Chapter 7. A related discussion about convergence toward single-point goals
can be found in [LaValle, 2006]. There, the concept of so-called asymptotic
solution plans is also described. Since exponential stability is a stronger
requirement than asymptotic stability [Slotine and Li, 1991], Theorem 5.2
implies that temporally coupled DMPs are asymptotic solution plans.
Stability for the original DMP framework was shown in [Perk and Slo-
tine, 2006; Wensing and Slotine, 2017] by utilizing that f ( x) converged to 0,
which followed from the fact that x decayed exponentially, regardless of any
deviation from the intended movement. However, this is true only if tempo-
ral coupling is not used, and the convergence of f ( x) is less obvious for the
DMP version studied in this chapter. Due to the adaptive time parameter
τa in (5.4), it can not be assumed that x decays exponentially for temporally
coupled DMPs. Instead, the stability of temporally coupled DMPs has now
been established in the proof of Theorem 5.2.
In the experiments, a force sensor was mounted on each wrist of the
robot, and the measured forces were scaled and added to the reference
accelerations as disturbances, which allowed for the human to introduce
perturbations in an intuitive way. This arrangement worked well for the
sake of evaluating the stability properties of temporally coupled DMPs, and
it was straightforward to implement. However, for the purpose of interacting
physically with the robot, one could also consider passive force control.
Such an approach was presented in [Denisa et al., 2016], though temporal
coupling was not included. As suggested in [Denisa et al., 2016], it would
therefore be interesting to combine temporal coupling with passive force
control in the DMP framework in future research.
The effect of temporal coupling can be seen both in the simulated and
the experimental data. For example, consider Fig. 5.20 where the robot was
grasped by the human at t ( 2.5 s. During the perturbation, the evolution
of x, yc , and z was slowed down. Such behavior is desired in general, since it
helps to retain the coordination between the degrees of freedom of the robot.
It resulted from the fact that e had a significant magnitude, which in turn
caused τa to be significantly larger than τ , see (5.9). The robot was released
at t ( 10 s, and the controller described by (5.8) started to drive y to yc ,
whereby e was driven to 0. After e had converged to insignificant magnitude,
at approximately 12 s, it can be seen that x, yc , and z evolved considerably
faster. The same phenomenon could be seen in each experiment, as well as
in the simulations.
It is necessary to implement saturation on the control signals, to prevent
too large acceleration and velocity for large perturbations. Such boundaries
were implemented, but never reached in the experiments in this work.

77
Chapter 5. Temporally Coupled Dynamical Movement Primitives in Rn

In the current implementation, the actual trajectory returned to the

reference trajectory, approximately where it started to deviate. This might
not always be desired. For instance, it might sometimes be more practical
to connect further along the reference trajectory, e.g., after avoiding an
obstacle. A lower value of k c would result in such behavior. However, it
must then be known what value of k c that should be used. Further, one
could think of scenarios where it would not be desirable to connect to the
reference trajectory, e.g., if a human would modify the last part of the
trajectory to a new end point. Hence, future work includes development
of a method to determine the desired behavior after a perturbation. The
method presented in this chapter would be useful for executing the desired
behavior, once it could be determined. Nevertheless, one can think of various
scenarios where the recovery presented here would be desirable, such as
those in Sec. 5.7.
The proposed control algorithm in Sec. 5.4 includes some parameters
that should be chosen, but these have intuitive meanings. To avoid over-
shoot, the control gains in (5.8) should be constrained as described in
Sec. 5.4. This yields a negative real double pole for the system (5.8), lo-
cated in − k v /2, and a larger magnitude of the double pole corresponds to a
faster control system. Similarly, α and β should be constrained as described
in Chapter 3 and [Ijspeert et al., 2013]. The control system is intentionally
compliant to unforeseen disturbances, such as contact with a human, and
therefore no integral parts are used in the control loops that operate out-
side the internal robot controller. Introducing integral parts could also yield
overshoot in the position domain, which in the context of robotics might re-
sult in collision. The constant α e determines the extent of low-pass filtering
from position error to e, and k c determines how much the DMP evolution
should slow down in presence of position deviations.
Part II of the video in [Karlsson, 2017a] shows Setups 5.3 to 5.6, but the
original DMP formulation, without temporal coupling, was used instead.
Setups 5.5 and 5.6 were deliberately designed to require coordination be-
tween both arms, in order to highlight the necessity of temporal coupling
for DMPs. Without temporal coupling, the coordination was lost in presence
of any significant perturbation, and hence these tasks failed in Part II. In
fact, the coordination between all degrees of freedom of the robot, and not
only between the two arms, might be lost. For instance, only the left arm
moved in Setup 5.1, and yet the behavior was not satisfactory in Part II of
the video. It can also be noted that the coordination was retained in Part III,
despite perturbations, thanks to the temporal coupling.
In the experiments, the DMPs were defined in joint space. The im-
plementation in joint space allowed for the operator to specify the entire
configuration on each seven-degree-of-freedom robot arm, rather than just
the tool pose. The DMPs could also have been defined in Cartesian space,

78
 5.10 Conclusion

see for instance [Ijspeert et al., 2013; Ude et al., 2014]. However, whereas the
assumption of a contractible state space in Sec. 5.6 is correct for the robot
joint space and the position in Cartesian space, some care must be taken
regarding the orientation in Cartesian space, as discussed in, e.g., [Mayhew
et al., 2011]. Restrictions on the orientation must be invoked to guarantee
that it is contained in a contractible space. Finding such restrictions, that
are not unnecessarily conservative, is the subject of Chapter 6.

5.10 Conclusion
In this research, it was shown how perturbations of DMPs could be recovered
from, while preserving the characteristics of the original DMP framework
in the absence of significant perturbations. Feedforward control was used
to track the reference trajectory generated by a DMP. Feedback control
with moderate gains was used to suppress deviations. This design is the
first, to the best of the author’s knowledge, that takes the following aspects
into account. In the absence of significant disturbances, the position error
must be small enough, so that the dynamical system would not slow down
unnecessarily due to the temporal coupling. Very large controller gains
would result in small errors under ideal conditions, but in practice they
yield control signals that are not realizable. On the other hand, if the gains
are moderate and only feedback control is used, too large errors occur.
Feedforward enabled the controller to act even without significant error,
which in turn allowed for moderate controller gains. The proposed control
system was verified in simulations and experimentally. Videos of the exper-
iments are available in [Karlsson, 2016; Karlsson, 2017a]. Further, it was
shown mathematically that the proposed control system is globally expo-
nentially stable, which implies exponential convergence to a steady state in
which the goal position is reached by the actual robot position.

Appendix A. Passivity of Temporally Coupled Dynamical

Movement Primitives
We here consider a passivity property of temporally coupled DMPs. The
passivity formalism is described in, e.g., [Slotine and Li, 1991; Khalil, 2002].
The entire control system for temporally coupled DMPs in Sec. 5.6 is given

79
Chapter 5. Temporally Coupled Dynamical Movement Primitives in Rn

below for convenience.

ẏ − ẏc
   
y − yc
 − kp ( y − yc ) − kv ( ẏ − ẏc ) 
 ẏ − ẏ   
 c
   α e ( y − yc − e)

 e   
   α

d   x 
− x

 x = 
(5.25)
dt 
  
 τ a


  
   1
z
 y − ˆ  
 c 
  τa 
α
  
1 
z (β (ˆ − yc ) − z) + f ( x)
τa τa
Theorem 5.3
Consider the system defined by (5.25) for DMP operation. The virtual force
f ( x) can be seen as an external input to the system. The mapping from
f ( x) to z/τa is passive. 2
Proof. The system is passive if f (0) = 0, ξ˙ = 0 for ξ = f ( x) = 0, and there
exists a C 1 function V : R5n+1 → R, called a storage function, for which the
following criteria hold [Khalil, 2002].

V (0) = 0 (5.26a)
V (ξ ) ≥ 0, ∀ξ ,= 0 (5.26b)
V̇ (ξ ) ≤ f ( x) z/τa , ∀ξ , f ( x)
T
(5.26c)

Since x is a factor of f ( x) and the remaining part of f ( x) is bounded,

f (0) = 0, see Chapter 3 and [Ijspeert et al., 2013]. It can further be seen
that ξ˙ = 0 for ξ = f ( x) = 0.
The states y − yc , ẏ − ẏc , and e have the same dimension, which we
denote by n. Let In be the identity matrix, and 0n the zero matrix, each of
size n. The upper part of (5.25) is given by the following linear system.
    
y − yc 0n In 0n y − yc
d 
 ẏ − ẏc  = − kp In − k v In   ẏ − ẏc 
   
0n (5.27)
dt
e α e In 0n −α e In e

Denote by ξ¯ and Ā the state vector and the system matrix in (5.27), respec-
tively. Because k p = k2v /4 (see Sec. 5.4) the eigenvalues are given by
kv
λ1,...,2n = − (5.28a)
2
λ2n+1,...,3n = −α e (5.28b)

Since k v , α e > 0, the eigenvalues are strictly negative. Let Q be a symmetric,

positive definite matrix of size 3n. Since Ā has all eigenvalues strictly in

80
 5.A Passivity of Temporally Coupled Dynamical Movement Primitives

the left half-plane, it follows [Glad and Ljung, 2000] that there exists a
symmetric, positive definite matrix P solving the Lyapunov equation

P Ā + ĀT P = − Q (5.29)

Now, assume the following storage function V .

1 2 1 1
V (ξ ) = ξ¯T P ξ¯ + x + αβ ( yc − ˆ)T ( yc − ˆ) + zT z (5.30)
2 2 2

It holds that V (0) = 0. Since P is positive definite, any deviation from ξ¯ = 0

will give a positive contribution to V (ξ ). Similarly, the quadratic terms in
(5.30) guarantee that any deviation from ( x, yc − ˆ , z) = (0, 0, 0) gives a
positive contribution to V (ξ ). Therefore, V (ξ ) > 0, ∀ξ ,= 0. Further,

d
V (ξ ) = ξ¯T P ξ˙¯ + ξ˙¯T P ξ¯ + x ẋ
dt
d
+ αβ ( yc − ˆ)T ( yc − ˆ) + żT z
dt
αx 2 1
= ξ ( P Ā + Ā P )ξ −
¯ T T ¯ x + αβ ( yc − ˆ)T z
τa τa
3 4T
α 1
+ (β (ˆ − yc ) − z) + f ( x) z
τa τa
αx 2 α T
= −ξ¯T Qξ¯ − x − z z
τa τa
1 1
+ f ( x)T z ≤ f ( x)T z, ∀ξ , f ( x) (5.31)
τa τa

Hence, a passive mapping from f ( x) to z/τa can be concluded, according to

the passivity theory in [Khalil, 2002] and the criteria in (5.26). Thus, the
increase in stored energy is not larger than the externally added energy;
see also (5.32). 2

It is noteworthy that it is not straightforward to establish stability by

finding a Lyapunov function [Glad and Ljung, 2000] of ξ . A common attempt
is to investigate whether a weighted sum of the squared states qualifies as
a Lyapunov function. However, even though V (ξ ) consists of such a sum, it
does not qualify as a Lyapunov function. Furthermore, there is no obvious
way to modify V (ξ ) to form a Lyapunov function.
Since f ( x) represents a virtual force and z/τa represents a velocity, the
product of these can be interpreted as supplied power to the system. Time
integration up to a time t 1 yields the following mechanical interpretation

81
Chapter 5. Temporally Coupled Dynamical Movement Primitives in Rn

of (5.31).
Z t1
V (ξ ( t 1 )) = V (ξ (0)) + f ( x)T z/τa dt
| {z } | {z } 0 | {z }
Stored energy Energy at start Supplied power
Z t1
αx 2 α T
− ξ¯T Qξ¯ + x + z z dt
0 τa τa
| {z }
Dissipation power
Z t1
≤ V (ξ (0)) + f ( x)T z/τa dt (5.32)
| {z } 0 | {z }
Energy at start Supplied power

Here, integrating the supplied power with respect to time yields total work
done by f ( x). Hence, the increase in stored energy is not larger than the
externally added energy, which indicates passivity.
The global exponential stability established in Sec. 5.6 does not rely on
the passivity shown in this appendix. However, this appendix may contribute
with some intuition of the role of f ( x), and of why finding a Lyapunov
function of ξ is not trivial.

82
6
Temporally Coupled
Dynamical Movement
Primitives in 3D Orientation
Space

6.1 Introduction
In this chapter, we extend the results in Chapter 5 to support orienta-
tion in Cartesian space. Higher levels of robot control typically operate in
Cartesian space, for instance to control the pose of a robot end-effector or
an unmanned aerial vehicle. However, control of orientation in Cartesian
space is fundamentally limited: The rotation group SO(3) is not contractible
(defined in Sec. 6.4), and only globally contractible state spaces support
continuous and globally asymptotically stable feedback control systems. In
this chapter, a control system for temporally coupled dynamical movement
primitives (DMPs) in Cartesian space is designed. This is an extension of
the control law in Chapter 5 to support 3D orientation, and equivalently an
extension of [Ude et al., 2014] to include temporal coupling. Unit quater-
nions are used to represent orientations, and it is shown that the unit
quaternion set minus one point is contractible. Finally, the efficacy of the
control system is verified experimentally on an industrial robot.

6.2 Previous Research

The fundamentals of DMPs have been described in [Ijspeert et al., 2013], and
earlier versions have been introduced in [Schaal et al., 2000; Ijspeert et al.,
2002; Ijspeert et al., 2003]. The concept of temporal coupling for DMPs was
introduced in [Ijspeert et al., 2013], and extended in Chapter 5. However,

83
Chapter 6. Temporally Coupled DMPs in 3D Orientation Space

these previous DMP formulations are applicable only if the robot state space
is Euclidean. As a consequence, previous applications and extensions of
DMPs typically rely on Euclidean state spaces; see, e.g., [Prada et al., 2014;
Karlsson et al., 2017b; Talignani Landi et al., 2018a; Papageorgiou et al.,
2018; Yang et al., 2018]. In [Perk and Slotine, 2006; Wensing and Slotine,
2017], a helicopter and a legged robot were controlled using DMPs, and
global exponential stability was shown assuming Euclidean state spaces.
The assumption of a Euclidean state space is correct for position in
Cartesian space and joint space, but not for orientation in Cartesian space.
This circumstance motivated the DMP formulation proposed in [Ude et
al., 2014], which supports Cartesian space including orientations. It was
applied for robot programming by demonstration in [Nemec et al., 2018].
This chapter extends [Ude et al., 2014] by including temporal coupling
and addressing the stability properties of the proposed control algorithm.

6.3 Problem Formulation

In this chapter, we address the question of whether the control algorithm
in Chapter 5 could be extended also to incorporate orientations. Because a
contractible state space is necessary for design and analysis of a continuous
globally asymptotically stable control law (see Sec. 6.4), we first investigate
the contractibility properties of the quaternion set used to represent orien-
tations. Thereafter, we propose a control algorithm for temporally coupled
DMPs in Cartesian space including orientations. Finally, we address the
question of whether the proposed algorithm is exponentially stable. This is
addressed mathematically as well as experimentally.

6.4 The Unit Quaternion Set Minus One Single Point is

Contractible
Global exponential stability for temporally coupled DMPs was shown in
Chapter 5, but under the assumption that the DMP state space was con-
tractible. This is true for the spaces of joint positions as well as Cartesian
positions, because these are Euclidean spaces. However, global contractibil-
ity can not directly be assumed for orientations in Cartesian space. As
noted in [Mayhew et al., 2011], the rotation group SO(3) is not contractible,
and therefore it is not possible for any continuous state-feedback control
law to yield a globally asymptotically stable equilibrium point in SO(3)
[Koditschek, 1988; Bhat and Bernstein, 2000]. A space is contractible if
and only if it is homotopy equivalent to a one-point space [Crossley, 2006],
which intuitively means that the space can be deformed continuously to a

84
 6.4 The Unit Quaternion Set Minus One Single Point is Contractible

single point; see, e.g., [Crossley, 2006] for a definition of homotopy equiv-
alence. Contractibility is necessary for applying the contraction theory of
[Lohmiller and Slotine, 1998], as done in Chapter 5. In this chapter, unit
quaternions are used to parameterize SO(3). Similarly to SO(3), the unit
quaternion set, H, is not contractible. In this section, however, it is shown
that it is sufficient to remove one point from H to yield a contractible space.
Table 6.1 lists some of the notation used in this chapter.

Preliminary Topology
The fundamentals of mathematical topology and set theory are described
in, e.g., [Hatcher, 2002; Levy, 2002; Crossley, 2006; Schwarz, 2013]. We will
use the fact that homeomorphism [Crossley, 2006] is a stronger relation
than homotopy equivalence.

Lemma 6.1
If two spaces X and Y are homeomorphic, then they are homotopy equiva-
lent. 2

Proof. See Lemma 6.11 in [Crossley, 2006]. 2

Lemma 6.2
Assume that X ∼= Y, with a homeomorphism f1 : X → Y. Then X minus a
point p ∈ X, denoted X \ p, is homeomorphic to Y \ f1 (p). 2

Proof. Consider the function f2 : X \ p → Y \ f1 (p), and let f2 ( x) =

f1 ( x) ∀ x ∈ X \ p. It can be seen that f2 is a restriction of f1 . Since a
restriction of a homeomorphism is also a homeomorphism [Lehner, 1964],
f2 is a homeomorphism, and hence X \ p ∼ = Y \ f1 (p). 2

We will also use that homeomorphism preserves contractibility.

Lemma 6.3
If X ∼
= Y, and X is contractible, then Y is also contractible. 2

Proof. Since X ∼= Y, they are homotopy equivalent according to Lemma 6.1.

In turn, X is contractible and therefore homotopy equivalent to a one-point
space. Hence, Y is also homotopy equivalent to a one-point space, and
therefore contractible. 2

The Largest Contractible Subset of the Unit Quaternion Set

First, it will be shown that the unit sphere Sn (see Definition 6.1) minus a
point is contractible. This will then be applied to H, which is homeomorphic
to S3 [LaValle, 2006].

85
Chapter 6. Temporally Coupled DMPs in 3D Orientation Space

Table 6.1 Notation used in this chapter. All quaternions represent orienta-
tions and are therefore unit quaternions, i.e., quaternions of norm one. For
a unit quaternion, its inverse is equal to its conjugate.

Notation Description
H Unit quaternion set
Sn ∈ Rn+1 Unit sphere of dimension n
y ∈ R3 Actual robot position
ˆ ∈ R3 Goal position
yc ∈ R3 Coupled robot position
qa ∈H Actual robot orientation
qˆ ∈H Goal orientation
qc ∈H Coupled robot orientation
q0 ∈H Initial robot orientation
ωa ∈ R3 Actual angular velocity
ωc ∈ R3 Coupled angular velocity
z, ω z ∈ R3 DMP states
h Orientation difference space
dc ˆ ∈h Difference between q c and q ˆ
α, β , k v , k p ∈ R+ Constant control gains
τ ∈ R+ Time parameter
τa ∈ R+ Adaptive time parameter
x ∈ R+ Phase variable
α x , αe , kc ∈ R+ Positive constants
f ( x) ∈ R6 Learnable virtual forcing term
fp ( x), fo ( x) ∈ R3 Position and orientation components
Nb ∈ Z+ Number of basis functions
m ∈ Z+ Dimension of Cartesian configuration
Ψ j ( x) ∈ R6 The j:th basis function vector
wj ∈ R6 The j:th weight vector
e ∈ R3 $ h Low-pass filtered pose error
ep ∈ R3 Position component of e
eo ∈h Orientation component of e
ÿr , ω̇ r ∈ R3 Reference robot acceleration
ξ ∈ R22 $ h3 DMP state vector
qξ i q ∈ R+ 2-norm of DMP state i
q̄ ∈H Inverse of quaternion q (same as conjugate)
h ∈ R+ Sample period (4 ms in experiments)
) Homotopy equivalence
∼
= Homeomorphic relation

86
 6.4 The Unit Quaternion Set Minus One Single Point is Contractible

Definition 6.1
Let n be a non-negative integer. The unit sphere with dimension n is defined
as ) *
Sn = p ∈ Rn+1 p qpq2 = 1 (6.1)
2

Theorem 6.1
Let n be a non-negative integer. The unit sphere Sn minus a point p ∈ Sn ,
denoted Sn \ p, is contractible. 2

Proof. Consider first the case n ≥ 1. There exists a mapping from Sn \ p

to Rn called stereographic projection from p, which is a homeomorphism.
Thus, Sn \ p ∼= Rn [Huggett and Jordan, 2009; Schwarz, 2013]. See Fig. 6.1
for a visualization of these spaces. Since Rn is a Euclidean space it is
contractible, and it follows from Lemma 6.3 that Sn \ p is also contractible.
Consider now the case n = 0. The sphere S0 consists of the pair of points
{−1, 1} according to Definition 6.1. Thus S0 \ p consists of one point only,
and homotopy equivalence with a one-point space is trivial. Hence S0 \ p is
contractible. 2
Remark 6.1
Albeit we consider unit spheres in this chapter, it is not necessary to assume
radius 1 in Theorem 6.1. Further, it is arbitrary which point p ∈ Sn to
remove. 2

The main result of this section is concluded by the following theorem.

Theorem 6.2
The set of unit quaternions H minus a point q̃ ∈ H, denoted H \ q̃, is
contractible. 2

Proof. The set H is homeomorphic to S3 [LaValle, 2006]. Therefore H \ q̃ ∼

=
S3 \ p for some point p ∈ S3 , according to Lemma 6.2. Theorem 6.1 with
n = 3 yields that S3 \ p is contractible, and because of the homeomorphic
relation, Lemma 6.3 yields that H \ q̃ is also contractible. 2

It is noteworthy that the contractible subset H \ q̃ is the largest possible

subset of H, because one point is the smallest possible subset to remove.
Hence, it is guaranteed that no unnecessary restriction is made in Theo-
rem 6.2, though there are other, more limited, subsets of H that are also
contractible. Sometimes only half of H, for instance the upper half of the
quaternion hypersphere, is used to represent orientations. However, instead
of continuous transitions between the half spheres this results in disconti-
nuities within the upper half sphere [LaValle, 2006]. In the context of DMPs
and automatic control, such discontinuities would cause severe obstructions,

87
Chapter 6. Temporally Coupled DMPs in 3D Orientation Space

1 1

0.5 S0 \ p 0.5 R0 = p̂

0 ∼
= 0

−0.5 −0.5

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
1 1

0.5 S1 \ p 0.5 R1 ) p̂

0 ∼
= 0

−0.5 −0.5

−1 −1
−1 −0.5 0 0.5 1 −1 −0.5 0 0.5 1
2
S \p
1 1
R2 ) p̂

0 = 0
∼

−1 −1
1 1
1 1
0 0 0 0
−1 −1 −1 −1

Figure 6.1 Visualization of Sn \p (left) and Rn (right) for n = 0, 1, 2. The red

cross marks a point p removed from the unit sphere. Each space to the left
is homeomorphic to the corresponding space to the right, i.e., Sn \ p ∼= Rn . In
n
turn, R is homotopy equivalent to a point (for instance p̂ marked by a violet
dot in each plot to the right) and therefore Sn \ p is contractible according to
Lemma 6.3. Higher dimensions are difficult to visualize, and therefore S2 is
commonly used to visualize parts of the quaternion set, as done in Fig. 6.9.

88
 6.5 Method

which motivates the search for the largest possible contractible subset of H.
One of the experiments (Setup 6.3 in Sec. 6.6) provides an example of when
both half spheres are necessary for a continuous representation of the robot
orientation.

6.5 Method
In this section, we augment the controller in Chapter 5 to incorporate
orientation in Cartesian space. The resulting algorithm can also be seen as
a temporally coupled version of the Cartesian DMPs proposed in [Ude et al.,
2014]. In this chapter, we consider the control of one end-effector to limit
the notation, but it would be straightforward to include several end-effectors
and robots in the proposed control algorithm. The pose in Cartesian space
consists of position and orientation. The position control in this chapter is
the same as described in Chapter 5, except that it is also affected by the
orientation through the shared time parameter τa in this chapter.
Similar to the approaches in [Ude, 1999; Ude et al., 2014], we define a
difference between two quaternions, q 1 and q 2 , as

d (q 1 q̄ 2 ) = 2 · Im[log(q 1 q̄ 2 )] ∈ h (6.2)

where h is the orientation difference space, defined as the image of d ,

and Im denotes the operator that extracts the imaginary quaternion part,
assuming for now that q 1 q̄ 2 ,= (−1, 0, 0, 0). This is elaborated on in Sec. 6.8.
Further, we will use a shorter notation, so that for instance

d c ˆ = d (q c q̄ ˆ ) = 2 · Im[log(q c q̄ ˆ )] (6.3)

represents the difference between coupled and goal orientations. This map-
ping preserves the contractibility concluded in Sec. 6.4, as established by
the following theorem.
Theorem 6.3
The orientation difference space h is contractible. 2

Proof. The mapping

d : H \ (−1, 0, 0, 0) → h (6.4)

has the properties necessary to qualify as a homeomorphism. It is one-to-

one [Ude, 1999] and onto, continuous, and its inverse (division by 2 followed
by the exponential map) is also continuous.
Further, its domain H \ (−1, 0, 0, 0) is contractible (see Theorem 6.2),
and therefore its image h is contractible (see Lemma 6.3). 2

89
Chapter 6. Temporally Coupled DMPs in 3D Orientation Space

Using the function d , a coupled DMP pose trajectory is modeled by the

dynamical system

τa ż = α(β (ˆ − yc ) − z) + fp ( x) (6.5a)
τa ẏc = z (6.5b)
τaω̇ z = α(β (− d c ˆ ) − ω z ) + fo ( x) (6.5c)
τaω c = ω z (6.5d)

Here, x is a phase variable that evolves as

τa ẋ = − α x x (6.6)

Further, fo ( x) is a virtual forcing term in the orientation domain, and each

element of fo ( x), denoted foi ( x), is given by
P Nb
j=1 Ψi, j ( x)wi, j
foi ( x) = P Nb x · d i (q ˆ q̄ 0 ) (6.7)
j =1 Ψi, j ( x)

where each basis function, Ψi, j ( x), is determined as

3 4
1
Ψi, j ( x) = exp − ( x − c i, j )2 (6.8)
2σi,2j

Here, σ and c denote the width and center of each basis function, re-
spectively. The forcing term fp ( x) is determined accordingly, see Chapter 3.
Further, the parameters of f ( x) can be determined based on a demonstrated
trajectory by means of locally weighted regression [Atkeson et al., 1997], as
described in [Ijspeert et al., 2013] and Chapter 3.
All dimensions of the robot pose are temporally coupled through the
shared adaptive time parameter τa . Denote by y the actual position of the
robot, and by q a the actual orientation. The adaptive time parameter τa
is determined based on the low-pass filtered difference between the actual
and coupled poses as follows.

ė p = α e ( y − yc − e p ) (6.9a)
ė o = α e ( d ac − e o ) (6.9b)
e= [ eTp eTo ]T (6.9c)
τa = τ (1 + k c e e) T
(6.9d)

In (6.9c), the contributions from the position and orientation errors are
weighted equally to limit the notation, though this is not necessary in an
implementation. Further, we assume that the length unit for the position

90
 6.5 Method

error has been compensated for, so that e is a dimensionless quantity.

Moreover, the controller below is used to drive y to yc , and q a to q c .

ÿr = k p ( yc − y) + k v ( ẏc − ẏ) + ÿc (6.10a)

ω̇ r = − k p d ac − k v (ω a − ω c ) + ω̇ c (6.10b)

This can be seen as a pose PD controller (see Sec. 6.8) together with the
feedforward terms ÿc and ω̇ c . Here, ÿr and ω̇ r denote reference accelerations
sent to the internal controller of the robot, after conversion to joint values
using the robot Jacobian [Spong et al., 2006]. We let k p = k2v /4, so that
(6.10) represents a critically damped control loop (see Chapter 5). Similarly,
β = α/4 (see Chapter 3 and [Ijspeert et al., 2013]). The control system
is schematically visualized in Fig. 6.2. We model the ’Robot’ block as a
double integrator, so that ÿ = ÿr and ω̇ a = ω̇ r , as justified in Chapter 5 for
accelerations with moderate magnitudes and time derivatives. In summary,
the proposed control system is given by

ÿ = − k p ( y − yc ) − k v ( ẏ − ẏc ) + ÿc (6.11a)

ω̇ a = − k p d ac − k v (ω a − ω c ) + ω̇ c (6.11b)
3è éT 4
T
ė = α e [ y − yc ] d ac − e
T
(6.11c)

τa = τ (1 + k c eT e) (6.11d)
τa ẋ = −α x x (6.11e)
τa ẏc = z (6.11f )
τa ż = α(β (ˆ − yc ) − z) + fp ( x) (6.11g)
τaω c = ω z (6.11h)
τaω̇ z = α(β (− d c ˆ ) − ω z ) + fo ( x) (6.11i)

We introduce a state vector ξ as

 
y − yc
 ẏ − ẏc 
 
 
 dac 
 
ω a − ω c 
 
 
 e 
ξ =  ∈ R22 $ h3
 
(6.12)

 x 

 y −ˆ 
 c 
 

 z 

 dc ˆ 
 

wz

91
Chapter 6. Temporally Coupled DMPs in 3D Orientation Space

5 6
Feedforward ÿc
ω̇ c 5 6
5 6 5 6 5 6
ˆ yc ÿr y
qˆ qc ω̇ r qa
PD P
DMP Robot
controller

Feedback

Figure 6.2 The control structure for temporally coupled Cartesian DMPs.
The block denoted ’Robot’ includes the internal controller of the robot, to-
gether with transformations between Cartesian and joint space for low-level
control. The ’DMP’ block corresponds to the computations in (6.5) to (6.9).
The PD controller and the feedforward terms are specified in (6.10). This
forms a cascade controller, with the DMP as outer controller and the PD as
the inner.

Since R22 $h3 is a product of contractible spaces (see Theorem 6.3), and such
a product is itself contractible [Kahn, 1975], the state space is contractible.
In Sec. 6.A, it is shown that an approximation of the proposed control system
in discrete time is exponentially stable.

6.6 Experiments
The control law in Sec. 6.5 was implemented in the Julia programming lan-
guage [Bezanson et al., 2014], to control an ABB YuMi [ABB Robotics, 2018]
robot. The Julia program communicated with the internal robot controller
through a research-interface version [Bagge Carlson and Haage, 2017] of
Externally Guided Motion (EGM) [ABB Robotics, 2019c] at a sampling rate
of 250 Hz.
Three different setups were used to investigate the behavior of the con-
troller. As preparation for each setup, a temporally coupled Cartesian DMP
had been determined from a demonstration by means of lead-through pro-
gramming, which was available in the YuMi product by default. In each
trial, the temporally coupled DMP was executed while the magnitudes of
the states in (6.12) were logged.
Perturbations were introduced by physical contact with a human. This
was enabled by estimating joint torques induced by the contact, and map-
ping these to Cartesian contact forces and torques using the robot Jacobian.
A corresponding acceleration was then added to the reference acceleration
as a load disturbance. However, we emphasize that this chapter is not
focused on how to generate the perturbations themselves. Instead, that
functionality was used only as an example of unforeseen deviations, and to

92
 6.6 Experiments

(a) (b)
Figure 6.3 Photographs of a trial of Setup 6.1. The robot was initially
released from the pose in (a), with an offset to the goal pose. In (b), the goal
pose was reached.

investigate the stability properties of the proposed control algorithm.

A video of the experimental arrangement is available in [Karlsson, 2019].
The setups were as follows.

Setup 6.1 This setup is visualized in Fig. 6.3. Prior to the experiment, a
test DMP that did not perform any particular task was executed, and the
robot then converged to the goal pose, i.e., to y = yc = ˆ and d ac = d c ˆ = 0.
Thereafter, the operator pushed the end-effector, so that the actual pose
deviated from the coupled and goal poses. The experiment was initialized
when the operator released the robot arm. The purpose of this procedure
was to examine the stability of the subsystem in (6.11a) to (6.11c). A total
of 100 perturbations were conducted.

Setup 6.2 See Fig. 6.4. The task of the robot was to reach a work object
(in this case a gore-tex graft used in cardiac and vascular surgery) from
its home position. A DMP defined for this purpose was executed, and the
operator introduced two perturbations during the robot movement. The
purpose of this setup was to investigate the stability of the entire control
system in (6.11). A total of 10 trials were conducted.

Setup 6.3 See Fig. 6.5. The task of the robot was to hand over the work
object from its right arm to its left. The movement was specifically designed
to require an end-effector rotation angle of more than π , thus requiring both
the upper and the lower halves of the quaternion hypersphere (see Fig. 6.9),
and not only one of the halves which is sometimes used [LaValle, 2006].
Such movements motivate the search for the largest possible contractible
subset of H in Sec. 6.4. Similar to Setup 6.2, the purpose was to investigate
the stability of (6.11), and 10 trials were conducted.

93
Chapter 6. Temporally Coupled DMPs in 3D Orientation Space

(a) (b)

(c) (d)
Figure 6.4 Photographs of a trial of Setup 6.2. The DMP was executed
from the home position (a), and was perturbed twice on its way toward the
goal (b). It recovered from these perturbations (c), and reached the goal at
the work object (d).

6.7 Results
Figures 6.6 to 6.9 display data from the experiments. Figure 6.6 shows the
magnitude of the states during a trial of Setup 6.1, and it can be seen
that each state converged to 0 after the robot had been released. This is
consistent with Theorem 6.4 in Sec. 6.A. Similarly, Figs. 6.7 and 6.8 show
data from Setups 6.2 and 6.3 respectively, and it can be seen that the robot
recovered from each of the perturbations. Further, each state subsequently
converged to 0. This is consistent with Theorem 6.5 in Sec. 6.A. All trials in
a given setup gave similar results. Further, these results suggest that the
control system (6.11) is exponentially stable.
Figure 6.9 shows orientation data from Setup 6.2 (left) and Setup 6.3
(right). The upper plots show quaternions for the demonstrated paths, q d ,
determined using lead-through programming prior to the experimental tri-
als, relative to the goal quaternions q ˆ . The middle plots show coupled
orientations q c relative to q ˆ . It can be seen that the paths of q d and q c
were similar for each of the setups, which was expected given a sufficient

94
 6.8 Discussion

(a) (b)

(c) (d)
Figure 6.5 Photographs of a trial of Setup 6.3. The robot started its move-
ment from the configuration in (a). The end-effector was rotated as indicated
by the red arrows, which resulted in a rotation larger than π from start
to goal. The robot was perturbed twice by the operator (b), recovered and
continued its movement (c), and accomplished the handover (d).

number of DMP basis functions. The perturbations can be seen in the bot-
tom plots, which show q a relative to q c . Though q a q̄ c was very close to the
identity quaternion for most of the time, it deviated significantly twice per
trial as a result of the perturbations. Setup 6.3 is an example of a movement
where it would not be possible to restrict the quaternions to the upper half
sphere, without introducing discontinuities. This is shown in Figure 6.9, as
quaternions were present not only on the upper half sphere, but also on the
lower, for Setup 6.3.

6.8 Discussion
In each of the experiments, the robot recovered from the perturbations
and subsequently reached the goal pose, which was the desired behavior.
Because each state converged to 0, we conclude that the experimental results
verify Theorem 6.4 and Theorem 6.5; see Sec. 6.A. Further, the behavior

95
Chapter 6. Temporally Coupled DMPs in 3D Orientation Space

1.5 q y − yc q [m]
q ẏ − ẏc q [m/s]

qξ i q
1 q dac q [rad]
qω a − ω c q [rad/s]
0.5
q eq [1]
0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5
Time [s]

Figure 6.6 Data from a trial of Setup 6.1. The notation q · q represents
the 2-norm, and the unit symbol [1] indicates dimensionless quantity. The
experiment was initialized with some position error y − yc and orientation
error d ac . The operator released the robot at t = 0. It can be seen that each
state converged to 0.

q y − yc q [m]
1
q ẏ − ẏc q [m/s]
qξ i q

q dac q [rad]
0.5 qω a − ω c q [rad/s]
q eq [1]
0
0 5 10 15 20 25 30
2
x [1]
1.5 q yc − ˆq [m]
qξ i q

1 q zq [m/s]
q dc ˆ q [rad]
0.5 qω z q [rad/s]
0
0 5 10 15 20 25 30
Time [s]

Figure 6.7 Data from a trial of Setup 6.2. Consider first the upper plot.
The two perturbations are clearly visible, and these were recovered from as
the states converged to 0. In the lower plot, it can be seen that the time
evolution of the states was slowed down in the presence of perturbations. It
can further be seen that each of the states converged to 0.

corresponds to that in Chapter 5, except that orientations in Cartesian

space are now supported. Most of the discussion in Chapter 5 is therefore
valid also for these results, and is not repeated here.
There are two reasons for discretizing the control system in Sec. 6.A.
Firstly, it shows that the system can be integrated without violating the unit

96
 6.8 Discussion

1.5
q y − yc q [m]
1 q ẏ − ẏc q [m/s]
qξ i q q dac q [rad]
0.5 qω a − ω c q [rad/s]
q eq [1]
0
0 5 10 15 20 25 30 35
4
x [1]
3 q yc − ˆq [m]
qξ i q

2 q zq [m/s]
q dc ˆ q [rad]
1 qω z q [rad/s]
0
0 5 10 15 20 25 30 35
Time [s]

Figure 6.8 Data from a trial of Setup 6.3. The organization is the same as
in Fig. 6.7, and similar conclusions can be drawn. In addition, the required
rotation angle from start to goal was larger than π in this setup, which
corresponds to q d c ˆ q being larger than π initially.

length of the quaternions, which is sometimes problematic. This is possible

because the quaternion difference in (6.2) can be updated by, for instance,
the forward Euler method, without violating quaternion unit length. Ap-
plying such updates directly on quaternions would have resulted in loss of
unit length. This is particularly important for q c , which is not updated by
measurements but by simulated integration also during DMP operation on
real robots. Secondly, the discretization yields conditions on the sampling
period h, which should be short enough in relation to the control gains to
retain exponential stability. These conditions can easily be extracted from
the eigenvalues in (6.20), (6.22) and (6.26), which should be inside the unit
circle. Further, these conditions are easy to satisfy with reasonable gains
and sampling periods, and they were therefore assumed to be satisfied in
the proofs in this chapter.
It should be emphasized that the control system analyzed in Sec. 6.A
is a simplified version of the proposed control system. The view of the
feedback control in (6.10b) as PD control and the representation of the
control system in Sec. 6.A rely on the approximation that the time derivative
of the quaternion difference in (6.2) is equal to the angular velocity, i.e.,
ḋ (q) ( ω . This is exactly true only when the rotation takes place around
a fixed axis [Boyle, 2017]. For the sake of proving stability, however, the

97
Chapter 6. Temporally Coupled DMPs in 3D Orientation Space

H q d q̄ ˆ
Start Identity quaternion

1 1
Re[q]

0 0

−1 −1
−1 1 −1 −1
0 0 0 0
1 −1 1 1

1 1
Re[q]

0 0

−1 −1
−1 q c q̄ ˆ 1 −1 −1
0 0 0 0
1 −1 1 1
0.5 0.5

0 0

−0.5 q a q̄ c −0.5
−0.5 0 0.5 −0.5 0 0.5

Figure 6.9 Orientation data from Setup 6.2 (left) and Setup 6.3 (right).
Quaternions have been projected onto S2 for the purpose of visualization.
Vertical axes represent quaternion real parts, and horizontal axes represent
the first two imaginary elements with magnitudes adjusted to yield unit
length of the resulting projection. The bottom plots show the quaternion set
seen from above, and hence their real axes are directed out from the figure.

98
 6.8 Discussion

approximation was only useful for showing that q a is driven to q c , and q c to

q ˆ after the virtual forcing term in practice has vanished. Hence, detours
due to the virtual forcing term, which indeed might violate the assumption
of a fixed rotation axis, are not as problematic as it might seem. A formal
proof that the exact proposed control system is exponentially stable would
enhance the contribution of this chapter, but remains as future research.
Nevertheless, it has now been shown that the topology of h does not prohibit
a globally exponentially stable control system.
The magnitude of the difference between two quaternions, q d (q 1 q̄ 2 )q,
corresponds to the length of a geodesic curve connecting q 1 and q 2 [Ude,
1999]. This results in proper scaling between orientation difference and
angular velocity in the DMP control algorithm, as explained in [Ude et al.,
2014]. This is the reason why the quaternion difference in (6.2) was used in
[Ude et al., 2014] and in this chapter.
In Sec. 6.4, the largest possible contractible subset of H was found as H\
q̃. Hence, it is not necessary to remove a large proportion of the quaternion
set, which is sometimes done. For instance, sometimes the lower half of the
quaternion hypersphere is removed [LaValle, 2006], which is unnecessarily
limiting. The results from Setup 6.3 show that this proposed method works
also when it is necessary to use both half spheres, see Fig. 6.9. In Sec. 6.5,
the removed point q̃ was chosen as (−1, 0, 0, 0), which corresponds to a full
2π rotation from the identity quaternion. A natural question is therefore
how to handle the case where (−1, 0, 0, 0) is visited by q a q̄ c or q c q̄ ˆ . In
theory, almost any control signal could be used to move the orientations
away from this point, and in practice a single point would never be visited
because it is infinitely small. However, in practice some care should be
taken in a small region around (−1, 0, 0, 0), because of possible numerical
difficulties and rapidly changing control signals.
In this chapter, the same control gains were used in the position domain
as in the orientation domain. This was done in order to limit the notation,
but is not actually required.
The functionality presented in this chapter was demonstrated at the
final review of the SARAFun project, see Chapter 2 and [SARAFun, 2019].
It was used for motion control in combination with the obstacle avoidance
in [Stavridis et al., 2017] on the main demonstration platform, and in
combination with the friction estimation in [Bagge Carlson, 2019] on the
separate demonstration platform.
An interesting direction of future research is to use the proposed con-
troller to warm-start reinforcement learning for robotic manipulation. Re-
inforcement learning with earlier DMP versions has been investigated in,
e.g., [Li et al., 2018].

99
Chapter 6. Temporally Coupled DMPs in 3D Orientation Space

6.9 Conclusion
In this chapter, it was first shown that the unit quaternion set minus one
point is contractible. A contractible state space allows for continuous and
exponentially stable control systems. Thereafter, a control algorithm for
DMPs with temporal coupling in Cartesian space was designed. The pro-
posed DMP functionality was verified experimentally on an industrial robot.
Further, exponential stability was proven mathematically for an approxi-
mation of the proposed control system. A video that shows the experiments
is available in [Karlsson, 2019].

Appendix A. Stability Analysis

In this appendix, an approximation of the proposed control system is shown
to be exponentially stable. The entire control system in Sec. 6.5 is given by

ÿ = − k p ( y − yc ) − k v ( ẏ − ẏc ) + ÿc (6.13a)

ω̇ a = − k p d ac − k v (ω a − ω c ) + ω̇ c (6.13b)
3è éT 4
T
ė = α e [ y − yc ] d Tac − e (6.13c)

τa = τ (1 + k c eT e) (6.13d)
τa ẋ = −α x x (6.13e)
τa ẏc = z (6.13f )
τa ż = α(β (ˆ − yc ) − z) + fp ( x) (6.13g)
τaω c = ω z (6.13h)
τaω̇ z = α(β (− d c ˆ ) − ω z ) + fo ( x) (6.13i)

The state vector is given by

 
y − yc
 ẏ − ẏc 
 
 
 dac 
 
ω a − ω c 
 
 
 e 
ξ =  ∈ R22 $ h3
 
(6.14)

 x 

 y −ˆ 
 c 
 

 z 

 dc ˆ 
 

wz

100
 6.A Stability Analysis

Because R22 $ h3 is a product of contractible spaces (see Theorem 6.3),

and such a product is itself contractible [Kahn, 1975], we note that the
state space in (6.14) is contractible. This is a necessary property for the
contraction theory applied in this appendix.
In the following, we will consider a representation of the system in
discrete time, based on the approximation that the time derivative of the
quaternion difference in (6.2) is equal to the angular velocity, i.e., ḋ (q) ( ω .
This is elaborated on in Sec. 6.8. Using the forward Euler method, a discrete
state-space representation of the system is given by
 
y − yc + h( ẏ − ẏc )
 

 − hk p ( y − yc ) + (1 − kv h)( ẏ − ẏc ) 
d ac + h(ω a − ω c )
 
 
 
− hk p dac + (1 − hk v )(ω a − ω c ) 
 

 

 3è éT 4 

T
e + hα e [ y − yc ] d ac − e 
T
 

 
3 4
 
 
 αx 
ξ ( t + h) = 
 1−h x  ( t)

(6.15)
 τa 
 
 h 
 yc − ˆ + z 

 τa 

h
 
 
 z+ [α(β (ˆ − yc ) − z) + fp ( x)] 

 τa 

 h 
dc ˆ + ω z
 
τa
 
 
 
 h 
wz + [α(β (− dc ˆ ) − ω z ) + fo ( x)]
τa
Here, h represents the sampling period, and t is discrete time so that
t = kh for some integer k. Because of the adaptive time scale from τa , and
the forcing terms fp ( x) and fo ( x), the system is nonlinear. To structure the
stability analysis, we utilize that (6.15) is a hierarchical system, and analyze
one subsystem at a time. A similar strategy was applied in Chapter 5,
though for a purely Euclidean state space. In this context, the following
relation is useful.
Proposition 6.1
If x1 ( t + h) = ˆ 1 ( x1 ( t)) is contracting, and x2 ( t + h) = ˆ 2 ( x1 ( t), x2 ( t)) is
contracting for each fixed x1 , then the hierarchy
3 4 3 4
x1 ( t + h) ˆ 1 ( x1 ( t))
= (6.16)
x2 ( t + h) ˆ 2 ( x1 ( t), x2 ( t))

101
Chapter 6. Temporally Coupled DMPs in 3D Orientation Space

is contracting. 2

This is a discrete version of Proposition 2 in [Wensing and Slotine, 2017]

applied to autonomous systems.
In the following, we assume that h is sufficiently short in relation to the
control gains, so that the discretization itself does not cause instability; see
Sec. 6.8.
The first five lines in (6.15) do not depend on the rest of the system, and
can be analyzed separately as the following linear subsystem. Denote by ξ 1
the state vector of the subsystem, so that

ξ 1 ( t + h) = Φ1ξ 1 ( t) (6.17)

where ξ 1 is given by
y − yc
 
 ẏ − ẏc 
 
 dac 
 
ξ1 = 
  (6.18)
ω a − ω c 

 
 ep 
eo
Further, denote by I the identity matrix and by 0 the zero matrix, of due
sizes. The system matrix is given by

I hI
 
0 0 0 0
− hk p I (1 − hk v ) I 0 0 0 0 
 
I hI
 
 0 0 0 0 
Φ1 =  
 0
 0 − hk p I (1 − hk v ) I 0 0 

 
 hα e I 0 0 0 (1 − hα e ) I 0 
0 0 hα e I 0 0 (1 − hα e ) I
(6.19)

Theorem 6.4
The dynamical system defined by (6.17) for Cartesian DMP operation is a
contraction and has ξ 1 = 0 as fixed point. 2

Proof. If ξ 1 ( t) = 0, it can be seen in (6.17) that also ξ 1 ( t + h) = 0, and

ξ 1 = 0 is therefore a fixed point. The eigenvalues of Φ1 are given by

hk v
λ1,...,2m = 1 − (6.20a)
2
λ2m+1,...,3m = 1 − hα e (6.20b)

102
 6.A Stability Analysis

Each eigenvalue is strictly within the unit circle since h, k v , α e > 0, and the
system is therefore globally exponentially stable. Because it is also a linear
time-invariant system, contraction can be concluded. 2

Next, we will show contraction for a larger subsystem of (6.15).

Proposition 6.2
The system defined by

ξ 1 ( t + h) = Φ1ξ 1 ( t) (6.21a)
αx
x( t + h) = (1 − h ) x( t) (6.21b)
τa
is contracting and has the origin as fixed point. 2

Proof. The origin is a fixed point since (ξ 1 ( t), x( t)) = (0, 0) [ (ξ 1 ( t +

h), x( t + h)) = (0, 0). Further, it is known from Theorem 6.4 that (6.21a)
is contracting with the origin as fixed point. Hence, for the fixed point of
(6.21a), it holds that τa = τ . For the fixed point of (6.21a), x therefore
evolves as the linear system
αx
x( t + h) = (1 − h ) x( t) (6.22)
τ
which is a contraction because the eigenvalue is strictly within the unit
circle (since h, α x , τ > 0). Because (6.21a) is contracting and (6.21b) is
contracting for the fixed point of (6.21a), it follows from Proposition 6.1 that
the hierarchy (6.21) is also contracting. 2

Finally we address the stability of the whole DMP system in (6.15).

Theorem 6.5
The system in (6.15) for Cartesian DMP operation has the origin as globally
exponentially stable equilibrium point. 2
Proof. The system in (6.15) is a hierarchy from (6.21) to the rest of (6.15)
(through τa and x). It is known from Proposition 6.2 that (6.21) is contracting
with the origin as fixed point. For the fixed point of (6.21), we specifically
have τa = τ and x = 0, and the rest of (6.15) is linear and can be written
as
ξ 2 ( t + h) = Φ2ξ 2 ( t) (6.23)
where
yc − ˆ
 
 z 
ξ2 = 
 
 (6.24)
 dc ˆ 
wz

103
Chapter 6. Temporally Coupled DMPs in 3D Orientation Space

so that ξ = [ξ 1T x ξ 2T ]T , and

h
 
I I 0 0
τ
3 4
 
 
 hαβ hα 
− I 1− I
 
0 0 
 τ τ 
Φ2 = 
  (6.25)
h 
 0 0 I I 
τ
 
3 4
 

 hαβ hα 
0 0 − I 1− I
τ τ

The eigenvalues of Φ2 are given by

hα
λ1,...,2m = 1 − (6.26)
2τ
and because they are all strictly within the unit circle since h, α, τ > 0,
(6.23) is contracting. It now follows from Proposition 6.1 that the hierarchy
(6.15) is contracting. Hence, it converges globally exponentially to a single
trajectory. Since one solution to (6.15) is ξ = 0, any trajectory must converge
exponentially to this point. 2

104
7
Detection and Control of
Contact-Force Transients

7.1 Introduction
Robot programming is traditionally done using position-based control. Se-
quences of movements are specified with details such as velocities, target
positions, etc. Robot positions in free space are uncomplicated to visualize
and simulate, for instance with software tools such as rviz [ROS, 2018] and
Gazebo [Gazebo, 2018], and successful movement can be verified by com-
paring measured position with target position. Control in the contact-force
domain is also possible [Olsson et al., 2002; García et al., 2006], but this is
a more difficult control problem and forces are only indirectly visible [Jo-
hansson et al., 2015]. Nevertheless, many applications rely on force control.
In friction-stir welding, for instance, a rotating tool is pushed against work
pieces with a specified contact force to generate heat from friction [Cook
et al., 2004; De Backer, 2014; De Backer and Bolmsjö, 2014; Bagge Carl-
son et al., 2016; Karlsson et al., 2016]. In lead-through programming, motor
torques are commanded to help the programmer in overcoming joint friction
[Stolt et al., 2015a].
Whereas the previous chapters focused on algorithms for motion control,
the aim of this chapter is to use contact forces for validation of manipulation
subtasks. In robotic manipulation tasks, such as assembly tasks, it might
not be possible to validate subtasks based on robot position. One reason
is that the position tolerances are typically lower than the position uncer-
tainties. The uncertainties are introduced both by positioning of the robot
end-effector, which can be done within 1 mm on a typical industrial robot,
and of the work objects. Further, early detection is usually desirable, and in
some cases force data may indicate completion earlier than position data.
Force measurements were used to validate subtasks in robotic assembly in,
e.g., [Stolt et al., 2011; Stolt et al., 2012b; Linderoth, 2013; Stolt, 2015].

105
Chapter 7. Detection and Control of Contact-Force Transients

Conditions for the validation consisted of contact-force thresholds and were

programmed manually. This programming requires expert knowledge and
is time consuming.
In contrast, the research in [Stolt et al., 2015b] allowed for the program-
mer to provide the robot with examples instead of programming explicitly.
Thereby, [Stolt et al., 2015b] contributed to more accessible robot program-
ming. Measured contact-force transients generated during robotic assembly
were recognized based on a classification model acquired using machine
learning, and used for validation. The final classification model consisted
of a support vector machine (SVM) [Bishop, 2007; Murphy, 2012], trained
by solving a convex optimization problem in CVX [Grant and Boyd, 2008;
CVX Research Inc., 2019]. Some hyperparameters were determined in an
initial training phase, where the classification was based partly on a lin-
ear model learned using the least-squares method [Bishop, 2007; Murphy,
2012], and partly on several learnable threshold detectors combined into
a boosting classifier. This approach reduced the assembly time, compared
to using force thresholds only. It also removed the necessity to determine
and specify any force threshold manually, which would have required con-
siderable engineering work and explicit robot programming. As compared
to using position criteria, the approach in [Stolt et al., 2015b] also enabled
robots to switch between movements at the right time, despite any position
uncertainties, thus avoiding to push unnecessarily hard on work objects or
to leave any task unfinished.
In this chapter, we continue the work presented in [Stolt et al., 2015b],
with the following extensions. In [Stolt et al., 2015b], a force/torque sensor
was used to measure the contact force/torque. Such sensors and systems
are usually expensive, with costs comparable to the robot itself. If attached
to the wrist of the robot, it would introduce extra weight that the robot
would have to lift and move. Further, some robot models do not support any
seamless attachment of such sensors. If the force sensor would be attached
to an object in the work space, e.g., a table, this would imply restrictions
on where the assembly could take place. In this work, we therefore propose
a method based on robot joint-torque measurements on the motor side for
the detection, thus avoiding the requirement of an external sensor. Motor-
torque measurements are typically available, at least for the manufacturers,
which is a benefit. However, measuring on the motor side also introduces
a new difficulty. Due to friction, backlash, and elasticity between the arm
side and the motor side of each joint, some information is lost when using
motor torques as compared to an external force/torque sensor. A robot joint
model is visualized in Fig. 6.1 in [Olofsson, 2015]. Motor torque data could
be useful also when force/torque sensors are available, since the redundant
information could increase the confidence of state estimates. In this thesis,
the gear ratio between the motor side and the arm side has been compen-

106
 7.2 Problem Formulation

sated for, so that the presented joint torques are of the same magnitude as
on the arm side even though they have been measured on the motor side.
The research presented in this chapter also extends [Stolt et al., 2015b]
by investigating whether a given detection model could generalize to tasks
that involve new objects, from which no data have been used to deter-
mine the model. In order to investigate whether robot joint torques contain
enough information to distinguish transients, we follow a machine-learning
procedure. First, we gather data to determine a recurrent neural network
(RNN), which is an artificial neural network specialized in processing se-
quential data [Graves, 2012; Goodfellow et al., 2016], and subsequently we
evaluate the performance of the RNN on new data.
The procedure proposed does not rely on any assumption of what tasks
or work objects are considered, as long as distinguishable joint-torque tran-
sients are generated. In this chapter, we evaluate the procedure on the
manipulation scenarios shown in Figs. 7.1 and 7.2. These scenarios also
serve as examples of how the procedure could be used in practice.
Machine learning for analyzing contact forces in robotic assembly was
also applied in [Rodriguez et al., 2010], where force measurements were
used as input to an SVM, to distinguish between successful and failed
assemblies. A verification system, specialized in snap-fit assembly, was de-
veloped in [Rojas et al., 2012]. Similar to [Stolt et al., 2015b] and [Rodriguez
et al., 2010], a force/torque sensor was assumed in [Rojas et al., 2012]. Such
a requirement has been avoided in some previous research, by using in-
ternal robot sensors instead. For instance, a method to estimate contact
forces from joint torques was presented in [Linderoth et al., 2013]. Further,
force-controlled assembly without a force sensor was achieved in [Stolt et
al., 2012b], by estimating contact forces from position errors in the internal
controller of the robot.

7.2 Problem Formulation

Autonomous robotic manipulation requires in each subtask a validation
procedure, to determine when to switch to the next subtask. Without such
validation, subtask completion may not be verified, thus jeopardizing suc-
cessive steps and, eventually, the successful completion of the entire task.
In this research, we address the question of whether robot joint torques on
the motor side could be used to recognize contact-force transients during
robotic assembly, despite uncertainties introduced by, e.g., joint friction, in
order to automatically confirm completion of manipulation tasks. We fur-
ther investigate how long parts of the transients that need to be included
as input for the detection algorithm, to achieve satisfactory performance of
the detection. Finally, we investigate whether a detection model trained on

107
Chapter 7. Detection and Control of Contact-Force Transients

(a) (b)

(c) (d)
Figure 7.1 An ABB YuMi robot [ABB Robotics, 2018] was used in the
experiments (a). The experimental setup for the switch assembly task is
shown in (b), (c), and (d). The robot gripper, box, and switch are shown
in (b). The robot grasped the switch, and attached it to the box by pushing
downwards. The downward motion began in (c), where the switch was not yet
snapped into place. It ended when the assembly was completed, in (d). These
photos were taken during the experimental evaluation (see Sec. 7.4), and the
same setup was used for gathering training and test data (see Sec. 7.3).

(a) (b) (c)

Figure 7.2 Experimental setups for evaluation of the RNN model on objects
that had not been used for acquiring test or training data (Setup 7.2). The
pocket calculator and its battery cover are shown in (a), the spectacle case
is shown in (b), and the power switch on the extension cord is shown in (c).
For each setup, the robot was programmed to move the gripper downwards
until a transient was detected, and subsequently move the gripper upwards.

108
 7.3 Method

Table 7.1 Notation used in this chapter.

Notation Description
+
T ∈Z Number of samples in torque sequence
n pre ∈ Z+ Number of samples before transient peak
n post ∈ Z+ Number of samples after transient peak
n ch ∈ Z+ Number of input channels
tanh(·) Rn → Rn Element-wise hyperbolic tangent function
τ̄ ∈ Rnch Robot joint torque
h ∈ Rnch Activation of hidden RNN neurons
U, W ∈ Rnch $nch RNN weight matrices
b ∈ Rnch RNN bias vector
V ∈ R2$nch RNN weight matrix
c ∈ R2 RNN bias vector
o ∈ R2 RNN output vector
e ∈R Base of the natural logarithm
p̂ ∈ R2 Estimated probability distribution
p ∈ R2 Target label for given torque sequence
L ∈R Loss function

data from manipulation of a certain object could be used for detection in

similar tasks but with different objects and robots.

7.3 Method
In this section, the proposed machine-learning procedure to determine a
force/torque transient-detection model from training and test data is de-
tailed. The assembly scenario used to acquire the data consisted of attach-
ing a switch to a box, see Fig. 7.1. The objective for the robot was to move
toward the box while holding the switch, thus pushing the switch against
the box, until it snapped into place. The robot should detect the completion
of the task automatically, stop moving toward the box, and possibly start a
new movement. Table 7.1 lists some of the notation used in this chapter.

Sequence Model
An RNN [Graves, 2012; Goodfellow et al., 2016] was used as a sequence
classifier. This choice is discussed in Sec. 7.6. It had a sequence of joint
torques as input, one single output indicating whether the sequence con-
tained a given transient or not, one hidden layer, and recurrent connec-
tions between its hidden neurons. Each input torque sequence consisted of
T = n pre + 1 + n post time samples, where n pre and n post were determined

109
Chapter 7. Detection and Control of Contact-Force Transients

as explained later in this section. In turn, each time sample consisted of

n ch = 7 channels, i.e., one channel per robot joint. The dimension of the
hidden layer was chosen to be the same as the number of input channels,
n ch .
Denote by h(t) the activation of the hidden neurons at time step t. The
activation was defined recursively as

h(1) = tanh(b + Uτ̄ (1) ) (7.1a)

( t) ( t−1) ( t)
h = tanh(b + W h + Uτ̄ ) t ∈ [2; T ] (7.1b)

where U and W are weight matrices, both of size n ch $ n ch , b is a bias vector

with dimension n ch , and τ̄ (t) is the input joint torque at time t. Further,
tanh(·) represents the hyperbolic tangent function. After reading an entire
input sequence, the RNN produced one output o(T ) given by

o(T ) = c + V h(T ) (7.2)

where V is a weight matrix of size 2 $ n ch , and c is a bias vector with

dimension 2. Finally, the softmax operation was applied to generate p̂, a
vector that represents the normalized probabilities of whether a transient
is present.
C (T ) (T )
DT
T eo 1 eo 2
p̂ = [p̂ 1 p̂ 2 ] = (T ) (T ) (T ) (T )
(7.3)
eo 1 + eo 2 eo 1 + eo 2
(T )
Here, o i represents the i:th element of the output vector, p̂ 1 represents the
estimated probability that a transient is present, and p̂ 2 is the estimated
probability that a transient is not present, so that p̂ 1 + p̂ 2 = 1. If p̂ 1 was
larger than a desired decision boundary (0.5 was chosen, see Sec. 7.6), the
data point was classified as positive, i.e., it was indicated that the task was
completed within the sequence. Vice versa, if the first element was less than
the decision boundary, the data point was classified as negative. The RNN
architecture is visualized in Fig. 7.3.

Acquisition of Training Data and Test Data

Training data and test data were obtained as follows. The right arm of
the robot was used to grasp the switch, just above the box, as shown in
Fig. 7.1. Thereafter, a reference velocity was sent to the internal controller
of the robot, causing the robot gripper to move toward the box at 1.5 mm/s,
thus pushing the switch against the box. Once the switch was snapped
into place, the robot was stopped manually by the robot operator. The robot
joint torques were recorded in 250 Hz. The torque transient, induced by the
snap-fit, was labeled manually, and used to form a positive data point. This

110
 7.3 Method

Weighted
L
cross-entropy

Target Normalized
p p̂
label probability
softmax
o(T ) Output vector
V

W W W
h(1) h(2) h(3) h(T ) Hidden layer

U U U U

Input
τ̄ (1) τ̄ (2) τ̄ (3) τ̄ (T ) torque

Figure 7.3 The RNN visualized as an unfolded computational graph,

where each node is associated with a certain time step. The biases b and c,
as well as the activation function tanh(·), are omitted for a clearer view, but
the computations are detailed in (7.1) to (7.3). The input torque was used
to determine the hidden state, which was updated each time step. The last
hidden state was used to determine the normalized probability p̂ of whether
a transient was present in the time sequence or not.

procedure was repeated 50 times, which yielded 50 positive data points.

Data prior to each transient were used to form negative data points.
Given n pre and n post , a positive data point was formed by extracting
a torque sequence, from n pre samples previous to the peak value of the
transient (inclusive), to n post samples after (inclusive). Negative data points,
with the same sequence length T as the positive ones, were extracted
from torque measurements that ranged from a couple of seconds before the
transient, until the positive data point (exclusive). The negative data points
were chosen so that overlap was avoided. For each data point, the target was
labeled as a two-dimensional one-hot vector p, where p = [1 0]T represented
a positive data point indicating that a transient was present, and p = [0 1]T
represented a negative data point indicating that no transient was present.
Note that with the approach above, it was possible to extract several
negative data points, but only one positive data point, for every assembly
trial experienced by the robot.
Half of the positive and negative data points were used in the training
set, and the other half was used in the test set.

111
Chapter 7. Detection and Control of Contact-Force Transients

Model Training
Given the training set, the model parameters U, V, W, b, and c were deter-
mined by minimizing a loss function L. The training set contained much
more negative data points than positive ones. If not taken into account, this
type of class imbalance has been reported to obstruct the training procedure
of several different classifiers. The phenomenon has been described in more
detail in [Japkowicz, 2000; Japkowicz and Stephen, 2002], and should be
taken into account when designing the loss function. Consider first the fol-
lowing loss function L̃, which is the ordinary cross-entropy between training
data and model predictions, averaged over the training examples.
D A
1 XX d ! "
L̃ = − p a log p̂da (7.4)
D
d =1 a=1

Here, a and d are indices for summing over the vector elements and training
data points, respectively. This cross-entropy is commonly used as a loss
function in machine learning [Rubinstein and Kroese, 2013; Goodfellow et
al., 2016]. Because of the class imbalance in the present training set, it
would be possible to yield a relatively low loss L̃ by simply classifying all
or most of the data points as negative, regardless of the input, even though
that strategy would not be desirable.
In order to take the class imbalance into account, weighted cross-entropy
[Japkowicz and Stephen, 2002; Panchapagesan et al., 2016] was used as loss
function. Denote by r the ratio between negative and positive data points
in the training set, and introduce the weight vector wr = [r 1]T . The loss
function was defined as
D A
1 XX d ! "
L=− p a log p̂da · wTr yd (7.5)
D
d =1 a=1

The values of n pre and n post were determined using both the training
set and the test set as follows. All positive data points available were used,
and r = 20 times as many negative data points. Starting with n pre =
n post = 1, the model was trained using the training set, and its performance
was measured using the test set. Subsequently, both n pre and n post were
increased by 1, and the training and evaluation procedure was repeated.
This continued until perfect classification was achieved, or until the values
of n pre and n post were large; 30 was chosen as an upper limit, though it
was never reached in the experiments presented here. Thereafter, n pre was
kept constant, and it was investigated how much n post could be lowered with
retained performance. This was done by decreasing n post one step at a time,
while repeating the training and evaluation procedure for each value. Once
the performance was decreased, the value just above that was chosen for

112
 7.3 Method

n post . This way, the lowest possible value of n post was found, that resulted
in retained performance.
Once n pre and n post were determined, new model parameters were ob-
tained by training on a larger data set, with r = 100. The reason for using
a lower value for the other iterations, was that it took significantly longer
computation time to use such a large data set.
Because of the class imbalance in the test set, ordinary classification ac-
curacy, as defined by the number of correctly classified test data points
divided by the total number of test data points, would not be a good
model-performance measurement. Instead, the F-measurement [Hripcsak
and Rothschild, 2005] was used, defined as

PR
F1 = 2 (7.6)
P+R

where P is the precision, i.e., the number of correctly classified positive

data points divided by the number of all data points classified as positive
by the model, and R is the recall, i.e., the number of correctly classified
positive data points divided by the number of all data points that were
truly positive. The value of F1 ranges from 0 to 1, where 1 indicates perfect
classification.

Software Implementation
The RNN in Sec. 7.3 was defined as a computational graph in the Julia pro-
gramming language [Bezanson et al., 2014], using TensorFlow [Abadi et al.,
2016; TensorFlow, 2019] and the wrapper TensorFlow.jl [Malmaud, 2017].
The Adam algorithm [Kingma and Ba, 2014] was used for minimization of
the loss function L.
After training, the RNN model was saved on a server, and loaded into a
Julia program on a PC, which communicated with the internal controller of
the robot as described in Chapter 2. The sample frequency was 250 Hz, and
hence the sample period was 4 ms. The robot joint torques were logged and
saved, and for each time sample t, a joint-torque sequence was formed by the
samples in [ t − n pre − n post ; t], and sent as input to the RNN. Measurements
after time t were not available at t, which is why the input only contained
samples up until t. Each sequence was classified in real time. The compu-
tation time for one classification was short; well below the sample period.
To move the robot, desired velocity references for the gripper in Cartesian
space were first specified in the Julia program. Then, the corresponding
joint velocities were computed using the robot Jacobian, and these were
sent as references to the internal controller of the robot.

113
Chapter 7. Detection and Control of Contact-Force Transients

7.4 Experiments
Two ABB YuMi robots, shown in Figs. 7.1 and 7.4 and described in Chapter 2,
were used for experimental evaluation. The implementation in Sec. 7.3 was
used for robot control and transient detection. To facilitate understanding
of the experimental setups and results, a video is available on [Karlsson,
2017b]. The same RNN, obtained as described in Sec. 7.3, was used for
detection in each setup. The following three setups were considered.

Setup 7.1 The same robot as used for RNN training was used. Since the
test set had been used to determine the hyperparameters of the RNN, i.e.,
n pre and n post , it was necessary to gather new measurements to evaluate the
general performance. The experimental setup was similar to that in Sec. 7.3,
except that the measured torque sequences were saved and classified by
the RNN, instead of just saved to the training and test sets. The robot was
programmed to first move its gripper down, thus pushing the switch against
the box. Once a transient was recognized, it was programmed to stop its
downward motion, and instead move the box to the side. The assembly was
repeated 50 times, to evaluate the robustness of the proposed approach.
The experimental setup is visualized in Fig. 7.1.

Setup 7.2 The same robot as used for RNN training was used for manipu-
lation tasks that involved objects not used during model training, in order
to test its generalizability. Again, the robot was programmed to move its
gripper down until a transient was detected. Once a transient was detected,
it was programmed to move its gripper upwards. Three tasks were consid-
ered: snapping a battery cover into place on a pocket calculator, closing a
spectacle case, and pushing a power switch on an extension cord. The setup
for each task is shown in Fig. 7.2. For each of these tasks, 50 attempts were
made for validation.

Setup 7.3 A different robot individual but of the same model as used for
RNN training was used for manipulation of a power switch. The setup is
shown in Fig. 7.4. These experiments were done at ABB Corporate Research
in Västerås, Sweden. The RNN was downloaded on a local PC, and used for
the detection. A total of 5 trials were done.

7.5 Results
The performance of the RNN on the test set, for different values of the
hyperparameters, is shown in Table 7.2. The abbreviations are as follows:
number of true positives (TP), true negatives (TN), false positives (FP),
and false negatives (FN). The hyperparameters were increased until n pre =

114
 7.5 Results

(a) (b)
Figure 7.4 Experimental setup at ABB Corporate Research in Västerås,
Sweden (Setup 7.3). The YuMi robot was different from the one used to
acquire training and test data. The task began in (a) and finished in (b).

n post = 5, for which perfect classification was obtained. Then, n post was
decreased until the performance decreased at n post = 2. With this value of
n post , larger values of n pre were tested (see second last row in Table 7.2),
which did not yield perfect classification for any value, i.e., F1 < 1. Thus,
(npre , npost ) = (5, 3) was chosen for the final model.
After training, the RNN detected all 50 snap-fits in Setup 7.1 correctly,
without any false positives prior to the snap. The torque data and RNN
output from one of the trials are shown in Figs. 7.5 and 7.6. The other trials
gave qualitatively similar results. The RNN also detected each transient

Table 7.2 RNN performance on the test set, for different values of the
hyperparameters. The row with the lowest value of n post that yielded perfect
classification is marked in blue. With these values, the model was trained
and tested again, but now with more negative data points (see last row,
marked in red).
n pre n post TP TN FP FN P R F1
1 1 20 500 0 5 1 0.80 0.89
2 2 22 500 0 3 1 0.88 0.94
3 3 23 498 2 2 0.92 0.92 0.92
4 4 23 500 0 2 1 0.92 0.96
5 5 25 500 0 0 1 1 1
5 4 25 500 0 0 1 1 1
5 3 25 500 0 0 1 1 1
5 2 22 500 0 3 1 0.88 0.94
6, 7, . . . 30 2 - - - - - - <1
5 3 25 2500 0 0 1 1 1

115
Chapter 7. Detection and Control of Contact-Force Transients

6
τ̄1
τ̄2

Torque [Nm]
4
τ̄3
2 τ̄4
τ̄5
0 τ̄6
τ̄7
−2
0 1 2 3 4 5 6 7 8
1 p̂ 1
Probability

DB
0.5

0
0 1 2 3 4 5 6 7 8
Time [s]

Figure 7.5 Data from Setup 7.1. The robot joint torques (upper plot) were
used as input for the RNN. The torque of joint i is represented by τ̄i . The
first element of the RNN output (p̂ 1 in the lower plot) was close to 0 before
the assembly was completed, and increased to close to 1 once the task was
finished. Completion was indicated when p̂ 1 was above the decision boundary
(DB, at 0.5) for the first time. Thus, for detection purposes, the RNN output
generated after this event was not relevant.

successfully, without any false positives, for each trial in Setup 7.2. Data
from one trial of each task are shown in Fig. 7.7, and the other trials gave
qualitatively similar results. Similarly, successful detection was achieved
for the 5 trials on the robot that had not been used for training (Setup 7.3).
Data from one of these trials are shown in Fig. 7.8.

7.6 Discussion
There are several alternatives to RNNs for classification. Using an RNN
is motivated as follows. Two less complicated models, matched filter and
logistic regression, were tried initially, without achieving satisfactory per-
formance on test data. A linear model, an SVM, and learned force thresholds
for transient detection have been evaluated in [Stolt et al., 2015b]. RNNs
are specialized in processing sequential data, and the torque measurements
used in this work were sequential. Two properties of the RNN contribute to
lower its complexity, as compared to a standard artificial neural network.
These consist of parameter sharing, which means that the same parame-

116
 7.6 Discussion

6
τ̄1
τ̄2

Torque [Nm]
4
τ̄3
2 τ̄4
τ̄5
0 τ̄6
τ̄7
−2
6.27 6.28 6.29 6.3 6.31 6.32 6.33 6.34
1 p̂ 1
Probability

DB
0.5

0
6.27 6.28 6.29 6.3 6.31 6.32 6.33 6.34
Time [s]

Figure 7.6 Same data as in Fig. 7.5, but zoomed in on the time when the
task finished. The transient was detected at time t = 6.308 s. The first torque
sequence to be classified as positive was that within the vertical dashed lines
in the upper plot. Figure 7.5 might give the impression that p̂ 1 increased
from 0 to 1 instantly upon detection. In this figure, however, it can be seen
that the confidence was actually built up during a couple of sample periods,
which was also the case for the remaining trials.

ters are used in several connections, and sparsity, which means that only
some of the neurons are connected to each other. This can be seen in
Fig. 7.3. Thanks to the lower complexity, it is possible to estimate a model
with significantly fewer training examples than would be needed otherwise.
Compared to models that are not specialized in sequential data, e.g., logit
models, SVMs, and ordinary neural networks, the RNN is less sensitive to
variations of the exact time step in which some information in the input
sequence appears. An RNN can also be generalized to classify data points
of sequence lengths not present in the training set, though this was not
used in this present work. General properties of RNNs are well described
in [Goodfellow et al., 2016].
The concept of RNNs is well known from previous research [Graves,
2012; Goodfellow et al., 2016], and hence it should not be seen as a con-
tribution from this chapter. Instead we have used it to evaluate whether
transients in robot motor torques could be used for validation of manipu-
lation tasks, which in turn is the main purpose of this chapter. The RNN
architecture described in Sec. 7.3 serves as an example of how joint torques
could be used for validation in practice.

117
Chapter 7. Detection and Control of Contact-Force Transients

τ̄1
5 τ̄2

Torque [Nm]
τ̄3
τ̄4
0 τ̄5
τ̄6
τ̄7
−5
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
1 p̂ 1
DB
Probability

0.5

0
0 0.5 1 1.5 2 2.5 3 3.5 4 4.5 5
Time [s]

5 5
Torque [Nm]

0 0

−5 −5
0 2 4 0 2 4
1 p̂ 1 1 p̂ 1
DB DB
Probability

0.5 0.5

0 0
0 2 4 0 2 4
Time [s] Time [s]

Figure 7.7 Experimental data from the manipulation tasks in Setup 7.2;
snapping the battery cover into place on the pocket calculator (upper), closing
the spectacle case (lower left), and pushing the power switch on the extension
cord (lower right). The organization of each plot is the same as in Figs. 7.5
and 7.6.

118
 7.6 Discussion

τ̄1
5
τ̄2

Torque [Nm]
τ̄3
τ̄4
0 τ̄5
τ̄6
τ̄7
−5
0 0.5 1 1.5 2 2.5 3
1 p̂ 1
Probability

DB
0.5

0
0 0.5 1 1.5 2 2.5 3
Time [s]

Figure 7.8 Experimental data from Setup 7.3, where a power switch on
an extension cord was pushed by a YuMi robot not used for RNN training.
Detection occurred at time t ( 1 s.

As compared to [Stolt et al., 2015b], the approach proposed here had four
new benefits. A force sensor was no longer required, the detection delay was
reduced, the computation time for model training was shortened, and the
confidence level for detection could be set through one parameter. In [Stolt
et al., 2015b], n post > 10 (corresponding to > 40 ms) was required for perfect
classification, whereas our proposed method required n post = 3 (12 ms). The
training time of the RNN was in the order of minutes on an ordinary PC,
which was an improvement compared to days in [Stolt et al., 2015b]. The
decision boundary for p̂ 1 in Sec. 7.3 corresponds to a desired confidence level
for detection. It was set to 0.5 in Sec. 7.3, meaning that a task was classified
as completed when the estimated probability of completion was above 50 %.
However, the meaning of p̂ 1 is intuitive, and the decision boundary can
easily be adjusted to a desired level of confidence. For example, it might be
worse to fail to detect a force transient and risk to damage the hardware,
than to finish a subtask too early and possibly detect this mistake later on.
Such a cost asymmetry would motivate a lower decision boundary. In some
scenarios, however, backup detectors may confirm completion reliably but
late, and it might then be more important to avoid false positives. This was
the case in the experimental scenario in [Stolt et al., 2015b]. The ability to
adjust the confidence level according to specific circumstances is valuable.
In some approaches, the confidence level is acquired by weighting false

119
Chapter 7. Detection and Control of Contact-Force Transients

positive and false negative points differently in the cost function used in the
model training, see, e.g., [Stolt et al., 2015b]. This requires a new training
and validation procedure to adjust the confidence level, which is avoided
with the method proposed here.
Further, this research extended [Stolt et al., 2015b; Karlsson, 2017e]
by verifying generalization of the classification model to manipulation of
new objects, that had not been used to generate training and test data. It
is expected that this generalization is limited to similar tasks. Therefore,
the idea is that the training procedure described in Sec. 7.3 should be
gone through whenever a detection model for a completely different task is
required.
Given a certain contact force/torque acting on the end-effector of the
robot, the corresponding motor torques depend on the configuration, as
well as gravity and friction between arm side and motor side of each joint.
It is expected that different robot individuals of the same model would
experience slightly different motor torques for a given task, because of small
uncertainties in the robot manufacturing and different wear and tear. In
this chapter, it has been shown that generalization of the RNN to a new
robot individual is possible. This fact was also used during the final review
of the SARAFun project [SARAFun, 2019], where Setup 7.3 was used to
demonstrate the proposed detection algorithm.
However, changing robot model or configuration would change the joint
torques significantly, and generalization can not be expected with this cur-
rent implementation. Figure 7.9 shows the success rate for the setup in
Fig. 7.4, but where the task was carried out at different distances from the
position used during training. As expected, the success rate deteriorated
with increased distance. The failures consisted both of false detections and
undetected transients. However, already this approach allowed for some
uncertainty in the work-space configuration, in the order of 0.1 m. Even
though generalization to the specific configurations tested in Fig. 7.9 could
be achieved by first including those configurations in the training proce-
dure, it would not be a feasible strategy for generalization to the entire
work space. The reason is that this would require more training examples
than one could provide in reasonable time.
To enable generalization also to new configurations and robot models, it
would be a good idea to first model and compensate for gravity and friction-
induced motor torques, then estimate the external contact forces/torques,
and subsequently use these as input for the detection model. Furthermore,
if external force/torque sensors are available in a given setup, it would be
straightforward to include those measurements in this approach. Gener-
alization to new robots and configurations would then easily be achieved.
Motor torques could still be relevant for validation of the external measure-
ments.

120
 7.6 Discussion

100

80

Success rate [%]

60

40

20

0
0 0.05 0.1 0.15 0.2 0.25 0.3
Distance [m]
(a) (b)

Figure 7.9 Evaluation of success rate depending on distance from work-

object position used during training. The setup is shown in (a), where the
task was performed 0.3 m away from the position used during training.
Furthermore, the robot and the work object were different from those used
during training. The success rates for different distances are shown in (b).
Five trials were done for each distance.

In this architecture, task completion was concluded the first time p̂ 1 was
greater than the decision boundary. Therefore, the values of p̂ 1 after that
event were not relevant for detection purposes, and torque measurements
after any transient were not used as training or test data. Further, the
torques after task completion depend not only on the interaction forces, but
also on the commanded behavior of the robot.
The concept of weighted loss to compensate for class imbalance in ma-
chine learning has been evaluated in [Japkowicz and Stephen, 2002], and
successfully applied to a deep neural network in [Panchapagesan et al.,
2016]. The loss function (7.5) extends (7.4) by the factor wTr yd , which evalu-
ates to r for positive data points, and to 1 for negative ones.
In Fig. 7.5, it should be noted that p̂ 1 raised significantly above 0 at
t ( 4.2 s, even though the task did not yet finish at that time. Even though
the values were still well below the decision boundary, this leaves room for
improvement in terms of reliability of the detection.
In a complete setup, the validation procedure described here could
be combined with more sophisticated motion controllers, such as DMPs.
Whereas the evaluation of the model was done in real time in this work,
the RNN training was performed offline. A complete product should have a
user interface allowing for an operator to gather training data and test data,
label them, and run the training procedure. This would shorten the time
required for acquisition of training and test data, which was approximately
one working day with the current arrangement. It would also be valuable if

121
Chapter 7. Detection and Control of Contact-Force Transients

labels could be suggested automatically. This could for instance be achieved

using a simple classifier trained on some of the first labels, which would
have to be determined manually. It is also important to shorten the reaction
time of the robot, i.e., to further reduce the value of n post . This could be done
by including more sensors. For instance, the snap-fit assembly generates a
sound, easily recognized by a human. Adding a microphone to the current
arrangement would therefore add information to the detection approach. In
turn, this could be used to decrease the required amount of training data,
improve robustness of the detection, or detect the transient earlier.

7.7 Conclusion
In this research, we have addressed the question of whether robot joint-
torque measurements could be used for detection of force/torque transients
in robotic manipulation. We have shown that such detection is possible,
which is the main contribution of this chapter. In contrast, the concept of
RNNs is well known from previous research, and was used in this chapter
for the purpose of evaluation, as well as to exemplify how joint torques could
be used for detection in a practical setup. First, training and test data were
gathered and labeled. Then, these were used to determine the parameters of
the detection model. Finally, a real-time application for transient detection
was implemented and tested, both on the assembly task used to generate
training and test data, and on new tasks and using a new robot, that had not
been used to determine any model parameters. The method presented seems
promising, since the resulting model had high performance, both on the test
data and during the experiments. A video that shows the functionality is
available on [Karlsson, 2017b].

122
8
A Dual-Arm
Haptic Interface
for Task Demonstration

8.1 Introduction
Whereas purely position-based programs can be mediated to robots through
ordinary lead-through programming, tasks that involve contact forces be-
tween a robot and its surroundings are not as straightforward to demon-
strate. Ordinary lead-through programming, using only one robot arm,
yields forces between the operator and the robot which can not be distin-
guished from the task-specific forces in general. In this chapter, an immer-
sive haptic interface for task demonstration is presented and evaluated. The
interface allows for an operator to act and sense remotely, thereby demon-
strating desired movements and interaction forces. The interface consists of
two robot arms, one of which can be moved directly by an operator through
physical contact, and one that moves accordingly, while contact forces are
reflected to either arm as haptic feedback. A control law enforces a constant
pose offset in Cartesian space between the end-effectors of the robot arms.
We refer to the robot arm in direct contact with the operator as the master
side, and to the other robot arm as the slave side. However, the presented
algorithm is symmetric with respect to the robot arms, i.e., either arm could
be used as master or slave without any specification. The concept is visu-
alized in Fig. 8.1 and in a video available online [Ghazaei Ardakani et al.,
2019].
Haptic feedback for robotic teleoperation [Hokayem and Spong, 2006]
and for virtual reality [Constantinescu, 2008] has become an important
research topic. Four-channel haptic systems represent the most general
form [Aliaga et al., 2004], where position and force data are transferred
between the master and slave sides in both directions. In addition to a

123
Chapter 8. A Dual-Arm Haptic Interface for Task Demonstration

Figure 8.1 Example use case of the proposed interface. An operator demon-
strates a peg-in-hole task by moving the left robot arm through physical
contact. The right robot arm follows the demonstrated movement in real
time.

haptic system, visual feedback from the slave side to the master side is
typically necessary. In the arrangements considered in this chapter, the
slave side can be seen directly by the operator, see Fig. 8.1. If this is not the
case, cameras on the slave side are commonly used. A teleoperated interface
is transparent if movement and contact forces are reflected to either side
through the interface, so that the operator has the feeling of interacting
with the remote environment.
There are several factors that limit the transparency. In [Lawrence,
1993], it has been shown that transparency and stability are conflicting fea-
tures. Delays are detrimental both to stability and transparency [Anderson
and Spong, 1989; Lee and Spong, 2006; Hatanaka et al., 2015]. Light-weight
collaborative robots may be used as haptic devices without external sensors.
Joint friction should then be compensated for, and uncertainties in the fric-
tion model would limit the transparency. In this chapter, we describe how
measurements of external forces can be included in the control law, but
in the experiments we consider a worst-case scenario where external sen-
sors are not available. Transparency is further limited by the structures
of the master and slave devices. The common work space is limited to
the configurations reachable by both devices simultaneously. At the work-
space boundary, at least one device has reduced manipulability because of
a singular configuration or limits of reach or joint angles. It is desirable to
provide feedback regarding these internal limitations, though this contra-
dicts transparency with respect to the environment. The operator typically
expects force feedback both from limitations of the robotic system, and from
interaction with the environment. The proposed algorithm supports deploy-

124
 8.2 Problem Formulation

ment of two not-necessarily-similar robots, and the ability to reflect the

internal limitations of the system then becomes of major importance.
The main contribution of this chapter is the implementation and evalu-
ation of a haptic interface between two not-necessarily-similar robot arms,
which can be interacted with at any point of their structures. Whereas
in previous methods, the master and slave devices have commonly been
mounted and configured to avoid difficulties in case of dissimilarities be-
tween their individual work spaces, this proposed method supports use of
any six-degree-of-freedom or redundant robot arms independently, mounted
on any surfaces and with any initial configurations. It further results in the-
oretically perfect pose tracking during free-space motion, and perfect force
tracking in static contact with the environment except at the boundary of
the common work space where internal limitations are sensed by the oper-
ator. Any of the arms can pass through singular configurations while the
haptic feedback is adjusted accordingly. The control law is based on virtual
constraints, and suits well for kinesthetic teaching. The robot arms are
coupled in Cartesian space to retain a constant pose offset between their
end-effectors. This chapter is based on our previous publication [Ghazaei
Ardakani et al., 2018], in which we extended the research in [Ghazaei
Ardakani, 2016] by implementing the proposed algorithm and evaluating
it experimentally. Simulation results are available in [Ghazaei Ardakani,
2016].

8.2 Problem Formulation

It is difficult to program robots for tasks that require force interaction,
because both knowledge of the specific task and the robot dynamics are
necessary. Whereas positions can be visualized and specified graphically
in user interfaces, interaction forces are not directly visible [Johansson et
al., 2015]. A human typically does not know required interaction forces in
quantitative terms as physical entities, but may still be able to perform
and demonstrate the task. Programming by demonstration is a natural way
for humans to mediate behavior to robots [Lee et al., 2015]. If interaction
forces could be extracted from a demonstration, intuitive specification of
required forces would be achieved. Lead-through programming [Pan et al.,
2010; Stolt et al., 2015a; Capurso et al., 2017] has the benefit that it allows
for the operator to perceive the mechanical properties and limitations of
the robot during the demonstration. Unfortunately, however, ordinary lead-
through programming implies contact forces between the operator and the
robot, which are in general impossible to distinguish from the demonstrated
interaction forces between the robot and the work objects.
In this chapter, we therefore implement a dual-arm haptic interface for

125
Chapter 8. A Dual-Arm Haptic Interface for Task Demonstration

task demonstration, where one robot arm can be moved directly by the
operator, and the other robot arm moves accordingly. Because interaction
with the operator and with the work objects have been separated to dif-
ferent robot arms, demonstrated forces between robot and work objects are
easily distinguished. In the context of task demonstration through a haptic
interface, transparency is an important aspect. Further, it is important to
consider stability, especially during physical contact with a stiff environ-
ment. Therefore, this chapter also aims to evaluate the movement and force
tracking between the two robot arms, as well as the stability of the proposed
system.

8.3 Previous Research

A survey of haptic feedback in teleoperation is given in [Hokayem and
Spong, 2006]. There are currently several approaches to coupling a slave
and a master device. In [Rebelo and Schiele, 2012], the placement of the
slave device was optimized to maximize manipulability, and position offset
and scaling were used to avoid work-space limitations. A point-to-point
kinematic mapping to maximize the common work space was presented in
[Chen et al., 2007]. The trajectory of the master robot was mapped into
positions that were reachable for the slave robot.
A passive control law for teleoperation in joint space, capable of inertia
scaling and obstacle avoidance, was proposed in [Lee and Li, 2005]. Virtual-
constraint forces have been introduced using potential fields, which can be
used to help the operator to remain in a region of interest, and avoid ob-
stacles and singularities [Turro et al., 2001]. Furthermore, force feedback
from virtual damped springs have been introduced in virtual-reality appli-
cations [Constantinescu et al., 2006], to simulate contact with the virtual
environment.
In the context of kinesthetic teaching, it is desirable to compensate
for joint friction. Advancements in sensorless force estimation have been
reported in [Linderoth et al., 2013; Wahrburg et al., 2014; Stolt et al.,
2015c], and lead-through programming using friction compensation has
been developed in [Stolt et al., 2015a; Ghazaei Ardakani, 2016; Capurso
et al., 2017].
The control law described in this chapter is synthesized by constraining
the poses of the robot end-effectors virtually. Virtual constraints have been
used for control synthesis in previous research, see, e.g., [Shiriaev et al.,
2007; Westervelt et al., 2007; Freidovich et al., 2008]. Further, several
researchers have taken the view of a teleoperation system as a passive
mechanical tool [Itoh et al., 2000; Lee and Li, 2003], which is also done in
this chapter.

126
 8.4 Method

Table 8.1 Notation used in this chapter. The subscripts 1 and 2 are used
to distinguish each robot arm. The subscript i indicates one arbitrary arm.

Notation Description
+
ni ∈Z Number of joints on arm i
n ∈ Z+ n1 + n2
q ∈ Rn Joint positions
pi ∈ R3 Cartesian position
Ri ∈ SO(3) Cartesian orientation
Ji ∈ R6$ni Robot Jacobian
G ∈ R6$n [− J1 J2 ]
M ∈ Rn$n Mass matrix
C ∈ Rn$n Centripetal and Coriolis matrix
Qe ∈ Rn External torques
Qm ∈ Rn Motor torques scaled to arm side
Qvc ∈ Rn Torques due to virtual constraint
λ ∈ R6 Forces due to virtual constraint
e ∈ R6 Pose deviation from offset constraint
ẏ ∈ R6 Difference between end-effector velocities
Γ ∈ R6$6 G M −1 GT
x, y, z ∈R Cartesian coordinates, see Fig. 8.3

8.4 Method
The control law for the haptic interface implemented in this chapter was
first described in [Ghazaei Ardakani, 2016], and is briefly described in this
section for convenience. The main idea is to virtually constrain the two
end-effectors to have a constant pose offset, as if they were physically and
rigidly attached to each other, and subsequently design a control law that
realizes this constraint. Table 8.1 lists some of the notation used in this
chapter.

Virtual Constraint
Denote by n 1 and n 2 the number of joints on the first and second arm,
respectively, and let n = n 1 + n 2 represent the total number of joints.
Throughout this chapter, we will continue to use the subscripts 1 and 2
to associate quantities with each of the two arms. Further, let q 1 ∈ Rn1
and q 2 ∈ Rn2 represent joint positions, and introduce q = [qT1 qT2 ]T . The

127
Chapter 8. A Dual-Arm Haptic Interface for Task Demonstration

geometric constraint between the end-effectors can be expressed as

p 2 − p 1 = ∆p (8.1a)
RT1 R 2 = ∆R (8.1b)

where p i ∈ R3 represents position, R i ∈ SO(3) represents orientation, and

∆p and ∆ R are constant offsets in position and orientation, respectively. For
the control design, ∆p and ∆ R will be seen as references for the pose offset.
Theorem 8.1
The geometric constraint in (8.1) implicates the kinematic constraint given
by
Gq̇ = 0 (8.2)
where G = [− J1 J2 ] ∈ R6$n consists of the robot Jacobians Ji . 2

Proof. See Sec. 8.A. 2

Control Algorithm
The equations of motion for a rigid-body model of each robot arm i can be
written as [Spong et al., 2006]

Mi (q i )q̈ i + C i (q i , q̇ i )q̇ i = Qei + Qm

i − µi q̇ i (8.3)

where torques induced by Coulomb friction and gravity have been omitted
since we compensate for these separately. Here, Mi (q i ) is the mass matrix,
C i (q i , q̇ i ) represents the effect of centripetal and Coriolis forces, Qei rep-
resents external torques, Qm i represents motor torques scaled to the arm
side, and µi represents coefficients for viscous friction. Denote by Qvc motor
torques that enforce the virtual constraint. Our aim is to use these as con-
trol signals, so that Qm = Qvc . The equations of motion of the constrained
system can be derived from (8.2) and (8.3) as

M (q)q̈ + C (q, q̇)q̇ = Qe + Qvc − µq̇ (8.4a)

Gq̇ = 0 (8.4b)

where

M (q) = blkdiag ( M1 (q 1 ), M2 (q 2 )) (8.5a)

C (q, q̇) = blkdiag (C 1 (q 1 , q̇ 1 ), C 2 (q 2 , q̇ 2 )) (8.5b)

Furthermore, we introduce forces at the end-effectors that enforce the vir-

tual constraint, denoted by λ. These are related to the motor torques due
to the virtual constraint as follows.

Qvc = GT λ (8.6)

128
 8.4 Method

Intuitively, −λ and λ can be interpreted as interaction forces between the

end-effectors, if they were physically attached to each other. The interaction
forces have equal magnitude and opposite direction, as predicted by New-
ton’s third law of motion. The opposite directions are taken into account by
the different signs of the Jacobians in G, so that
C D C D
− J1T J1T (−λ)
Q vc
=G λ=
T
λ= (8.7)
J2T J2T λ

By introducing the state vector ξ = [qT q̇T ]T and replacing λ with u, (8.4)
can be written as
C D
q̇
ξ˙ = ! " (8.8a)
M −1 (q) −C (q, q̇)q̇ + Qe − µq̇ + GT (q)u
ẏ = G(q)q̇ (8.8b)

where ẏ denotes the difference between the end-effector velocities. Define

Γ = G M −1 GT . The control signal u∗ (ξ ) that yields zero relative acceleration,
i.e., ÿ = 0, can be obtained by differentiating (8.8b) with respect to time
and solving for u which yields
! "
u∗ = Γ −1 G M −1 (C q̇ − Qe + µq̇) − Ġq̇ (8.9)

This solves the servo problem. To also solve the regulation problem, i.e.,
to drive the physical offset to the reference in case of any deviation, we
introduce the state feedback

λ = u = u∗ − Γ −1 ( K d ẏ + K p e) (8.10)

This results in asymptotic stability of (8.8) [Ghazaei Ardakani et al., 2018].

Further, if one of the arms is in contact with the environment and standing
still, and no internal limits of the robots are reached, the forces acting on
the end-effectors are theoretically perfectly transferred [Ghazaei Ardakani
et al., 2018]. Here, K p and K d are positive definite gain matrices. These
matrices are block diagonal, i.e., K p = blkdiag( K tp , K op ), where K tp is a
positive definite gain matrix for translation, and K op is a positive definite
gain matrix for orientation. The matrix K d has a corresponding structure.
Further, e is the pose difference between the end-effectors beyond the de-
sired offset, i.e., e = [ eTP eTO ]T , where e P = p 2 − (p 1 + ∆p) and e O is the
imaginary part of the unit quaternion corresponding to R 2 ( R 1 ∆ R)T . Hence,
e represents an error which is 0 when the geometric constraint in (8.1) is
fulfilled. The error dynamics can be obtained by inserting the control law
(8.10) into the equations of motion (8.8) which yields

ÿ + K d ẏ + K p e = 0 (8.11)

129
Chapter 8. A Dual-Arm Haptic Interface for Task Demonstration

Q̂e m
Control Q = Q
vc
q, q̇
+ Mm (q)
Manipulators
Law

Gravity and friction

compensation

Figure 8.2 Control architecture. Estimated external torques Q̂e and the
states of the robots are inputs to the controller for the calculation of Qvc ,
the torques from the virtual constraint. The manipulators are gravity and
Coulomb-friction compensated.

The motor-torque references sent to the robots are given by

Qm = Qvc = GT λ (8.12)

The control architecture is visualized in Fig. 8.2.

8.5 Experiments
For the experiments, the ABB YuMi prototype was used. The control algo-
rithm in Sec. 8.4 was implemented on a PC that communicated with the
internal controller of the robot. According to Fig. 8.2, the sum of the gen-
eralized forces from the virtual constraint Qvc , and the gravity and friction
compensation signals, was sent as a torque reference to the internal robot
controller. For the friction compensation at zero joint velocities, a dithering
technique was used [Capurso et al., 2017]. The worst-case scenario was ex-
perimentally evaluated, where no external sensors were assumed and the
estimated external forces and torques were set to zero. Each end-effector
of the robot was equipped with a force/torque sensor, and these were used
for evaluation only. The experimental arrangement is shown in Fig. 8.3.
Additionally, a video that shows the setup and the experiments is available
online [Ghazaei Ardakani et al., 2019]. Two different setups were consid-
ered.

Setup 8.1 The left robot arm was used as master. First, free-space motion
was conducted, in order to evaluate the motion tracking between the two
arms. The right arm was then brought to collision with a concrete block
in front of the robot, in order to evaluate force tracking and stability in
contact with a stiff environment. When in contact, the operator pushed the
end-effector of the master arm in the positive y-direction, defined in Fig. 8.3,
so that the slave arm was pushed against the block in order to excite the
interaction forces.

130
 8.6 Results

z
y
x

(a) (b)
Figure 8.3 Robot configurations for experimental evaluation. Setup 8.1 is
shown in (a), where the left arm of the robot was used as master, and the
movement was such that the end-effector of the right arm established contact
with the concrete block in front of the robot. Setup 8.2 is shown in (b), where
the right robot arm was used as master, and the left arm was first brought
upwards to its reach limit.

Setup 8.2 The right robot arm was used as master. By moving the master
arm, the left arm was first brought to its upper reach limit. This configu-
ration was also a singularity, see Fig. 8.3. Thereafter, it was moved slightly
to the side thus reaching the angle limit of Joint 1, the joint closest to
the body. The purpose of this experiment was to evaluate stability proper-
ties and interaction forces when the slave arm reached internal physical
limitations.

8.6 Results
Data from Setup 8.1 are shown in Fig. 8.4. It can be seen that motion track-
ing was achieved. The bottom plot compares the forces in the y-direction, as
measured by the wrist-mounted force sensors. The force measured on the
left arm has been negated for easier comparison, and corresponds to the in-
teraction force experienced by the operator. After 50 s of free-space motion,
a contact was established between the right end-effector and the concrete
block. It can be seen that force tracking was achieved, meaning that the
operator experienced the same contact forces as the slave arm. Moreover,
the contact force from the concrete block prevented both arms from moving
further in the y-direction, despite the force exerted by the operator. This
behavior is the desired result of the virtual constraint.

131
Chapter 8. A Dual-Arm Haptic Interface for Task Demonstration

0.4 R.p x
R.p y

Position [m]
0.2
R.p z
0 L.p x
−0.2 L.p y
L.p z
−0.4
0 10 20 30 40 50 60 70 80 90
0.2
R.ṗ y
Velocity [m/s]

0.1 L.ṗ y
0
−0.1
−0.2
0 10 20 30 40 50 60 70 80 90
10
R. Fy
Force [N]

0 L. Fy

−10

−20
0 10 20 30 40 50 60 70 80 90
Time [s]

Figure 8.4 Results from Setup 8.1. For an uncluttered view, only some
of the dimensions are shown. The first plot from above shows the position
of the arms, where the position of the left arm has been compensated for
with the desired offset between the arms. The second plot shows velocities
in the y-direction, and the third plot shows forces in the y-direction. R.p x
represents the right arm’s position in the x-direction, and so on.

Data from Setup 8.2 are visualized in Fig. 8.5. The left arm was first
brought to its upper reach limit. Despite a large force exerted upwards on
the master arm by the operator, it can be seen that the master arm was
prevented from moving in the z-direction by the virtual constraint. Joint 1
of the left arm was thereafter brought to its angle limit. Then, further
motion of the arms in the limited direction was prevented despite pushing
the master arm, since the joint limit of the slave arm was being propagated
to the master arm thanks to the virtual constraint.
As expected from an impedance-type controller, stability was retained
both during free-space motion, during contact with a stiff environment, and

132
 8.7 Discussion

0.6 R.p x
R.p y
Position [m]
0.4
R.p z
0.2 L.p x
0 L.p y
L.p z
−0.2
0 5 10 15 20 25 30 35
−2
Joint 1 angle [rad]

L.q (1)
−2.2

−2.4

−2.6
0 5 10 15 20 25 30 35
30
R. Fy
20
Force [N]

R. Fz

10

0
0 5 10 15 20 25 30 35
Time [s]

Figure 8.5 Results from Setup 8.2. The master arm was subjected to large
forces twice; first when the left arm reached a work-space limit in the z-
direction, and thereafter when Joint 1 of the left arm was in its lower angle
limit.

while reaching stiff internal limitations. In addition, it can be concluded that

the external physical limitation caused by the concrete block in Setup 8.1
had a similar effect as the internal limitations in Setup 8.2; the motion of the
slave arm was limited directly, and the master arm was limited through the
virtual constraint. Moreover, the virtual constraint transferred the contact
forces between the arms, giving the operator the feeling of interacting with
the environment using a passive tool.

8.7 Discussion
The proposed control law solves the problem of capturing interaction forces
from demonstrations based on lead-through programming. The dual-arm

133
Chapter 8. A Dual-Arm Haptic Interface for Task Demonstration

lead-through programming is more user friendly and has less side effects
than single-arm arrangements, thanks to the separation between contact
forces from the operator and the work objects.
A possible drawback with the proposed approach is that the physical
dynamical properties of the robots may be undesired. This can be mitigated
by using local feedback to adjust the apparent dynamical properties, before
applying the proposed controller.
In the current sensorless implementation, the force tracking was mainly
limited by uncertainties in the friction model. A simple model was adopted,
consisting of Coulomb and viscous friction. In reality, many aspects af-
fect joint friction, such as load, configuration [Wahrburg et al., 2018], and
temperature [Bagge Carlson et al., 2015]. Especially zero-velocity friction
is difficult to estimate [Stolt et al., 2015a; Capurso et al., 2017]. In the
experiments, external sensors were not used for the haptic interface, to
investigate a worst-case scenario. The algorithm supports integration of ex-
ternal sensors, which would improve the overall performance significantly.
The sensorless approach is limited to light-weight collaborative robots, as
large industrial robots exhibit more joint friction and usually require exter-
nal sensing for any lead-through programming.
The prototype YuMi robot was used partly because it had force sensors
for validation, and partly because its research interface was more mature
than that of the product version of YuMi at the time, with predictable delays
and low jitter, which was important for the implementation.
This research has implications for the industry, as the robot company
Cognibotics AB, Lund, Sweden, aims to commercialize the interface. As a
future step, the implementation will be migrated to the product version
of YuMi, possibly using another version of the research interface. It will
then be a benefit that the product version of YuMi exhibits lower and more
predictable joint friction than the prototype.
Using a dual-arm robot for this interface has the benefit that two arms
are then naturally available. In such an arrangement, however, manipula-
tion tasks that require both arms leave no arm available as master side.
This can be solved by using yet another robot as master. For instance, it
would be straightforward to use an entire YuMi robot as master, controlling
all 14 degrees of freedom, and another one as slave. Enforcing the same
configuration on both robots would then be possible and perhaps more in-
tuitive for an operator, but would not be a requirement for the interface as
it also supports independent configurations of the master and slave.
The control signal in (8.9) results in zero dynamics of the system (8.8). It
was found by following the steps described in [Slotine and Li, 1991], i.e., time
differentiate the output until the control signal appears, and subsequently
choose the control signal to cancel nonlinearities and guarantee tracking.
Alternatively, the control signal for the zero dynamics could be derived

134
 8.7 Discussion

using the general results in [Isidori, 1995], as done in [Ghazaei Ardakani,

2016; Ghazaei Ardakani et al., 2018].
The interface also allows for humans to interact remotely with each
other. Early in the video in [Ghazaei Ardakani et al., 2019], two humans
shake hands remotely through the interface. This result highlights the fact
that the implementation is symmetric with respect to the robot arms, and
that the distinction between master and slave side is artificial. Haptic feed-
back and physical remote human–human interaction is likely to become im-
portant in, e.g., the video-game industry. Today, phone calls and video calls
are frequently used instead of physical meetings. Reducing transportation
needs could potentially save time and money, while decreasing the negative
impact on the environment. While the demand for faster and more efficient
transportation is high in general, in theory ideal telepresence could replace
almost any transportation of humans.
In the presented implementation, the robot arms were just virtually
constrained by the constant pose offset between their end-effectors. However,
additional constraints are easy to append. In the last part of the video in
[Ghazaei Ardakani et al., 2019], a virtual plane prevented the robot arms
from moving too low, thus avoiding collision. In [Ghazaei Ardakani et al.,
2018], it was shown how to add additional virtual constraints for redundant
robots, to allow for impacting all the degrees of freedom from either side.
Coupling between points other than the end-effectors, locking of certain
dimensions, and scaling can also be included to further assist an operator.
See, e.g., [Shiriaev et al., 2007; Freidovich et al., 2008] for previous research
on control design based on virtual constraints.
It would also be possible to use robot arms with less than six degrees of
freedom, though this would prevent movement in six degrees of freedom for
both end-effectors. Such a limitation would be reflected as haptic feedback
to the other side of the interface, just like joint and reach limits as well as
singularities were reflected in the current implementation.
In Chapter 7, joint torques from interaction forces were used as training
data for a recognition model. To gather the training data, the robot had
been programmed in advance to make a simple movement, so that the task
would finish. The interface in this chapter offers the possibility of demon-
strating a task directly, including interaction forces, without relying on prior
programming. It remains as future work to combine the functionalities in
Chapter 7 and this chapter into one program, allowing for an operator to
demonstrate contact forces, which can then be learned and subsequently
achieved autonomously by the robot. Further, it would be straightforward
to use this interface to demonstrate movement and represent it as DMPs.
In the context of DMPs in contact operations, an interesting direction of
research considers compliant DMPs, presented in [Denisa et al., 2016; Ba-
tinica et al., 2017].

135
Chapter 8. A Dual-Arm Haptic Interface for Task Demonstration

8.8 Conclusion
In this chapter, an immersive haptic interface for task demonstration using
two manipulators has been described. It allows for an operator to demon-
strate both movements and interaction forces in an intuitive way. The ap-
proach is based on introducing virtual constraints between the manipula-
tors, ensuring that the offset between the end-effectors remains constant.
Based on this, a control law coupling the motion of two nonlinear robotic
arms in task space has been derived. The interface supports using dissimilar
arms, mounted arbitrarily and starting in any configurations independently.
Interaction forces from the environment, as well as internal limitations such
as singularities, are naturally mediated to the operator as haptic feedback.
A video illustrating the interface is available online [Ghazaei Ardakani et
al., 2019].

Appendix A. Proof of Kinematic Constraint

In this appendix, a proof of Theorem 8.1 is presented. It was originally
presented in [Ghazaei Ardakani et al., 2018] where it had been formulated
by the first author, and is included here as well for the sake of completeness.
Consider first the orientation, and note that for a rotation matrix R the
inverse exists and is given by RT . Then the following relation holds.

RT1 R 2 = ∆ R \ R 2 = R 1 ∆ R (8.13)

Differentiating with respect to time yields

S (ω 2 ) R 2 = S (ω 1 ) R 1 ∆ R = S (ω 1 ) R 2 (8.14)

Here S (ω ) is the skew-symmetric matrix corresponding to the vector prod-

uct by the angular velocity ω , as explained in [Spong et al., 2006]. Hence,

S (ω 2 − ω 1 ) R 2 = 0 (8.15)

and multiplication with RT2 on both sides yields

S (ω 2 − ω 1 ) = 0 [ ω 2 − ω 1 = 0 (8.16)

Further, differentiating the position constraint in (8.1) yields ṗ 2 − ṗ 1 , and

hence the kinematic constraint is

ṗ 2 − ṗ 1 = 0 (8.17a)
ω2 − ω1 = 0 (8.17b)

136
 8.A Proof of Kinematic Constraint

Rewriting this using Jacobians and joint velocities yields

J2 (q 2 )q̇ 2 − J1 (q 1 )q̇ 1 = 0 (8.18)

or, equivalently,
Gq̇ = 0 (8.19)
This completes the proof.

137
9
Conclusion and Future
Research

This thesis presented research in the context of human–robot interaction.

It provided methods based on motion control, force control, and machine
learning to make robot programming more accessible, while enhancing re-
planning capabilities of robots. Each of the proposed methods was verified
through experiments on an industrial collaborative robot.
Robot programming by demonstration has been a theme throughout
the thesis. Lead-through programming was used to demonstrate desired
movements, and to modify faulty movements in existing programs thus sav-
ing time as compared to conventional programming. Temporal coupling for
DMPs was made realizable, allowing for robots to recover from unforeseen
disturbances and detours during task execution.
The proposed algorithms also enable demonstration of task-specific
forces. A haptic interface was developed, allowing for an operator to demon-
strate contact forces remotely, while experiencing the forces through feed-
back. Machine learning was used to recognize force transients based on
example data, to enable autonomous monitoring of manipulation tasks.
The proposed functionalities could be combined in future products, even
though they were evaluated separately in this thesis. For instance, the
haptic interface could be used to demonstrate movement and forces for a
given task. A DMP could then be used to execute the movement, and force
measurements could be used to monitor the progress and verify completion.
Even though this thesis mainly described applications for collaborative
robot arms, the insights could be used for control of other machines such as
mobile robots and unmanned aerial vehicles. For instance, DMPs were used
for control of a humanoid in [Ijspeert et al., 2002], a helicopter process in
[Perk and Slotine, 2006], and a legged robot in [Wensing and Slotine, 2017].
This thesis was limited to cases where the setting during the autonomous
operation was similar to that of the demonstration. An ability of robots

138
 Chapter 9. Conclusion and Future Research

to accomplish general manipulation tasks, without prior programming or

demonstration specific for a certain task, is still very distant. Because it
is not feasible to manually specify rules for general manipulation, machine
learning will probably continue to be an important field while approaching
this.
An interesting direction of future work would be to use demonstrated
programs to warm-start reinforcement learning, for instance to accomplish
assembly tasks. In such a setting, a reward is typically assigned to the
robot once the task is successfully finished. Detection of task completion
can be done manually, or through manually specified criteria. In this thesis,
autonomous detection of task completion, learned from previous demon-
strations, was developed. Such detection may be used to determine when to
assign rewards in reinforcement learning.
Further, DMPs could represent prior information, and exploration could
be achieved through small deviations from the preliminary movement. Dur-
ing the learning procedure, it would be likely that the assembly took more
time to accomplish as compared to a demonstrated movement. It would then
be useful to slow down the DMP evolution rate through temporal coupling.
Reinforcement learning and earlier DMP versions have been combined in,
e.g., [Stulp et al., 2012a; Stulp et al., 2012b; Li et al., 2018]. Further, the
reinforcement learning in [Levine and Koltun, 2013; Levine et al., 2015]
is a promising approach to achieve robotic assembly, though impractical
amounts of trials are required if prior information is not included. It would
therefore be valuable to explore combinations of these previous methods
and the methods proposed in this thesis.

139
Bibliography

Abadi, M., A. Agarwal, P. Barham, E. Brevdo, Z. Chen, C. Citro, G. S.

Corrado, A. Davis, J. Dean, M. Devin, et al. (2016). “TensorFlow:
large-scale machine learning on heterogeneous distributed systems”.
arXiv:1603.04467.
ABB Robotics (2018). YuMi — IRB 14000. url: http://new.abb.com/products/
robotics/yumi (visited on 2019-02-28).
ABB Robotics (2019a). ABB IRC5. url: http://new.abb.com/products/robotics/
controllers/irc5 (visited on 2019-02-28).
ABB Robotics (2019b). ABB RobotStudio. url: https://new.abb.com/products/
robotics/robotstudio/ (visited on 2019-03-02).
ABB Robotics (2019c). Application Manual — Controller Software IRC5.
url: https://us.v-cdn.net/5020483/uploads/editor/7y/p0a51v5nb8kg.pdf (vis-
ited on 2019-02-06).
Abu-Dakka, F. J., B. Nemec, J. A. Jørgensen, T. R. Savarimuthu, N. Krüger,
and A. Ude (2015). “Adaptation of manipulation skills in physical contact
with the environment to reference force profiles”. Autonomous Robots
39:2, pp. 199–217.
Aliaga, I., A. Rubio, and E. Sanchez (2004). “Experimental quantitative com-
parison of different control architectures for master-slave teleoperation”.
IEEE Transactions on Control Systems Technology 12:1, pp. 2–11.
Anderson, R. and M. Spong (1989). “Bilateral control of teleoperators with
time delay”. IEEE Transactions on Automatic Control 34:5, pp. 494–501.
Argall, B. D., S. Chernova, M. Veloso, and B. Browning (2009). “A survey of
robot learning from demonstration”. Robotics and Autonomous Systems
57:5, pp. 469–483.
Åström, K. J. and B. Wittenmark (2013). Computer-Controlled Systems:
Theory and Design. Courier Corporation, Mineola, NY.
Åström, K. J. and T. Hägglund (1995). PID Controllers: Theory, Design, and
Tuning. Instrument Society of Automation, Research Triangle Park, NC.

140
 Bibliography

ATI (2019). ATI: Industrial Automation. url: https://www.ati-ia.com/ (visited

on 2019-03-03).
Atkeson, C. G., A. W. Moore, and S. Schaal (1997). “Locally weighted learn-
ing for control”. In: Lazy Learning. Springer, Dordrecht, Netherlands,
pp. 75–113.
Bagge Carlson, F. (2016). DynamicMovementPrimitives.jl. Dept. Auto-
matic Control, Lund University. url: https://github.com/baggepinnen/
DynamicMovementPrimitives.jl (visited on 2019-02-28).
Bagge Carlson, F. (2019). Machine Learning and System Identification for
Estimation in Physical Systems. PhD thesis. TFRT–1122–SE, Dept. Au-
tomatic Control, Lund University, Lund, Sweden.
Bagge Carlson, F. and M. Haage (2017). “YuMi low-level motion guidance
using the Julia programming language and externally guided motion re-
search interface”. Technical Reports TFRT-7651. Dept. Automatic Con-
trol, Lund University, Lund, Sweden.
Bagge Carlson, F., M. Karlsson, A. Robertsson, and R. Johansson (2016).
“Particle filter framework for 6D seam tracking under large external
forces using 2D laser sensors”. In: IEEE/RSJ International Conference
on Intelligent Robots and Systems (IROS). October 9–14, Daejeon, South
Korea, pp. 3728–3734.
Bagge Carlson, F., A. Robertsson, and R. Johansson (2015). “Modeling and
identification of position and temperature dependent friction phenomena
without temperature sensing”. In: IEEE/RSJ International Conference
on Intelligent Robots and Systems (IROS). September 28–October 2,
Hamburg, Germany, pp. 3045–3051.
Batinica, A., B. Nemec, A. Ude, M. Rakovi, and A. Gams (2017). “Compliant
movement primitives in a bimanual setting”. arXiv:1707.04629.
Bezanson, J., A. Edelman, S. Karpinski, and V. B. Shah (2014). “Julia: a
fresh approach to numerical computing”. arXiv:1411.1607.
Bhat, S. P. and D. S. Bernstein (2000). “A topological obstruction to contin-
uous global stabilization of rotational motion and the unwinding phe-
nomenon”. Systems & Control Letters 39:1, pp. 63–70.
Bishop, C. (2007). Pattern Recognition and Machine Learning. Springer,
New York.
Björkelund, A., L. Edström, M. Haage, J. Malec, K. Nilsson, P. Nugues, S. G.
Robertz, D. Störkle, A. Blomdell, R. Johansson, M. Linderoth, A. Nilsson,
A. Robertsson, A. Stolt, and H. Bruyninckx (2011). “On the integration of
skilled robot motions for productivity in manufacturing”. In: IEEE/CIRP
International Symposium on Assembly and Manufacturing (ISAM). May
25–27, Tampere, Finland.

141
Bibliography

Blomdell, A., G. Bolmsjö, T. Brogårdh, P. Cederberg, M. Isaksson, R. Jo-

hansson, M. Haage, K. Nilsson, M. Olsson, T. Olsson, A. Robertsson, and
J. Wang (2005). “Extending an industrial robot controller: implementa-
tion and applications of a fast open sensor interface”. IEEE Robotics &
Automation Magazine 12:3, pp. 85–94.
Blomdell, A., I. Dressler, K. Nilsson, and A. Robertsson (2010). “Flexible ap-
plication development and high-performance motion control based on ex-
ternal sensing and reconfiguration of ABB industrial robot controllers”.
In: IEEE International Conference on Robotics and Automation (ICRA).
May 3–8, Anchorage, Alaska, pp. 62–66.
Boyle, M. (2017). “The integration of angular velocity”. Advances in Applied
Clifford Algebras 27:3, pp. 2345–2374.
Capurso, M., M. M. Ghazaei Ardakani, R. Johansson, A. Robertsson, and
P. Rocco (2017). “Sensorless kinesthetic teaching of robotic manipulators
assisted by observer-based force control”. In: IEEE International Con-
ference on Robotics and Automation (ICRA). May 29–June 3, Singapore,
pp. 945–950.
Chen, N., M. Karl, and P. van der Smagt (2016). “Dynamic movement prim-
itives in latent space of time-dependent variational autoencoders”. In:
IEEE International Conference on Humanoid Robotics. IEEE. November
15–17, Cancun, Mexico, pp. 629–636.
Chen, Y., J. Zhang, C. Yang, and B. Niu (2007). “The workspace mapping
with deficient-DOF space for the PUMA 560 robot and its exoskele-
ton arm by using orthogonal experiment design method”. Robotics and
Computer-Integrated Manufacturing 23:4, pp. 478–487.
Constantinescu, D. (2008). “The feel of virtual mechanisms”. In: Product
Engineering. Springer, Berlin Heidelberg, Germany, pp. 231–242.
Constantinescu, D., S. E. Salcudean, and E. A. Croft (2006). “Haptic manipu-
lation of serial-chain virtual mechanisms”. Journal of Dynamic Systems,
Measurement, and Control 128:1, pp. 65–74.
Cook, G. E., R. Crawford, D. E. Clark, and A. M. Strauss (2004). “Robotic
friction stir welding”. Industrial Robot: An International Journal 31:1,
pp. 55–63.
Crossley, M. D. (2006). Essential Topology. Springer, London, UK.
CVX Research Inc. (2019). CVX: Matlab software for disciplined convex
programming. url: http://cvxr.com/cvx/ (visited on 2019-01-07).
De Backer, J. (2014). Feedback Control of Robotic Friction Stir Welding.
PhD thesis. University West, Dept. Engineering Science, Trollhättan,
Sweden.

142
 Bibliography

De Backer, J. and G. Bolmsjö (2014). “Deflection model for robotic friction

stir welding”. Industrial Robot: An International Journal 41:4, pp. 365–
372.
Denisa, M., A. Gams, A. Ude, and T. Petri (2016). “Learning compliant move-
ment primitives through demonstration and statistical generalization”.
IEEE/ASME Transactions on Mechatronics 21:5, pp. 2581–2594.
Fox, E., M. I. Jordan, E. B. Sudderth, and A. S. Willsky (2009). “Sharing
features among dynamical systems with beta processes”. In: Advances
in Neural Information Processing Systems (NIPS). December 6–9, Van-
couver, Canada, pp. 549–557.
Franka Emika (2018). Panda. url: https://www.franka.de/panda/ (visited on
2018-10-30).
Freidovich, L., A. Robertsson, A. Shiriaev, and R. Johansson (2008). “Peri-
odic motions of the pendubot via virtual holonomic constraints: theory
and experiments”. Automatica 44:3, pp. 785–791.
García, J. G., A. Robertsson, J. G. Ortega, and R. Johansson (2006). “Gen-
eralized contact force estimator for a robot manipulator”. In: IEEE In-
ternational Conference on Robotics and Automation (ICRA). May 15–19,
Orlando, FL, pp. 4019–4024.
Gazebo (2018). Robot simulation made easy. url: http://gazebosim.org/ (vis-
ited on 2018-10-31).
Ghazaei Ardakani, M. M. (2016). On Trajectory Generation for Robots.
PhD thesis. TFRT–1116–SE, Dept. Automatic Control, Lund University,
Lund, Sweden.
Ghazaei Ardakani, M. M., M. Karlsson, K. Nilsson, A. Robertsson, and R.
Johansson (2018). “Master-slave coordination using virtual constraints
for a redundant dual-arm haptic interface”. In: IEEE/RSJ Interna-
tional Conference on Intelligent Robots and Systems (IROS). October
1–5, Madrid, Spain, pp. 8751–8757.
Ghazaei Ardakani, M. M., M. Karlsson, K. Nilsson, A. Robertsson, and R.
Johansson (2019). Master-slave coordination using virtual constraints
for a redundant dual-arm haptic interface. Dept. Automatic Control,
Lund University. url: https://www.youtube.com/watch?v=4u57bEA4bqg (visited
on 2019-01-10).
Giszter, S. F., F. A. Mussa-Ivaldi, and E. Bizzi (1993). “Convergent force
fields organized in the frog’s spinal cord”. Journal of Neuroscience 13:2,
pp. 467–491.
Glad, T. and L. Ljung (2000). Control Theory. CRC Press, Miami, FL.
Goodfellow, I., Y. Bengio, and A. Courville (2016). Deep Learning. (vis-
ited on 2019-02-28). MIT Press, Cambridge, MA. url: http://www.
deeplearningbook.org.

143
Bibliography

Grant, M. C. and S. P. Boyd (2008). “Graph implementations for nons-

mooth convex programs”. In: Recent Advances in Learning and Control.
Springer, Berlin Heidelberg, Germany, pp. 95–110.
Graves, A. (2012). “Neural networks”. In: Supervised Sequence Labelling
with Recurrent Neural Networks. Springer, Berlin Heidelberg, Germany,
pp. 15–35.
Haage, M., S. Profanter, I. Kessler, A. Perzylo, N. Somani, O. Sörnmo, M.
Karlsson, S. G. Robertz, K. Nilsson, L. Resch, and M. Marti (2016). “On
cognitive robot woodworking in SMErobotics”. In: ISR 2016: 47th Inter-
national Symposium on Robotics. VDE. June 21–22, Munich, Germany,
pp. 1–7.
Hatanaka, T., N. Chopra, and M. W. Spong (2015). “Passivity-based control
of robots: historical perspective and contemporary issues”. In: IEEE
Conference on Decision and Control (CDC). December 15–18, Osaka,
Japan, pp. 2450–2452.
Hatcher, A. (2002). Algebraic Topology. Cambridge University Press, Cam-
bridge, England.
Hokayem, P. F. and M. W. Spong (2006). “Bilateral teleoperation: an histor-
ical survey”. Automatica 42:12, pp. 2035–2057.
Hripcsak, G. and A. S. Rothschild (2005). “Agreement, the F-measure, and
reliability in information retrieval”. Journal of the American Medical
Informatics Association 12:3, pp. 296–298.
Huggett, S. and D. Jordan (2009). A Topological Aperitif. Springer, London,
UK.
Ijspeert, A., J. Nakanishi, and S. Schaal (2003). “Learning control policies
for movement imitation and movement recognition”. In: Neural Infor-
mation Processing System (NIPS). Vol. 15. December 5–10, Stateline,
Nevada, pp. 1547–1554.
Ijspeert, A. J., J. Nakanishi, H. Hoffmann, P. Pastor, and S. Schaal (2013).
“Dynamical movement primitives: learning attractor models for motor
behaviors”. Neural Computation 25:2, pp. 328–373.
Ijspeert, A. J., J. Nakanishi, and S. Schaal (2002). “Movement imitation
with nonlinear dynamical systems in humanoid robots”. In: IEEE In-
ternational Conference on Robotics and Automation (ICRA). Vol. 2. May
11–15, Washington, DC, pp. 1398–1403.
Isidori, A. (1995). Nonlinear Control Systems. Springer Verlag, London, UK.
Itoh, T., K. Kosuge, and T. Fukuda (2000). “Human-machine cooperative
telemanipulation with motion and force scaling using task-oriented vir-
tual tool dynamics”. IEEE Transactions on Robotics and Automation
16:5, pp. 505–516.

144
 Bibliography

Japkowicz, N. (2000). “The class imbalance problem: significance and strate-

gies”. In: International Conference on Artificial Intelligence. June 26–29,
Las Vegas, Nevada.
Japkowicz, N. and S. Stephen (2002). “The class imbalance problem: a
systematic study”. Intelligent Data Analysis 6:5, pp. 429–449.
Johansson, R. (1993). System Modeling and Identification. Prentice-Hall,
Englewood Cliffs, NJ.
Johansson, R., M. Annerstedt, and A. Robertsson (2009). “Stability of haptic
obstacle avoidance and force interaction”. In: IEEE/RSJ International
Conference on Intelligent Robots and Systems (IROS). October 11–15, St.
Louis, MO, pp. 3238–3243.
Johansson, R., K. Nilsson, and A. Robertsson (2015). “Force control”. In:
Handbook of Manufacturing Engineering and Technology. Springer Ver-
lag, London, UK, pp. 1933–1965.
Johansson, R. and A. Robertsson (2009). “Stability of robotic obstacle avoid-
ance and force interaction”. In: International Symposium on Robot Con-
trol (SYROCO’09). September 9–12, Gifu, Japan, pp. 561–566.
Kahn, D. W. (1975). Topology: An Introduction to the Point-Set and Algebraic
Areas. Courier Corporation, North Chelmsford, MA.
Karl, M., M. Soelch, J. Bayer, and P. van der Smagt (2016). “Deep variational
Bayes filters: unsupervised learning of state space models from raw
data”. arXiv:1605.06432.
Karlsson, F., M. Karlsson, B. Bernhardsson, F. Tufvesson, and M. Persson
(2015). “Sensor fused indoor positioning using dual band WiFi signal
measurements”. In: European Control Conference (ECC). July 15–17,
Linz, Austria, pp. 1669–1672.
Karlsson, M. (2016). Experimental evaluation of DMP perturbation recovery.
Dept. Automatic Control, Lund University. url: https://www.youtube.com/
watch?v=u8GwsSsL0TI (visited on 2019-02-28).
Karlsson, M. (2017a). Convergence of DMPs - experimental evaluation. Dept.
Automatic Control, Lund University. url: https://www.youtube.com/watch?
v=KRfjMXs4LjQ (visited on 2019-02-28).
Karlsson, M. (2017b). Detection of contact force transients from robot joint
torques. Dept. Automatic Control, Lund University. url: https://www.
youtube.com/watch?v=TE1q5lQr4nk (visited on 2018-12-18).
Karlsson, M. (2017c). DMP perturbation, simulation example. Dept. Auto-
matic Control, Lund University. url: https://gitlab.control.lth.se/cont-
mkr/dmp_perturbation_sim_example (visited on 2019-02-28).

145
Bibliography

Karlsson, M. (2017d). Modification of dynamical movement primitives. Dept.

Automatic Control, Lund University. url: https://www.youtube.com/watch?
v=q998JUwofX4&feature=youtu.be (visited on 2019-02-28).
Karlsson, M. (2017e). On Motion Control and Machine Learning for Robotic
Assembly. Licentiate Thesis, TFRT–3274–SE, Dept. Automatic Control,
Lund University, Lund, Sweden.
Karlsson, M. (2019). Temporally coupled DMPs in Cartesian space. Dept.
Automatic Control, Lund University. url: https://www.youtube.com/watch?
v=r0UMux3KH64 (visited on 2019-02-06).
Karlsson, M., F. Bagge Carlson, J. De Backer, M. Holmstrand, A. Roberts-
son, and R. Johansson (2016). “Robotic seam tracking for friction stir
welding under large contact forces”. In: 7th Swedish Production Sympo-
sium (SPS). October 25–27, Lund, Sweden.
Karlsson, M., F. Bagge Carlson, A. Robertsson, and R. Johansson (2017a).
“Two-degree-of-freedom control for trajectory tracking and perturbation
recovery during execution of dynamical movement primitives”. In: 20th
IFAC World Congress. July 9–14, Toulouse, France, pp. 1959–1966.
Karlsson, M. and F. Karlsson (2016). “Cooperative indoor positioning by
exchange of bluetooth signals and state estimates between users”. In:
European Control Conference (ECC). June 29–July 1, Aalborg, Denmark,
pp. 1440–1444.
Karlsson, M., A. Robertsson, and R. Johansson (2017b). “Autonomous in-
terpretation of demonstrations for modification of dynamical movement
primitives”. In: IEEE International Conference on Robotics and Automa-
tion (ICRA). May 29–June 3, Singapore, pp. 316–321.
Karlsson, M., A. Robertsson, and R. Johansson (2018a). “Convergence of
dynamical movement primitives with temporal coupling”. In: European
Control Conference (ECC). June 12–15, Limassol, Cyprus, pp. 32–39.
Karlsson, M., A. Robertsson, and R. Johansson (2018b). “Detection and
control of contact force transients in robotic manipulation without a force
sensor”. In: IEEE International Conference on Robotics and Automation
(ICRA). May 21–25, Brisbane, Australia, pp. 21–25.
Karlsson, M., A. Robertsson, and R. Johansson (2019). “Temporally cou-
pled dynamical movement primitives in Cartesian space”. Manuscript
prepared for submission to review for publication.
Keating, J. and W. Thach (1997). “No clock signal in the discharge of neurons
in the deep cerebellar nuclei”. Journal of Neurophysiology 77:4, pp. 2232–
2234.
Khalil, H. K. (2002). Nonlinear Systems. Prentice-Hall, Englewood Cliffs,
NJ.

146
 Bibliography

Khatib, O. (1986). “Real-time obstacle avoidance for manipulators and mo-

bile robots”. The International Journal of Robotics Research 5:1, pp. 90–
98.
Kingma, D. and J. Ba (2014). “Adam: a method for stochastic optimization”.
arXiv:1412.6980.
Kober, J., B. Mohler, and J. Peters (2008). “Learning perceptual coupling for
motor primitives”. In: IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS). September 22–26, Nice, France, pp. 834–
839.
Kock, S., T. Vittor, B. Matthias, H. Jerregard, M. Källman, I. Lundberg, R.
Mellander, and M. Hedelind (2011). “Robot concept for scalable, flexible
assembly automation: a technology study on a harmless dual-armed
robot”. In: International Symposium on Assembly and Manufacturing
(ISAM). IEEE. May 25–27, Tampere, Finland, pp. 1–5.
Koditschek, D. (1987). “Exact robot navigation by means of potential func-
tions: some topological considerations”. In: IEEE International Confer-
ence on Robotics and Automation (ICRA). Vol. 4. March 31–April 3,
Raleigh, NC, pp. 1–6.
Koditschek, D. E. (1988). “Application of a new Lyapunov function to global
adaptive attitude tracking”. In: 27th IEEE Conference on Decision and
Control. December 7-9, Austin, TX, pp. 63–68.
Kroemer, O., R. Detry, J. Piater, and P. J (2010). “Combining active learn-
ing and reactive control for robot grasping”. Robotics and Autonomous
Systems (RAS) 58:9, pp. 1105–1116.
Kröse, B. and P. van der Smagt (1993). An Introduction to Neural Networks.
Citeseerx, State College, PA.
KUKA (2018). LBR iiwa. url: https://www.kuka.com/en-de/products/robot-
systems/industrial-robots/lbr-iiwa (visited on 2018-10-30).
Kulvicius, T., K. Ning, M. Tamosiunaite, and F. Worgotter (2012). “Join-
ing movement sequences: modified dynamic movement primitives for
robotics applications exemplified on handwriting”. IEEE Transactions
on Robotics 28:1, pp. 145–157.
LaValle, S. M. (2006). Planning Algorithms. Cambridge University Press,
Cambridge, UK.
Lawrence, D. (1993). “Stability and transparency in bilateral teleoperation”.
IEEE Transactions on Robotics and Automation 9:5, pp. 624–637.
Lee, D. and M. Spong (2006). “Passive bilateral teleoperation with constant
time delay”. IEEE Transactions on Robotics 22:2, pp. 269–281.

147
Bibliography

Lee, D. and P. Y. Li (2003). “Passive bilateral feedforward control of linear

dynamically similar teleoperated manipulators”. IEEE Transactions on
Robotics and Automation 19:3, pp. 443–456.
Lee, D. and P. Y. Li (2005). “Passive bilateral control and tool dynamics
rendering for nonlinear mechanical teleoperators”. IEEE Transactions
on Robotics 21:5, pp. 936–951.
Lee, S. H., I. H. Suh, S. Calinon, and R. Johansson (2015). “Autonomous
framework for segmenting robot trajectories of manipulation task”. Au-
tonomous Robots 38:2, pp. 107–141.
Lehner, J. (1964). Discontinuous Groups and Automorphic Functions. Amer-
ican Mathematical Society, Providence, RI.
Levine, S. and V. Koltun (2013). “Guided policy search”. In: International
Conference on Machine Learning (ICML). June 16–21, Atlanta, GA,
pp. 1–9.
Levine, S., N. Wagener, and P. Abbeel (2015). “Learning contact-rich manip-
ulation skills with guided policy search”. In: IEEE International Con-
ference on Robotics and Automation (ICRA). May 26–30, Seattle, WA,
pp. 156–163.
Levy, A. (2002). Basic Set Theory. Dover Publications, Mineola, NY.
Li, P. Y. and R. Horowitz (1999). “Passive velocity field control of mechanical
manipulators”. IEEE Transactions on Robotics and Automation 15:4,
pp. 751–763.
Li, Z., T. Zhao, F. Chen, Y. Hu, C.-Y. Su, and T. Fukuda (2018). “Reinforce-
ment learning of manipulation and grasping using dynamical movement
primitives for a humanoidlike mobile manipulator”. IEEE/ASME Trans-
actions on Mechatronics 23:1, pp. 121–131.
Linderoth, M. (2013). On Robotic Work-Space Sensing and Control. PhD the-
sis. TFRT–1098–SE, Dept. Automatic Control, Lund University, Lund,
Sweden.
Linderoth, M., A. Stolt, A. Robertsson, and R. Johansson (2013). “Robotic
force estimation using motor torques and modeling of low velocity fric-
tion disturbances”. In: IEEE/RSJ International Conference on Intelli-
gent Robots and Systems (IROS). November 3–7, Tokyo, Japan, pp. 3550–
3556.
Lindsten, F., T. B. Schön, and M. I. Jordan (2013). “Bayesian semiparametric
Wiener system identification”. Automatica 49:7, pp. 2053–2063.
Ljung, L. (1987). System Identification: Theory for the User. Prentice-Hall,
Englewood Cliffs, NJ.
Lohmiller, W. and J. J. E. Slotine (1998). “On contraction analysis for non-
linear systems”. Automatica 34:6, pp. 683–696.

148
 Bibliography

Malmaud, J. (2017). A Julia wrapper for TensorFlow. url: https://github.

com/malmaud/TensorFlow.jl (visited on 2019-02-28).
Marchal, P. C., O. Sörnmo, B. Olofsson, A. Robertsson, J. G. Ortega, and R.
Johansson (2014). “Iterative learning control for machining with indus-
trial robots”. In: 19th IFAC World Congress. August 24–29, Cape Town,
South Africa, pp. 9327–9333.
Marzinotto, A., M. Colledanchise, C. Smith, and P. Ögren (2014). “Towards
a unified behavior trees framework for robot control”. In: IEEE Interna-
tional Conference on Robotics and Automation (ICRA). May 31–June 7,
Hong Kong, China, pp. 5420–5427.
Matsubara, T., S.-H. Hyon, and J. Morimoto (2011). “Learning parametric
dynamic movement primitives from multiple demonstrations”. Neural
Networks 24:5, pp. 493–500.
Matthias, B., S. Kock, H. Jerregard, M. Källman, and I. Lundberg (2011).
“Safety of collaborative industrial robots: certification possibilities for
a collaborative assembly robot concept”. In: International Symposium
on Assembly and Manufacturing (ISAM). IEEE. May 25–27, Tampere,
Finland, pp. 1–6.
Mattingley, J. and S. Boyd (2012). “CVXGEN: a code generator for embedded
convex optimization”. Optimization and Engineering 13:1, pp. 1–27.
Mayhew, C. G., R. G. Sanfelice, and A. R. Teel (2011). “Quaternion-based
hybrid control for robust global attitude tracking”. IEEE Transactions
on Automatic Control 56:11, pp. 2555–2566.
Meier, F., E. Theodorou, F. Stulp, and S. Schaal (2011). “Movement segmen-
tation using a primitive library”. In: IEEE/RSJ International Confer-
ence on Intelligent Robots and Systems (IROS). September 25–30, San
Francisco, CA, pp. 3407–3412.
Miyamoto, H., S. Schaal, F. Gandolfo, H. Gomi, Y. Koike, R. Osu, E. Nakano,
Y. Wada, and M. Kawato (1996). “A Kendama learning robot based on
bi-directional theory”. Neural Networks 9:8, pp. 1281–1302.
Murphy, K. P. (2012). Machine Learning: A Probabilistic Perspective. MIT
press, Cambridge, MA.
Mussa-Ivaldi, F. A. (1999). “Modular features of motor control and learning”.
Current Opinion in Neurobiology 9:6, pp. 713–717.
Nemec, B., L. lajpah, S. lajpa, J. Pikur, and A. Ude (2018). “An efficient PbD
framework for fast deployment of bi-manual assembly tasks”. In: IEEE-
RAS 18th International Conference on Humanoid Robots (Humanoids).
IEEE. November 6–9, Beijing, China, pp. 166–173.

149
Bibliography

Niekum, S., S. Osentoski, G. Konidaris, S. Chitta, B. Marthi, and A. G.

Barto (2015). “Learning grounded finite-state representations from un-
structured demonstrations”. The International Journal of Robotics Re-
search 34:2, pp. 131–157.
Norrlöf, M. (2002). “An adaptive iterative learning control algorithm with
experiments on an industrial robot”. IEEE Transactions on Robotics and
Automation 18:2, pp. 245–251.
Norrlöf, M. and S. Gunnarsson (2002). “Experimental comparison of some
classical iterative learning control algorithms”. IEEE Transactions on
Robotics and Automation 18:4, pp. 636–641.
Ögren, P. (2012). “Increasing modularity of UAV control systems using
computer game behavior trees”. In: AIAA Guidance, Navigation, and
Control Conference. August 13–16, Minneapolis, Minnesota, p. 4458.
Olofsson, B. (2015). Topics in Machining with Industrial Robot Manipu-
lators and Optimal Motion Control. PhD thesis. TFRT–1108–SE, Dept.
Automatic Control, Lund University, Lund, Sweden.
Olsson, T., J. Bengtsson, R. Johansson, and H. Malm (2002). “Force control
and visual servoing using planar surface identification”. In: IEEE In-
ternational Conference on Robotics and Automation (ICRA). May 11–15,
Washington, DC, pp. 4211–4216.
Pan, Z., J. Polden, N. Larkin, S. Van Duin, and J. Norrish (2010). “Re-
cent progress on programming methods for industrial robots”. In: 41st
International Symposium on Robotics (ISR). VDE. June 7–9, Munich,
Germany, pp. 1–8.
Panchapagesan, S., M. Sun, A. Khare, S. Matsoukas, A. Mandal, B.
Hoffmeister, and S. Vitaladevuni (2016). “Multi-task learning and
weighted cross-entropy for DNN-based keyword spotting”. Interspeech
2016, pp. 760–764.
Papageorgiou, D., A. Sidiropoulos, and Z. Doulgeri (2018). “Sinc-based dy-
namic movement primitives for encoding point-to-point kinematic be-
haviors”. In: IEEE/RSJ International Conference on Intelligent Robots
and Systems (IROS). IEEE. October 1–5, Madrid, Spain, pp. 8339–8345.
Park, D.-H., H. Hoffmann, P. Pastor, and S. Schaal (2008). “Movement
reproduction and obstacle avoidance with dynamic movement primitives
and potential fields”. In: Humanoids 2008—8th IEEE-RAS International
Conference on Humanoid Robots. December 1–3, Daejeon, South Korea,
pp. 91–98.
Pastor, P., H. Hoffmann, T. Asfour, and S. Schaal (2009). “Learning and gen-
eralization of motor skills by learning from demonstration”. In: IEEE In-
ternational Conference on Robotics and Automation (ICRA). May 12–17,
Kobe, Japan, pp. 763–768.

150
 Bibliography

Pastor, P., M. Kalakrishnan, F. Meier, F. Stulp, J. Buchli, E. Theodorou, and

S. Schaal (2013). “From dynamic movement primitives to associative
skill memories”. Robotics and Autonomous Systems 61:4, pp. 351–361.
Perk, B. E. and J. J. E. Slotine (2006). “Motion primitives for robotic flight
control”. arXiv:cs/0609140.
Persson, M., M. Karlsson, and F. Karlsson (2014). Method for improved in-
door positioning and crowd sourcing using PDR. International patent,
Publication number: WO 2015/118369 (A1), Applicant: Sony Corpora-
tion, Examined and granted under the patent cooperation treaty (PCT).
Sony. url: https://patents.google.com/patent/WO2015118369A1/en (visited on
2019-03-02).
Prada, M., A. Remazeilles, A. Koene, and S. Endo (2014). “Implementation
and experimental validation of dynamic movement primitives for object
handover”. In: IEEE/RSJ International Conference on Intelligent Robots
and Systems (IROS). September 14–18, Chicago, Illinois, pp. 2146–2153.
Rebelo, J. and A. Schiele (2012). “Master-slave mapping and slave base
placement optimization for intuitive and kinematically robust direct
teleoperation”. In: International Conference on Control, Automation and
Systems (ICCAS). October 17–21, Jeju Island, South Korea, pp. 2017–
2022.
Rethink Robotics (2018). Baxter. url: https://www.rethinkrobotics.com/baxter/
(visited on 2018-10-30).
Rodriguez, A., D. Bourne, M. Mason, G. F. Rossano, and J. Wang (2010).
“Failure detection in assembly: force signature analysis”. In: IEEE Con-
ference on Automation Science and Engineering (CASE). August 21–24,
Toronto, Canada, pp. 210–215.
Rojas, J., K. Harada, H. Onda, N. Yamanobe, E. Yoshida, K. Nagata,
and Y. Kawai (2012). “A relative-change-based hierarchical taxonomy
for cantilever-snap assembly verification”. In: IEEE/RSJ International
Conference on Intelligent Robots and Systems (IROS). October 7–12, Vil-
amoura, Portugal, pp. 356–363.
ROS (2018). rviz Package Summary. url: http://wiki.ros.org/rviz (visited
on 2018-10-31).
Rubinstein, R. Y. and D. P. Kroese (2013). The Cross-Entropy Method: A
Unified Approach to Combinatorial Optimization, Monte-Carlo Simula-
tion and Machine Learning. Springer Science & Business Media, New
York.
Rugh, W. J. (1996). Linear System Theory. Vol. 2. Prentice-Hall, Upper
Saddle River, NJ.
Sanderson, C. and R. Curtin (2016). “Armadillo: a template-based C++
library for linear algebra”. Journal of Open Source Software 1, p. 26.

151
Bibliography

SARAFun (2019). Horizon 2020 SARAFun — Smart Assembly Robot

with Advanced Functionalities. url: http://h2020sarafun.eu/ (visited on
2019-01-07).
Schaal, S. and C. G. Atkeson (1998). “Constructive incremental learning
from only local information”. Neural Computation 10:8, pp. 2047–2084.
Schaal, S., A. Ijspeert, and A. Billard (2003). “Computational approaches
to motor learning by imitation”. Philosophical Transactions of the Royal
Society of London B: Biological Sciences 358:1431, pp. 537–547.
Schaal, S., S. Kotosaka, and D. Sternad (2000). “Nonlinear dynamical sys-
tems as movement primitives”. In: IEEE International Conference on
Humanoid Robotics. September 7–8, Boston, MA, pp. 1–11.
Schön, T. B., F. Lindsten, J. Dahlin, J. Wågberg, C. A. Naesseth, A. Svens-
son, and L. Dai (2015). “Sequential Monte Carlo methods for system
identification”. In: IFAC Symposium on System Identification (SYSID).
October 19–21, Beijing, China, pp. 775–786.
Schwarz, A. S. (2013). Topology for Physicists. Vol. 308. Springer, Berlin
Heidelberg, Germany.
Shiriaev, A. S., L. B. Freidovich, A. Robertsson, R. Johansson, and A. Sand-
berg (2007). “Virtual-holonomic-constraints-based design of stable oscil-
lations of Furuta pendulum: theory and experiments”. IEEE Transac-
tions on Robotics 23:4, pp. 827–832.
Siciliano, B., L. Sciavicco, L. Villani, and G. Oriolo (2010). Robotics: Mod-
elling, Planning and Control. Springer Verlag, London, UK.
Slotine, J. J. E. and W. Li (1991). Applied Nonlinear Control. Vol. 199. 1.
Prentice-Hall, Englewood Cliffs, NJ.
Spong, M. W., S. Hutchinson, and M. Vidyasagar (2006). Robot Modeling
and Control. John Wiley & Sons, Hoboken, NJ.
Stavridis, S., D. Papageorgiou, and Z. Doulgeri (2017). “Dynamical system
based robotic motion generation with obstacle avoidance”. Robotics and
Automation Letters 2:2, pp. 712–718.
Stolt, A. (2015). On Robotic Assembly using Contact Force Control and
Estimation. PhD thesis. TFRT–1109–SE, Dept. Automatic Control, Lund
University, Lund, Sweden.
Stolt, A., F. B. Carlson, M. M. Ghazaei Ardakani, I. Lundberg, A. Roberts-
son, and R. Johansson (2015a). “Sensorless friction-compensated passive
lead-through programming for industrial robots”. In: IEEE/RSJ Inter-
national Conference on Intelligent Robots and Systems (IROS). Septem-
ber 28–October 2, Hamburg, Germany, pp. 3530–3537.

152
 Bibliography

Stolt, A., M. Linderoth, A. Robertsson, and R. Johansson (2011). “Force

controlled assembly of emergency stop button”. In: IEEE International
Conference on Robotics and Automation (ICRA). May 9–13, Shanghai,
China, pp. 3751–3756.
Stolt, A., M. Linderoth, A. Robertsson, and R. Johansson (2012a). “Adapta-
tion of force control parameters in robotic assembly”. In: 10th IFAC Sym-
posium on Robot Control. September 5–7, Dubrovnik, Croatia, pp. 561–
566.
Stolt, A., M. Linderoth, A. Robertsson, and R. Johansson (2012b). “Force
controlled robotic assembly without a force sensor”. In: IEEE Interna-
tional Conference on Robotics and Automation (ICRA). May 14–18, St.
Paul, MN, pp. 1538–1543.
Stolt, A., M. Linderoth, A. Robertsson, and R. Johansson (2015b). “Detection
of contact force transients in robotic assembly”. In: IEEE International
Conference on Robotics and Automation (ICRA). May 26–30, Seattle, WA,
pp. 962–968.
Stolt, A., A. Robertsson, and R. Johansson (2015c). “Robotic force estimation
using dithering to decrease the low velocity friction uncertainties”. In:
IEEE International Conference on Robotics and Automation (ICRA). May
26–30, Seattle, WA, pp. 3896–3902.
Stulp, F., J. Buchli, A. Ellmer, M. Mistry, E. A. Theodorou, and S. Schaal
(2012a). “Model-free reinforcement learning of impedance control in
stochastic environments”. IEEE Transactions on Autonomous Mental
Development 4:4, pp. 330–341.
Stulp, F. and S. Schaal (2011). “Hierarchical reinforcement learning with
movement primitives”. In: IEEE-RAS International Conference on Hu-
manoid Robots. October 26–28, Bled, Slovenia, pp. 231–238.
Stulp, F., E. A. Theodorou, and S. Schaal (2012b). “Reinforcement learning
with sequences of motion primitives for robust manipulation”. IEEE
Transactions on Robotics 28:6, pp. 1360–1370.
Stulp, F., E. Theodorou, M. Kalakrishnan, P. Pastor, L. Righetti, and S.
Schaal (2011). “Learning motion primitive goals for robust manipula-
tion”. In: IEEE/RSJ International Conference on Intelligent Robots and
Systems (IROS). September 25–30, San Francisco, CA, pp. 325–331.
Talignani Landi, C., F. Ferraguti, C. Fantuzzi, and C. Secchi (2018a). “A
passivity-based strategy for coaching in human–robot interaction”. In:
IEEE International Conference on Robotics and Automation (ICRA). May
21–25, Brisbane, Australia, pp. 3279–3284.

153
Bibliography

Talignani Landi, C., F. Ferraguti, C. Fantuzzi, and C. Secchi (2018b). A

passivity-based strategy for coaching in human–robot interaction. Uni-
versity of Modena and Reggio Emilia, ICRA Spotlight Video. url: https:
//www.youtube.com/watch?reload=9&v=ouEY7nrHAH4 (visited on 2018-12-16).
TensorFlow (2019). An open-source software library for machine intelligence.
url: https://www.tensorflow.org/ (visited on 2019-02-28).
Theorin, A. (2014). A Sequential Control Language for Industrial Automa-
tion. PhD thesis. TFRT–1104–SE, Dept. Automatic Control, Lund Uni-
versity, Lund, Sweden.
Thomas, U., S. Molkenstruck, R. Iser, and F. M. Wahl (2007). “Multi sensor
fusion in robot assembly using particle filters”. In: IEEE International
Conference on Robotics and Automation (ICRA). April 10–14, Roma,
Italy, pp. 3837–3843.
Thomas, U. and F. M. Wahl (2001). “A system for automatic planning,
evaluation and execution of assembly sequences for industrial robots”. In:
IEEE/RSJ International Conference on Intelligent Robots and Systems
(IROS). Vol. 3. October 29–November 3, Maui, HI, pp. 1458–1464.
Turro, N., O. Khatib, and E. Coste-Maniere (2001). “Haptically augmented
teleoperation”. In: IEEE International Conference on Robotics and Au-
tomation (ICRA). May 21–26, Seoul, South Korea, pp. 386–392.
Ude, A. (1999). “Filtering in a unit quaternion space for model-based object
tracking”. Robotics and Autonomous Systems 28:2-3, pp. 163–172.
Ude, A., B. Nemec, T. Petri, and J. Morimoto (2014). “Orientation in Carte-
sian space dynamic movement primitives”. In: IEEE International Con-
ference on Robotics and Automation (ICRA). May 31–June 7, Hong Kong,
China, pp. 2997–3004.
Wada, Y. and M. Kawato (2004). “A via-point time optimization algorithm for
complex sequential trajectory formation”. Neural Networks 17:3, pp. 353–
364.
Wadenbäck, M., M. Karlsson, A. Heyden, A. Robertsson, and R. Johansson
(2017). “Visual odometry from two point correspondences and initial au-
tomatic camera tilt calibration”. In: 12th International Joint Conference
on Computer Vision, Imaging and Computer Graphics Theory and Appli-
cations, Volume 6. VISIGRAPP. February 27–March 1, Porto, Portugal,
pp. 340–346.
Wahrburg, A., S. Klose, D. Clever, T. Groth, S. Moberg, J. Styrud, and
H. Ding (2018). “Modeling speed-, load-, and position-dependent fric-
tion effects in strain wave gears”. In: IEEE International Conference
on Robotics and Automation (ICRA). May 21–25, Brisbane, Australia,
pp. 2095–2102.

154
 Bibliography

Wahrburg, A., S. Zeiss, B. Matthias, and H. Ding (2014). “Contact force

estimation for robotic assembly using motor torques”. In: IEEE Inter-
national Conference on Automation Science and Engineering (CASE).
August 18–22, Taipei, Taiwan, pp. 1252–1257.
Wen, J. T. and S. H. Murphy (1990). PID Control for Robot Manipulators.
Rensselaer Polytechnic Institute.
Wen, J. T. and S. H. Murphy (1991). “Stability analysis of position and force
control for robot arms”. IEEE Transactions on Automatic Control 36:3,
pp. 365–371.
Wensing, P. M. and J. J. E. Slotine (2017). “Sparse control for dynamic
movement primitives”. In: The 20th World Congress of the International
Federation of Automatic Control (IFAC). July 9-14, Toulouse, France,
pp. 10531–10538.
Westervelt, E. R., J. W. Grizzle, C. Chevallereau, J. H. Choi, and B. Morris
(2007). Feedback Control of Dynamic Bipedal Robot Locomotion. CRC
Press, Boca Raton, FL.
Yang, D., Q. Lv, G. Liao, K. Zheng, J. Luo, and B. Wei (2018). “Learning from
demonstration: dynamical movement primitives based reusable suturing
skill modelling method”. In: Chinese Automation Congress (CAC). IEEE.
November 30–December 2, Xi’an, China, pp. 4252–4257.

155