See discussions, stats, and author profiles for this publication at: https://www.researchgate.

net/publication/325671423

A Data-Driven Path Planner for Small Autonomous Robots Using Deep

Regression Models

Chapter in Lecture Notes in Computer Science · June 2018

DOI: 10.1007/978-3-319-93803-5_56

CITATIONS READS

3 147

3 authors, including:

Fredy Martínez M. Arbulu

District University of Bogotá Corporación Universitaria del Huila
120 PUBLICATIONS 394 CITATIONS 56 PUBLICATIONS 253 CITATIONS

SEE PROFILE SEE PROFILE

All content following this page was uploaded by M. Arbulu on 15 February 2019.

The user has requested enhancement of the downloaded file.

 A Data-Driven Path Planner for Small
Autonomous Robots Using Deep
Regression Models

Fredy Martı́nez1(B) , Angelica Rendón1 , and Mario Arbulú2

1
District University Francisco José de Caldas,
Bogotá D.C., Colombia
[email protected], [email protected]
2
Corporación Unificada Nacional de Educación Superior,
Bogotá D.C., Colombia
mario [email protected]
http://www.udistrital.edu.co

Abstract. This paper proposes a navigation scheme for robots in indoor

environments which uses real-time analysis of raw data captured by an
infra-red sensor. The infra-red sensor captures distance data in real time
producing a large database that the robot analyzes according to previ-
ous experiences to directly define its movement strategy. Observing the
dependence of data with the topology of the environment, this research
proposes models based on a Long Short-Term Memory (LSTM) net-
work. We demonstrate not only that the navigation scheme works suc-
cessfully on a real prototype, but also that the strategy can be used in
new unknown navigation environments with a high success rate.

Keywords: Autonomous · Big data · Data-driven · Sensor

Motion planner

1 Introduction

When a person is in a shopping mall for the first time, looking at the informa-
tion in the environment can easily locate places of interest. By observing signs,
symbols, or colors this person can find the food area, a particular store or the
restroom. This capacity is also observed in other animals such as ants, fish and
birds [5]. This ability to perceive and understand new environments by extract-
ing important or relevant information from them is possible due to the previous
information stored by the organism in its knowledge base [12]. This knowledge
base is appropriate for the organism throughout its life through experiences in
other environments, which in new environments allows it to make rational deci-
sions [9].
In the case of robots, the objective is to make the robot to move in the
right direction. Traditionally this has been solved in known environments with
c Springer International Publishing AG, part of Springer Nature 2018
Y. Tan et al. (Eds.): DMBD 2018, LNCS 10943, pp. 596–603, 2018.
https://doi.org/10.1007/978-3-319-93803-5_56
 A Data-Driven Path Planner for Small Autonomous Robots 597

an analysis of the topology of the environment in conjunction with a geometric

strategy and robot odometry [6,8], and in unknown environments through reac-
tive strategies [3,11]. These two strategies can be combined into one according to
the pattern of behavior observed in living beings. Therefore, taking living beings
as an example, it is possible to implement a similar system of decision making.
A knowledge base for the robot can contain key information for its navigation
in unknown environments [1]. This information is read from the environment by
means of sensors, which depending on the quantity and resolution can provide
an enormous amount of information. From this large database the robot can
search for specific information that allows it to make rational decisions about
movement and navigation.
This robot programming strategy can be structured in three specific steps:

1. Navigation in unknown environments to sense the characteristics of the envi-

ronment (mapping).
2. Identification in the sensed data of the relevant characteristics for decision
making in navigation tasks (analysis).
3. Identification of the relevant characteristics in navigation tasks on new envi-
ronments (autonomous navigation).

Traditional navigation schemes require a global mapping of the environment

[4,13]. However, in the proposed approach this is not necessary, since the robot
takes the decisions of movement not based on the design of the environment, but
what it can perceive in it. The strategy is coordinated by information, which is
read locally by the robot, this makes it possible for the robot to navigate in
previously unseen environments.
Making movement decisions directly from raw data is a very complex task.
Among the most appropriate tools for the treatment of this information, Deep
Neural Networks (DNNs) are among the most promising because of their abil-
ity to model complex dependencies [2]. DNNs are widely used to solve a large
number of complex problems in machine learning, in particular to solve pattern
recognition and classification problems.
In the same way, it is possible to consider different tasks simultaneously. From
the same database it is possible to identify information for navigation, but it is
also possible to identify information for other secondary tasks such as battery
recharging, communication with detected elements, etc. Each of these tasks can
be analyzed as a different control surface, and for the overall control scheme it
is possible to use a hybrid scheme that binds all surfaces together [7,10].
We propose the use of a data-driven motion planner for autonomous robots
that besides being reliable, allows reducing the amount of data manually adjusted
by the designer, in particular, working in unknown environments. For this, we
analyze the captured sensor data for specific environmental conditions, identi-
fying characteristics that can then be interpreted by the robot, and from them
define its movement strategy.
The paper is organized as follows. In Sect. 2 presents a description of the
problem. Section 3 describes the strategies used to analyze raw data and generate
598 F. Martı́nez et al.

data models for specific environmental characteristics. Section 4 introduces some

results obtained with the proposed strategy. Finally, conclusion and discussion
are presented in Sect. 5.

2 Problem Statement
Currently there are many robot motion schemes that allow from sensor readings
to coordinate safe motion in an environment to the desired target position. How-
ever, most of these schemes separate movement policies from data processing.
The approach we propose seeks a unified solution to the problem.
Let W ⊂ R2 be the closure of a contractible open set in the plane that has a
connected open interior with obstacles that represent inaccessible regions. Let O
be a set of obstacles, in which each O ⊂ O is closed with a connected piecewise-
analytic boundary that is finite in length. Furthermore, the obstacles in O are
pairwise-disjoint and countably finite in number. Let E ⊂ W be the free space
in the environment, which is the open subset of W with the obstacles removed.
Let us assume an agent in this free space. The agent knows the environment
E from observations, using sensors. These observations allow him to build an
information space I. A information mapping is of the form:

q : E −→ S (1)

where S denote an observation space, constructed from sensor readings over time,
i.e., through an observation history of the form:

õ : [0, t] −→ S (2)

The interpretation of this information space, i.e., I × S −→ I, is that which

allows the agent to make decisions. The problem can be expressed as the search
for a function u for specific environment conditions from a set of sensed data in
the environment y ⊂ S and a target function g.

f : y × g −→ u (3)

3 Methodology
Our control scheme seeks to develop a single global control strategy based on
raw sensor data. The function between input data and movement policies cor-
responds to a complex model. The problem of model development for specific
environmental conditions (obstacle, clear path, and target position) is analyzed
as a regression problem. Throughout the tests with the robot in known environ-
ments, a sequence of temporary data is produced that must allow to estimate the
type of characteristics in the environment for any similar situation. The sequence
of data for an obstacle is shown in Fig. 1. The figure, in addition to detecting
the obstacle, shows a growing trend that indicates some type of memory at the
 A Data-Driven Path Planner for Small Autonomous Robots 599

Fig. 1. Normalized reading of infrared sensors for an obstacle (blue), LSTM model
(red), and predictions for the training dataset (green). (Color figure online)

output. Therefore, we apply a recurrent neural network type LSTM to develop

the behavioral model.
The reference models for each characteristic of the environment are used as
a reference for the calculation of similarity with respect to the data captured
by the robot during navigation. The comparison is made against all channels,
i.e. each of the nine channels is compared with each of the nine channels of the
model (81 comparisons in each iteration).
Similarity is the measure of how alike two data objects are. It is a power-
ful tool when we want to discover patterns in a great data set. Normally it is
described as a distance between features of an object. If this distance is small,
there will be high degree of similarity. And if instead the distance is large, there
will be low degree of similarity.
The comparison of similarity between the reference models and the raw data
captured by the set of sensors can be done using different metrics. Each metric
corresponds to a strategy that produces a quantitative assessment of similarity.
There is no a right similarity metric, but a set of metrics that are appropriated
for a particular application. In our problem we have selected metrics about the
two-dimensional space of the model that can be represented as an image. In this
type of application, different metrics can be classified into three categories:
600 F. Martı́nez et al.

– Image based
– Histrogram based
– Based on similarity of objects contained in images.

We use image based in this research. This category focuses on the dif-
ference of values of corresponding pixels (pixel based). Euclidean distance is
the most commonly-used distance measure, and is particularly useful when
the data are dense or continuous. It is the straight-line distance between two
points in Euclidean space. In Cartesian coordinates, if p = (p1 , p2 , · · · , pn ) and
q = (q1 , q2 , · · · , qn ) are two points in Euclidean n-space, then the distance d
from p to q, is given by (Eq. 4):


2 2 2
d (p, q) = d (q, p) = (q1 − p1 ) + (q2 − p2 ) + · · · + (qn − pn )

n
2
(4)
= (qi − pi )
i=1

Similarly, the distance between two imaging functions, f and g, which provide
a value for each pixel of the n × m pixels image, is defined as (Eq. 5):


 n n
d (f, g) = 
2
(f (i, j) − g (i, j)) (5)
i=1 j=1

This difference can be evaluated by other distances, such as quadratic, Man-

hattan, Chi-Square or Bhattacharyya distance. The distance used as metric for
the calculation of similarity was Chi-Square due to its better performance.
Control policies were simplified as a sequence of actions specific to the identi-
fied environmental characteristics. For example, an obstacle to the front activates
the Evation Policy, which in this case consists of stopping, turning a random
angle in a random direction, and finally moving forward. Each characteristic of
the environment has a control policy, which is adjusted according to how the
characteristic is detected by the sensors.

4 Results
We have tested our proposed algorithm on a dataset generated by a 45 cm ×
61 cm robot (ARMOS TurtleBot 1 equipped with a set of nine infrared sensors
and one DragonBoard 410C development board, Fig. 2). The robot has nine
uniformly distributed infrared sensors around its vertical axis, each with a range
of 0.2 to 0.8 cm. The data matrix delivered by the sensors corresponds to nine
standard columns from 0 to 1 with reading intervals between 900 ms.
The performance of the LSTM models are evaluated using cross validation.
To do this we separated each dataset in an orderly manner, creating a training
set and a test set. For training we use 70% of the data, and we use the rest to
test the model. The network has a visible layer with one input, a hidden layer
 A Data-Driven Path Planner for Small Autonomous Robots 601

Fig. 2. Top view of the robot and distribution of the distance sensors.

Fig. 3. ARMOS TurtleBot 1 equipped with a set of nine infrared sensors and one
DragonBoard 410C development board.

with eight LSTM blocks or neurons, and an output layer that makes a single
value prediction. The default sigmoid activation function is used for the LSTM
blocks. The network is trained for 100 epochs and a batch size of one is used.
The model fit code was written in Python using Keras. We setup a SciPy work
environment with Pandas support. Models were evaluated using Keras 2.0.5,
TensorFlow 1.2.0 and scikit-learn 0.18.2.
602 F. Martı́nez et al.

The navigation scheme has been tested in laboratory with different configura-
tions on a 6 × 6.5 m environment. We have performed more than 100 tests, 85%
of them completely successful, that is, the robot managed to navigate the envi-
ronment to the final destination (Fig. 3). In cases of failure, the most frequent
problem is the robot’s inability to find the target point satisfactorily. At this
point it is necessary to perform a deeper analysis of the algorithm performance,
as well as more reliable ways to detect the target point.

5 Conclusions
The aim of this paper is to present a new strategy for the design of a global data-
driven motion planner for small mobile robots in indoor environments. The goal
is to develop an autonomous navigation strategy for small robots for unknown
and dynamic environments, such as those found in industrial plants or com-
mercial establishments, that is, in service applications with direct interaction
with humans. From local readings of a set of infrared sensors, our proposal is
able to define a sequence of movements that allows the robot to navigate the
environment and reach the target position. Our strategy uses similarity testing
against LSTM recurrent networks based models. The similarity metric used is
supported in two-dimensional images built from the captured data, measuring
the distances with Chi-Square. We performed a wide range of laboratory tests on
a real prototype which showed successful navigation in unknown environments.

Acknowledgments. This work was supported by the District University Francisco

José de Caldas and Corporación Unificada Nacional de Educación Superior. The views
expressed in this paper are not necessarily endorsed by District University or Cor-
poración Unificada Nacional de Educación Superior. The authors thank the research
group ARMOS for the evaluation carried out on prototypes of ideas and strategies.

References
1. Bao, J., Shuguo, W., Fu, Y.: Sensor-based motion planning for robot manipulators
in unknown environments. In: IEEE/RSJ International Conference on Intelligent
Robots and Systems (IROS 2005), pp. 199–204 (2005)
2. Bazrafkan, S., Corcoran, P.M.: Pushing the AI envelope: merging deep networks to
accelerate edge artificial intelligence in consumer electronics devices and systems.
IEEE Consum. Electron. Mag. 7(2), 55–61 (2018)
3. Benitez, J., Parra, L., Montiel, H.: Diseño de plataformas robóticas diferenciales
conectadas en topologı́a Mesh para tecnologı́a Zigbee en entornos cooperativos.
Tekhnê 13(2), 13–18 (2016)
4. Chapoulie, A., Rives, P., Filliat, D.: Appearance-based segmentation of
indoors/outdoors sequences of spherical views. In: IEEE/RSJ International Con-
ference on Intelligent Robots and Systems (IROS 2013), pp. 1946–1951 (2013)
5. Gao, J., Wang, L., Yang, X.: A survey of desert ant navigation. In: IEEE Interna-
tional Conference on Information Acquisition, pp. 1–6 (2005)
6. González, O., Acero, D.: Robot semi-autónomo para el transporte de pacientes con
movilidad reducida. Tekhnê 13(2), 49–63 (2016)
 A Data-Driven Path Planner for Small Autonomous Robots 603

7. Grandi, R., Falconi, R., Melchiorri, C.: Coordination and control of autonomous
mobile robot groups using a hybrid technique based on particle swarm optimization
and consensus. In: IEEE International Conference on Robotics and Biomimetics
(ROBIO 2013), pp. 1514–1519 (2013)
8. Hadda, I., Knani, J.: Global mapping and localization for mobile robots using stereo
vision. In: 10th International Multi-Conference on Systems, Signals & Devices (SSD
2013), pp. 1–6 (2013)
9. Kasaei, H., Oliveira, M., Hyun, G., Seabra, L., Tome, A.: An adaptive object
perception system based on environment exploration and Bayesian learning. In:
Autonomous Robot Systems and Competitions (ICARSC 2015), pp. 221–226
(2015)
10. Martı́nez, F.: Algunas aplicaciones de la lógica difusa en el control de convertidores
DC/DC. Distrital University Francisco José de Caldas, CIDC (2011). ISBN 978-
958-8723-08-2
11. Rendón, A.: Evaluación de estrategia de navegación autónoma basada en compor-
tamiento reactivo para plataformas robóticas móviles. Tekhnê 12(2), 75–82 (2015)
12. Schott, C., Padmanabha, M., Robler, M., Frob, D., Heinkel, U.: Position estima-
tion with Bayesian filters by using 3-dimensional environment models. In: 14th
Workshop on Positioning, Navigation and Communications (WPNC 2017), pp.
1–6 (2017)
13. Tian, W.: The research into methods of map building and path planning on mobile
robots. In: IEEE 2nd Information Technology, Networking, Electronic and Automa-
tion Control Conference (ITNEC 2017), pp. 1087–1090 (2017)

View publication stats