Study of a Multi-Robot Collaborative Task through Reinforcement Learning - Foundations on Natural and Artificial Computation - page 184

Study of a Multi-Robot Collaborative Task
through Reinforcement Learning

Juan Pereda 1 , Manuel Martín-Ortiz 1 ,JavierdeLope 2 , and Félix de la Paz 3

1 ITRB Labs
Research, Technology Development and Innovation, S.L.
2 Computational Cognitive Robotics
Universidad Politécnica de Madrid
3 Dept. Artificial Intelligence
UNED
{juan.pereda,manuel.martin}@itrblabs.eu, [e-mail],
[e-mail]

Abstract. A open issue in multi-robots systems is coordinating the col-
laboration between several agents to obtain a common goal. The most
popular solutions use complex systems, several types of sensors and com-
plicated controls systems. This paper describes a general approach for
coordinating the movement of objects by using reinforcement learning.
Thus, the method proposes a framework in which two robots are able
to work together in order to achieve a common goal. We use simple
robots without any kind of internal sensors and they only obtain in-
formation from a central camera. The main objective of this paper is
to define and to verify a method based on reinforcement learning for
multi-robot systems, which learn to coordinate their actions for achieving
common goal.

1 Introduction

One of the major problems that we have to deal is to achieve that robots col-
laborate in order to obtain a common objective. Part of the approach present
separate mechanisms for controlling the robot and for defining the coordination
process. This kind of solutions has the problem that the robot control interferes
in the solution adopted by the team.
Matarić [4] considers that some problems of coordination multi-robot systems
is an interested area of research. She present some approach for multi-robot
collaborative situations using reinforcement learning for collective foraging [5].
Some modern studies, use hybrid controllers based in evolutionary algorithms
and reinforcement learning for autonomous robots [7]. This approach use the
optimization force of the evolutionary algorithm and the eciency of the rein-
forcement learning to resolve the navigation problem in environment with ob-
stacles. This kind of solutions are designed for single robots but do not analyze
the relation between different robots to do collaborative tasks.