Study of a Multi-Robot Collaborative Task through Reinforcement Learning - Foundations on Natural and Artificial Computation - page 185

Other studies like [1], present new approach for the reinforcement learning
using k-NN method to improve converge and memory use, two of the critical
points in this kind of algorithm.
Normally, the different studies present reinforcement learning only for indi-
vidual learning. Some theories propose co-learning solutions for the multi-robots
environments [3] based on the diculties of scaling up the individual solutions to
multi-robots one. Other approach with multi-robot systems [2] presents solutions
with two learning strategies: one individual and another one for the collaborative
tasks.
In this paper, we define a method for solving collaborative tasks by means
of reinforcement learning. In the sequel we propose the basics and expose the
experimental results.
2 Collaborative Object Movement
2.1 Description of the Experimental Task
The problem we want to resolve is to move a rectangular object from an ini-
tial random position to a goal position that we define at the beginning of the
experiment.
We define a environment without obstacles, with two robots and one rect-
angular object that the robots have to move from initial point to the goal. We
use two robots that do not have any kind of sensory input the system know
the robot position due to we are using a cenital camera. With this camera, the
system knows where the object in and the other members of the team are. At
this stage, we do not employ the effects of the natural forces. We consider just
four basic movements for the robots: advance, turn left, turn right and go back.
They do not have any kind of arm or external actuator to move the objects, they
only can push it.
Figure 1 shows a setup representation where the two robots (represented by
two circles) move the object (represented like a stick) to a predefined goal (rep-
resented with a rectangles). The triangular object represents the camera that
provides information of the environment to the robots.
Fig. 1. Experiment setup