LEARNING TO COORDINATE IN TOPOLOGICAL NAVIGATION TASKS

In this paper we address multi-robot coordinated navigation from a topological perspective. The adopted topological representation of the environment leads naturally to a Markov game model to describe the interaction of the multiple robots in the environment. In this setting, we combine the Q-learning algorithm with a powerful coordination mechanism (biased adaptive play). We show that this combined algorithm, coordinated Q-learning, converges to an optimal, coordinated solution for the navigation problem. This implies that the team of robots is able to coordinate without using any communication protocol to enforce the coordination. We illustrate our method in some simple navigation tasks.