Performance Evaluation of the Dyna-Q algorithm for Robot Navigation

This paper tackles the path planning of a team of robots moving in a partially known environment, with static obstacles within it. Given the initial positions of the set of robots and a set of destinations, the robots should safety reach them avoiding the obstacles. Our approach is based on Reinforcement Learning, which is suited to partial knowledge of the environment and its dynamics. We use specifically the Dyna-Q algorithm (based on the Dyna architecture), including Planning and Reinforcement Learning, initially developed to a single robot case, and extended here to a multi-robot system. We analyze the problem with extensive and thorough simulations, for single and multi-robot systems, using the Robot Motion Toolbox, with the goal of characterizing the behavior of the Dyna-Q algorithm with respect to its main parameters.