Learning to Avoid Obstacles Through Reinforcement

José del R. Millán, Carme Torras

Abstract: Motion planning and control involve mainly symbolic and subsymbolic processing, respectively, as does the learning of these capabilities. This paper focusses on a motion control aspect, namely, the learning of obstacle-avoidance abilities. We present a reinforcement-based connectionist system able to find and learn obstacle-avoiding paths for a mobile robot in a non-maze-like 2D environment. In the conclusions section, some directions on how to interface the subsymbolic system developed with a symbolic path planner are provided.