Learning Reachable Regions and Interception Trajectories by Reinforcement in Pursuit/Evasion Tasks

摘要

This paper addresses the problem of learning reachable regions and interception trajectories, in the context of two vision-guided mobile robots. A reinforcement-based learning algorithm is proposed to acquire the shape of the Dynamic Map representing the kinetic capabilities of the mobile robot. We choose to represent the mapping between states and actions by a neural network. The mapping thus learned does not only indicate sequences of actions for the pursuer, but also indentifies unreachable locations; one in which the pursuer has direct access to the relative pose of the two robots, and the other in which vision is the only source information available. Experimental results are presented with car-like robots using a simulated vision sensor.