Rapid Reinforcement Learning for Reactive Control Policy Design in Autonomous Robots

This paper describes work in progress on a neural-based reinforcement learning architecture for the design of reactive control policies for an autonomous robot. Reinforcement learning techniques allow a programmer to specify the control program at the level of the desired behavior of the robot, rather than at the level of the program that generates the behavior. In this paper, we explicitly begin to address the issue of state representation which can greatly affect the system's ability to learn quickly and to apply what has already been learned to novel situations. Finally, we demonstrate the architecture as applied towards a real robot that is learning to move safely about its environment.