Rho-learning: a robotics oriented reinforcement learning algorithm

Abstract

We present a new reinforcement learning system more suitable to be used in robotics than existing ones. Existing reinforcement learning algorithms are not specifically tailored for robotics and so they do not take advantage of the robotic perception characteristics as well as of the expected complexity of the task that robots are likely to face. In a robot, the information about the environment comes from a set of qualitatively different sensors and in the main par of tasks small subsets of these sensors provide enough information to correctly predict the effect of actions. Departing from this analysis, we outline a new reinforcement learning system that aims at determining relevant subsets of sensors for each action and we present an algorithm that partially implements this new reinforcement learning architecture. Results of the application of the algorithm to the problem of learning to walk with a six legged robot are presented and compared with a well known reinforcement learning algorithm (Q-learning) showing the advantages of our approach.