Learning coordinated maneuvers in complex environments: a sumo experiment

Abstract

This paper describes a dual-agent system capable of learning eye-body-coordinated maneuvers in playing a sumo contest. The two agents rely on each other by either offering feedback information on the physical performance of a certain selected maneuver or giving advice on candidate maneuvers for an improvement over the previous performance. At the core of this learning system lies in a multi-phase genetic-programming approach that is aimed to enable the player to gradually acquire sophisticated sumo maneuvers. As illustrated in the sumo learning experiments involving opponents of complex shapes and sizes, the proposed multi-phase learning allows the development of specialized strategic maneuvers based on the general ones, and hence demonstrates the efficiency of maneuver acquisition. We provide details of the problem addressed and the implemented solutions concerning a mobile robot for performing sumo maneuvers and the computational assistant for coaching the robot. In addition, we show the actual performances of the sumo agent, as a result of coaching, in dealing with a number of difficult sumo situations.