Cooperative Behavior Generation of Soccer Humanoid Robots by Multi-Agent Reinforcement Learning

— Physical Simulation Environment for Speeding Up Learning —

Abstract

The purpose of this research is to acquire action decision strategies for multiple soccer humanoid robots using multi-agent reinforcement learning. In this paper, we report the results of experiments in which we constructed a simulation environment for learning multiple robots and soccer games. Since the humanoid robots require more time to be processed by the simulator, we tried to shorten the learning time by dividing it into phases. As a result of the learning, we confirmed that the humanoid robot scored goals in a soccer game, and cooperative behavior within a team such as role switching was observed. The time required to run the experiment was significantly reduced by using pre-learning to shorten the time required, and more efficient learning became possible.