強化学習を用いた二足歩行ロボットの行動選択の最適化

概要

論文の詳細を見る
Development of the motion selection rule for a biped walking robot by human is difficult because it needs to consider a possibility of falling unlike a wheel type robot. In this paper, we propose a method that a motion rule is automatically developed for a biped walking robot by the reinforcement learning. We use the biped walking robot which has a visual sensor on the head. The task of the robot is recognizing a target by the visual sensor, and approaching the target by walking, and kicking the target by the toe. In this study, the motion rule consists of a ball tracking rule and a motion selection rule. These rules are learned separately first, and then they are learned simultaneously. Experimental results show that the ball tracking rule is constructed by the reinforcement learning. However, constructing the motion selection rule fails by the reinforcement learning because the environmental noise causes instability of robot behaviors.This problem is able to improve by generating robot motions robustly to the environmental noise.