2P1-F22 強化学習を用いた大車輪運動の報酬と運動の解明

概要

論文の詳細を見る
In our laboratory, we have studied a giant swing robot using Q learning, which is one of the reinforcement learning. In the previous study, we succeeded dynamic tasks that acquired giant swing motion by devising rewards. Most of the rewards for the tasks were unable to acquire the giant swing motion, but its reasons are unknown. Therefore, this study attempts to clarify the causes by investigating the several rewards and examining whether the specialized learning in the each motion is possible by giving the rewards of only forward rotation or backward rotation. Consequently, the defective sensation problem of the reinforced learning is deeply concerned with success of giant swing motion and for two motions, not all the generalization of learning is possible because of the difference of behavioral characteristic.