2A1-E10 学習に基づく2自由度1脚跳躍ロボットの運動制御

概要

論文の詳細を見る
The aim of this study is to evaluate the effectiveness of "Supervised Learning" and "Reinforcement Learning" in terms of the acquisition of a balancing control rule for our 1-legged hopping robot with 2 DOF. In the balancing control experiments for supervised learning, a sigmoid neural network successfully learned the relationship between input signals and teaching signals based on the robot's behaviour generated by expert's operation. In the experiments for reinforcement learning, Gauss-sigmoid neural networks employed as an "Actor" and a "Critic" were preliminarily trained in computer simulations for the sake of time saving, and then these neural networks were installed in the actual robot. As a result, it was confirmed that both of supervised learning and reinforcement learning were effective to acquiring a balancing control rule for our 1-legged hopping robot.
一般社団法人日本機械学会の論文