2P2-G14 階層型ニューラルネットワークに基づく3自由度1脚跳躍ロボットの学習制御に関する研究(進化・学習とロボティクス)

概要

論文の詳細を見る
In this study, we apply neural network-based balancing control to a 1-legged hopping robot with a spring element and seek continuous robot hopping under adequate balancing. We first built a remote control system for our robot that was composed of a hopping mechanism with 1-DOF and a joint mechanism with 2-DOF and created teacher signals for the neural network using a joystick-like interface. Then we embedded the obtained controllers into the robot and evaluated the performance for balancing control. As a result, we finally achieved six continuous hops; at the same time, we clarified the difficulty novices experience when they attempt to remotely keep our hopping robot in a balancing state. For supervised learning, the robot's behavior basically depends on the quality of the teacher signal. Our experiments also suggest that the enhancement of the method to generate teacher signals is crucial for more sophisticated neural network-based balancing control.
一般社団法人日本機械学会の論文
2008-06-06