Reinforcement Learning of Action Strategies and Joint Stiffness of Tendon-driven Biped Robot

点击翻译按钮获取中文摘要

In this research, a learning method for robots to learn appropriate actions by profits and penalties given from the environment was developed and applied to action learning in the robotic succor game and walking movement of a biped robot. To apply it to the real robots and to improve the efficiency, a method to decide the criterion for penalties was considered and states in that the robot already had learned sufficiently were treated as a fixed-mode state(deterministic action strategy is used). Furthermore, the mechanism and control of a biped robot driven with motors(muscles) and wires(tendons) were considered. The tensile force control of tendons was done with robust stiffness-adjustable device, since tensile force up to 400N(40kgf) is expected during walking.