Reinforcement learning for passive dynamic walking robot

A quasi-passive dynamic walking robot was built to study natural, energy-efficient biped walking. The robot was actuated by mechanically adjustable compliance and controllable equilibrium position actuators (MACCEPA). A reinforcement learning based method was used to control the robot to walk. The method firstly learned the desired gait for walking in ideal environment with a gait model based Q-learning algorithm. Then, a fuzzy advantage learning method was used to teach the robot to walk in uneven floor. Stable walking of the robot is achieved by using the learning result to control the action of the actuators when changes occur in the walking phase. The effectiveness of the method was verified by simulations.