The Study on the Learning of Walking Gaits for Biped Robots

This paper presents a study on biped walking and balance control by reinforcement learning. The robot learns how to walk without prior knowledge of explicit dynamics model by a reinforcement learning approach. The Q-learning sharpens up the robot's walking gaits so as that improve the stability of each gait and reduce the number of the posture patterns in a gait cycle. In this paper, the learning agent employs an intuitive evaluation knowledge to help the biped robot with a basic walking skill learn to improve its behavior in terms of restricting ZMP to a certain region on each foot.