Abstract: In the RoboCup3D competition, a flexible and stable gait pattern is one of the key for the humanoid robot to win the match, to achieve such walk gait, a machine learning method of optimizing the vertical Center of Mass(Com) trajectory is presented. Firstly, to generate bipedal walk slowly and unsteadily, the vertical Com trajectory is planed by multiple polynomial, the inverted pendulum model(IPM) and a numerical method are utilized to control the Zero Moment Point(ZMP). Secondly, to get a fast and stable bipedal walk, a overlapping layered learning method is proposed to optimize the walking parameters which is based on Covariance Matrix Adaptation Evolution Strategy(CMA-ES). Finally, flocking control is applied to verify the flexibilty of the optimized gait in multi-robot environment. The experimental and competition results show that the proposed method is effective.

Key words: humaniod robot, gait optimization, IPM, evolutionary algorithm, RoboCup