Omnidirectional walk design of humanoid robots using layered learning method based on CMA-ES

Abstract

In the RoboCup3D simulation environment, the omnidirectional walk of biped robots exerts a crucial impart on the result of the competition. For a Nao humanoid robot possessing 22 degrees of freedom, 12 main walking parameters are involved in its walking model, due to which it is overwhelmingly difficult to reach optimal effectiveness in a hand-coded way. This paper depicts a method to develop an omnidirectional walk based on the Covariance Matrix Adaptation Evolution Strategy (CMA-ES), which is used by Apollo3D humanoid robots. In order to achieve walking motion, a double linear inverted pendulum (D-LIP) model based on the predictive control is built, and then a layered learning based on CMA-ES is designed to optimize the walking parameters. Detailed experiments reveal that this paper proposed an effective method for the realization of the rapid and stable omnidirectional walk of biped robots in a complex and dynamic environment.