Parallel stochastic optimization for humanoid locomotion based on neural rhythm generator

Abstract

This paper proposes a parallel stochastic approach to neural oscillator based motion control for bipedal humanoid locomotion. The motion control is based on the Central Pattern Generator (CPG), and is optimized by Simulated Annealing. Optimization of parameters in the motion control based on the CPG is a very hard problem. The number of the parameters which should be optimized increases as a robot's link structure becomes complicated. To optimize all parameters simultaneously may cause explosion of search space. Therefore, we divide search space into upper-limbs optimization space and leg optimization space. We then propose a parallel optimization method by two processes, which handles the control parameters of upper-limbs and legs, and communicates them each other. In the experiments, our method succeeded in optimization of all parameters without explosion of search space, and performed superior gaits. © Springer-Verlag Berlin Heidelberg 2005.