Acquisition of adaptive walking behaviors using machine learning with Central Pattern Generator

Recently, biologically inspired approaches have received much attention for robot control. A typical example of them is control of rhythmic behaviors by Central Pattern Generator (CPG). However, this control has a problem that there are few theories to determine parameters of CPG. For this reason, they are determined experimentally. In this paper, we propose a combination method of Genetic Algorithm and Reinforcement Learning for determining parameters of CPG, and apply to a quadruped robot with the CPG controller. Simulation results show that the robot obtains walking behaviors automatically through learning process without using the parameters set by knowledge of designers.