Acquisition of walking behaviors for robots using Central Pattern Generator and machine learning

Abstract

Recently, biologically inspired approaches have received much attention for controlling behaviors of robots. A typical example of biologically inspired approaches is control of rhythmic behaviors using Central Pattern Generator (CPG). This method is useful to control adaptive walking behaviors of robots, however, this method has a problem that there are few theories to design CPG. Hence, in order to design CPG, we proposed a learning method called the Three Steps Method which is a combination method of Genetic Algorithms and Reinforcement Learning. In this study, in order to show the effectiveness of the Three Steps Method, we apply it to a quadruped robot with the CPG controller. The simulation results show that the Three Steps Method leads robots to acquire walking behaviors without the knowledge of designers. Furthermore, we also show that the robot can adapt to various environments using the proposed method.