Robotic Application of Reinforcement Learning to Central Pattern Generator

Abstract

Reinforcement learning is one of the machine learning algorithms that learns by trial and error. Its general purpose is to design the controllers(policy) for nonlinear systems. For now, it is widely used in high-dimensional and continuous action spaces since the advent of deep reinforcement learning to use deep neural network as function approximators. The animal-like robots have been studied to answer the biological questions. In partic- ular, the salamander robot has been developed to implement the evolution of vertebrates from aquatic to terrestrial. A numerical central pattern generator model is proposed in or- der to control the locomotion of the robot. Since its action spaces are continuous and there are too many uncertainties in environment, it is difficult for the robot to apply traditional reinforcement learning algorithm. In order to solve this, deep deterministic policy gradient among deep reinforcement learning algorithms is applied. The robotic salamander and its environments are imple- mented in Gazebo dynamic simulation with Robot Operating System. In this thesis, the per- formance of deep deterministic policy gradient is verified by the robotic salamander which overcomes the uncertainties in environments.