Research on Robot Control Based on Reinforcement Learning

This paper reviews the rise and the development of the deep reinforcement learning (DRL). Then, the deep reinforcement learning algorithms for the high-dimensional continuous action space are divided into three categories of the algorithm based on the value function approximation, the algorithm based on the strategy approximation and the algorithm based on other structures. The latest representative algorithms and their characteristics of the deep reinforcement learning are explained in details, and their ideas, advantages and disadvantages are emphasized. Finally, combined with the development direction of the deep reinforcement learning algorithm, the control mechanism of using the deep reinforcement learning method to solve the control mechanism in the robotics problems is prospected.