REINFORCEMENT LEARNING-BASED FAULT-TOLERANT CONTROL OF ROBOT ARMS

Manlu Liu, Xinmao Li, Qiang Ling, and Jian Zhou

Keywords

Fault-tolerant control, robot arm single joint failure, missing failure information, reinforcement learning

Abstract

The failure of a single joint in a robot arm, especially with unavailable failure information, usually increases the difficulty of control. To solve this problem, this paper proposes a fault-tolerant control (FTC) based on reinforcement learning for robot arms. First, the workspace of a robot arm with a single failed joint is analysed. Based on the analysis, a quaternion sequence of the target-reaching task is formulated in the framework of the Markov decision process. A deep deterministic policy gradient algorithm is adopted to capture the optimal control strategy for the failed robot arm during trials of the target-reaching task. Second, two schemes are designed in the training process: training with respect to the failure of individual joints one by one and training with respect to the failure of any single joint. Finally, the effectiveness of the proposed FTC is verified through simulations and physical experiments on a target-reaching task with the proposed two training schemes. The results show that the FTC can effectively handle the failure of any single joint of the robot arm when failure information is unavailable. Moreover, the training scheme based on the failure of individual joints one by one performs better with missing failure information.

Important Links:

Go Back