关键词: 深度强化学习, 深度确定性策略梯度, 机械臂, 柔顺捕获

Abstract: A compliant capture control method based on the principle of deep reinforcement learning is proposed to solve the in orbit capture problem by a space manipulator. In the simulation environment the controller is trained to use a 6 DOF manipulator to capture the target, which has specified mass and initial velocity. The controller learns to output appropriate control forces according to the states of the manipulator and it can make the target speed eventually approach to 0 and effectively reduce the impact force. The controller also shows good compliance interaction performance for targets with different mass and initial velocity. Compared with the traditional compliant control method based on the principle of impedance control, this method can reduce the max impact force effectively and realize the model independent control. After improving for space application, it is expected to be used in an intelligence capture task.

Key words: deep reinforcement learning, deep deterministic policy gradient (DDPG), manipulator, compliant capture