Learning to Control a Free-floating Space Robot using Deep Reinforcement Learning

摘要

With the complexity of the dynamic model of free-floating space robots (FFSR), it is difficult to design the control system to capture targets. This paper presents a controller for FFSR to capture targets without the kinematic and dynamic model equations, where the agent learns a closed-loop control policy from state information only. At first, the process of the task is described as the reinforcement learning process without the dynamic models of the space robot. Then, we use the deep deterministic policy algorithm (DDPG) to train the policy for space manipulator motion planning. And we present a skill named "pre-training" in the training process to further import the learning efficiency. Finally, a 3 degrees of freedom space robot is modeled and simulated to demonstrate the validity of the controller.