Study on Virtual Training Platform for Reinforcement Learning of Space Manipulator

In this paper, a training platform based on Unity engine of target-capturing control strategies for a five-DOF space manipulator with fixed base is built. The platform mainly includes model processing, communication module and algorithm module, which provides the basic functions needed for space missions. Then PPO algorithm and imitation learning method are used in this platform for virtual training of space manipulator. And the results show that compared with PPO algorithm, the hybrid method of PPO algorithm and imitation learning has higher training efficiency and faster convergence speed. Finally, the trained model is successfully applied to the control of a real manipulator. Research of this paper shows that virtual training platform has universality and application prospect to a certain extent, and has greater practical value for future space tasks with high complexity and intelligent requirements.