Satellite Attitude Control with Deep Reinforcement Learning

Abstract

Deep reinforcement learning has recently gained great interest in the field of intelligent control. In this paper, we develop a set of deep reinforcement learning algorithms on satellite attitude control. By improving Deep Deterministic Policy Gradient (DDPG) and Twin Delayed DDPG (TD3), we analyze three kinds of attitude control problems, including no state-constrained, state-constrained and online optimization problems. We compare Deep Reinforcement Learning Controller with traditional PD controller at the same time. Besides, we summarize a design process to apply deep reinforcement learning algorithms on satellite attitude control problems. It is shown in this paper that Deep Reinforcement Learning Controller has advantages on model-free control and online optimization.