Spacecraft Autonomous Rendezvous Control With Rotation Target on SE(3): Deep Reinforcement Learning Methods

This paper concentrates on the spacecraft autonomous rendezvous (SAR) control scheme within the deep reinforcement learning (DRL) framework, effectively addressing challenges such as external disturbances, model uncertainties, and fuel consumption. A nonlinear integrated orientation and orbit dynamic model utilizing Lie group SE(3) is developed with gravitational perturbations and gravity gradient torque for each spacecraft. In response to uncertainties, disturbances, and the optimization objective, a training environment, SE3SAREnv, and a corresponding reward function are meticulously designed to train DRL controllers. The efficacy and advantages of implementing the Actor-Critic methods for the SAR control problem are demonstrated through two comparative experiments on the proposed SE3SAREnv: the applications of a traditional control algorithm (sliding mode control, for example) versus three prevalent DRL algorithms.