Safety-Aware Residual Reinforcement Learning for Spacecraft Formation with Coupled Relative Translation and Attitude Dynamics

This paper studies leader–follower spacecraft formation control with coupled relative translation and attitude dynamics under actuator-effectiveness uncertainty, bounded disturbances, and safety constraints. A safety-aware residual reinforcement learning framework is developed by combining a nominal controller, a structured policy, and an online safety filter. Relative-degree decoupling is used to construct affine safe command inequalities, and a recursive safety filter maintains feasibility through backup-sequence update. Simulation results show improved tracking and training convergence over pure policy learning and single-step shielding, while maintaining near-zero safety violation and moderate online computation cost.