Reinforcement Learning for Docking Maneuvers with Prescribed Performance

Abstract: We propose a two-component data-driven controller to safely perform docking maneuvers for satellites. Reinforcement Learning is used to deduce an optimal control policy based on measurement data. To safeguard the learning phase, an additional feedback law is implemented in the control unit, which guarantees the evolution of the system within predefined performance bounds. We define safe and safety-critical areas to train the feedback controller based on actual measurements. To avoid chattering, a dwell-time activation scheme is implemented. We provide numerical evidence for the performance of the proposed controller for a satellite docking maneuver with collision avoidance.