Safe reinforcement learning based agile satellite attitude control for non-cooperative target tracking with operation constraints and actuator faults

Persistent tracking of space-based non-cooperative targets is a fundamental capability for achieving space situational awareness and enabling numerous downstream space applications. However, it presents significant challenges, particularly during agile attitude maneuvers, due to the need to respect operational constraints and handle potential actuator faults. In this work, we divide the agile satellite tracking task into three distinct stages and formulate it as a constrained spacecraft attitude control problem. To systematically address the complex practical challenges in agile satellite control–including attitude constraints from forbidden zones, angular velocity limitations, torque saturation, and potential actuator faults–we formulate the problem as a Constrained Markov Decision Process (CMDP). To enhance the adaptability and fault-tolerance of the controller, we integrate a data-driven fault identification module capable of estimating both actuator effectiveness loss and additive torque faults. A safe reinforcement learning framework is further employed to learn fault-tolerant control policies that ensure continuous and stable tracking performance while minimizing the risk of constraint violations. Extensive simulation results demonstrate that the proposed approach enables the agile imaging satellite to achieve accurate, persistent tracking under strict operational constraints, while maintaining safe and resilient attitude control in the presence of actuator anomalies. © 2025 Elsevier Masson SAS.