Reinforcement Learning for Autonomous Visual Inspection of Space Targets
- Paper ID
100422
- DOI
- author
- company
University of Luxembourg;
- country
Luxembourg
- year
2025
- abstract
The increasing complexity of space missions necessitates advanced autonomous systems capable of performing intricate tasks with minimal human intervention. Traditional control strategies often require extensive modeling and tuning, making them less adaptable to changing mission conditions and unanticipated disturbances. This study presents a novel approach employing Reinforcement Learning (RL) to control satellite proximity maneuvers, unlocking the potential of autonomous inspection of a 3D target in space. The Lunar Gateway is used as a demonstrative case study, showcasing the feasibility of RL-driven spacecraft navigation for on-orbit servicing and observation. Leveraging a model-free, physics-agnostic RL framework, our method provides adaptability across various satellite configurations and mission parameters. We utilize the Proximal Policy Optimization (PPO) algorithm to train the RL agent in a state-based environment, enabling precise maneuvering towards both predefined and dynamically generated waypoints. The satellite's control system operates with discrete thrusters, implementing bang-bang control to facilitate efficient 3D movement. Operational constraints, including fuel limitations and collision avoidance, are integrated into the training process through reward penalties, ensuring safe and resource-aware decision-making. Training is conducted using RANS, our framework built on top of NVIDIA's Isaac Lab, which enables massive parallelization for efficient policy learning and photorealistic rendering for high-fidelity evaluation. To ensure robust performance validation, we employ extensive simulated evaluations across varying environmental conditions, analyzing waypoint tracking accuracy, fuel consumption, and maneuver completion time under diverse initial conditions and disturbances. A statistically rigorous evaluation methodology ensures the reliability of the trained policy across different operational contexts. The results demonstrate the potential of reinforcement learning to enhance autonomy in spacecraft operations, paving the way for future applications in in-orbit servicing, spacecraft inspection, and cooperative assembly tasks. By providing a scalable and retrainable control framework, this work contributes to the broader vision of self-sufficient space infrastructure, where intelligent agents can autonomously perform critical tasks in future space stations, deep-space exploration missions, and planetary operations.