Sample-Efficient and Robust Visual Reinforcement Learning for Autonomous Spacecraft Proximity Operations

On-Orbit Servicing, Assembly, and Manufacturing (OSAM) demands robust autonomous Rendezvous and Proximity Operations (RPO) with non-cooperative targets. Traditional vision-based Guidance, Navigation, and Control pipelines, reliant on modular pose estimation and model-based control, suffer from sensitivity to environmental variations and errorpropagation. Visual Reinforcement Learning offers an end-to-end solution but faces challenges in sample efficiency, generalization, and safety for orbital applications. We propose the Sample-Efficient and Robust Visual Learner, SRVL, an end-to-end Visual Reinforcement Learning framework. SRVL employs a ResNet-based visual encoder to capture environmental dynamics, enhancing visual representation and accelerating convergence. A group-wise action masking mechanism, synergistically optimized with consistency loss and pseudo-label supervision, enables smoother control and reduces fuel consumption. Evaluations in a MuJoCo-based DMControl simulation show SRVL outperforms baselines in success rate, efficiency, and robustness, advancing safe and effective autonomous space operations.