Aman Arora

Reinforcement Learning applied to Proximity Operations of Small Spacecraft

Autonomous docking and on-orbit refueling are key enablers of a self-sustaining space mobility infrastructure, where spacecraft can be refilled in orbit instead of being replaced. The task is challenging since terminal contact demands centimetre-level precision, propellant transfer shifts the centre of mass and inertia of the stack, and servicers must handle clients with uncertain dynamics. The thesis develops reinforcement learning (RL) algorithms and RL-based guidance and control for close-proximity operations, docking and propellant transfer of small spacecraft. Safety requirements derived from ESA close-proximity guidelines, such as approach corridors and range-dependent velocity envelopes, are embedded directly in training through constrained RL, producing compact policies that need no online optimisation at deployment. The work spans massively parallel GPU-accelerated simulation, benchmarking of classical, safe model-free and model-based reinforcement learning controllers, and is validated through zero-shot sim-to-real transfer on the floating platforms of the SpaceR Zero-G Lab, with a final goal of a complete hardware-in-the-loop in-orbit servicing demonstration in the laboratory.