Reinforcement Learning for Rendezvous and Proximity Operations

As the advancement of smaller spacecraft and satellite constellations continues to grow, faster and cheaper control systems need to match the production and complex operations of these systems. In particular, the need to dynamically model and optimize rendezvous and proximity operations (RPO) for satellite constellations are critical for modern space missions. Reinforcement learning (RL) offers an autonomous decision-making process on-board these systems that can utilized to find optimized controls to solve various mission designs. This paper presents an RL model that generates control magnitudes of a deputy spacecraft attempting to dock to a chief satellite with varying degrees of complexity including keep-out zones, randomized initial positions, and a manual override to simulate a collision avoidance maneuver. By developing a model to solve a simpler controls problem, the feasibility of using autonomous control systems for RPO is shown with the potential of further optimized solutions to be calculated.