DEEP REINFORCEMENT LEARNING FOR MULTI-AGENT AUTONOMOUS SATELLITE INSPECTION

The commodification of space is resulting in larger numbers of satellites requiring increasingly complex operation and logistics and servicing support. The current standard of planning and executing maneuvers days in advance by teams of operators is not tenable, and will require endowing satellites with greater autonomy. We consider one such use case, specifically the problem of autonomous multiagent collaborative satellite rendezvous and proximity operations, where multiple satellites collaboratively inspect a target platform for on-orbit servicing and manufacturing missions. In this scenario, the satellites must, in a distributed and coordinated way, determine how to maneuver in order to perform a full inspection of the platform. We present a hierarchical, deep reinforcement learning solution to this problem. We show that reinforcement learning provides an alternative to traditional optimal control methods for generating high-performance policies requiring minimal computational overhead once deployed.