Two Approaches to Autonomous Spacecraft Inspection: Shielded Reinforcement Learning versus Probabilistic Roadmaps

Abstract The close-proximity inspection of objects in low Earth orbit (LEO) is important to operations such as rendezvous, debris removal, servicing, and resident space object (RSO) characterization, all of which are of increasing interest to commercial and government organizations. Complex relative motion dynamics in eccentric LEO make the problem of path planning for autonomous inspection challenging. Agents must be able to fully inspect an object subject to illumination constraints while avoiding collision with the RSO. In this work, two approaches to the problem are developed and compared: First, a deep reinforcement learning (RL)-based solution that allows for closed-loop autonomous inspection is formulated, using analysis of equations of motion to derive a shield that guarantees safe actions. Second, a solver using probabilistic roadmaps (PRMs) and mixedinteger quadratically-constrained quadratic programs (MIQCQPs) is designed to provide more performant but computationally expensive open-loop inspection trajectories. The overall performance of solutions produced by the two methods is compared across the fuel-time trade space in a high-fidelity simulation environment: These experiments demonstrate that while PRMs result in more fuel- and time-efficient solutions than RL, this comes at the cost of offline, open-loop solutions as opposed to the closed-loop, onboard control provided by RL-based policies.