Spacecraft Command and Control with Safety Guarantees using Shielded Deep Reinforcement Learning

Increasingly complex space missions have motivated the development of autonomous command and control approaches which must handle high-dimensional, continuous observation and action spaces with hard-to-analyze behavior. Deep reinforcement learning (DRL) techniques are a rising area of research for dealing with such problems, but lack performance and safety guarantees which reduce their applicability for spacecraft operations. This work identifies promising strategies from the DRL literature for providing safety and performance guarantees using correct-by-construction shield synthesis, and techniques for identifying the robustness and stability of trained agents in a computationally expedient manner. Additionally, open-source spacecraft simulation tools derived from the AVS Basilisk astrodynamics simulation package are presented and discussed. Shielded learning agents are presented against a naive DRL agent approach for the command and control of a LEO ground observation mission and compared on the basis of performance, computational efficiency, and safety.