Stable RL Control for Satellites with Learned Dynamics

Deep Reinforcement Learned (RL) controllers are finding expanded applications in numerous fields due to their capabilities to optimize in complex, high-dimensional settings. However, most deep RL techniques lack guarantees for stability or other controller behavior, limiting their application in high-stakes contexts such as satellite control. In this proposed effort, we will explore using deep RL techniques to model unknown system dynamics and provide provably stable control. In particular, we will investigate learning Lyapunov functions to guarantee controller stability. Efforts in Phase I will include conducting a review of relevant literature, sketching a proof for a guaranteed stable RL controller, and verifying our RL methods with simulated tests. In phase II, a formal proof for guaranteed stable RL control will be completed. In addition, our method will be verified extensively with more complex satellite simulation scenarios, we will add methods to learn and account for uncertain and unknown dynamics, and we will investigate additional proofs for guaranteed controller passivity and bounded disturbance rejection.