Safe Reinforcement Learning on Attitude Control of Launch Vehicles

This paper delineates a safe reinforcement learning framework for the attitude control task of launch vehicles. The framework attains an equilibrium between exploratory maneuvers, error trails, and safety-critical control during the flight by integrating reinforcement learning algorithms with a suite of backup control law. The reinforcement learning algorithm facilitates the progressive learning and evolutionary refinement of the launch vehicle's control strategies. This method holds promise as a prospective control methodology for the ongoing learning of launch vehicles within the context of future reusable flight missions and the transition towards routine space transportation. It is envisioned that this method will lead to enhanced adaptability of launch vehicles to varying flight environments and to a continual increase in control precision.