How to Guarantee Safety for Neural Network based Rocket Engine Controllers

Kurzfassung

This paper introduces an intelligent reinforcement learning controller for the LUMEN rocket engine, trained entirely in simulation to meet complex control requirements such as reusability and throttling. To ensure safe real-world deployment, the focus is on detecting when the agent is operating in situations it hasn’t seen during training, which is important to ensure safe use in real-world applications. By introducing unexpected disturbances into the simulation and comparing the agent’s actions with real LUMEN engine test data, the method identifies deviations that signal when the controller’s performance may be unreliable. This approach represents a crucial first step toward validating the practical applicability and safety of reinforcement learning agents for rocket engine control