RL-STPA provides a practical toolkit for systematically finding safety hazards in RL systems before deployment, even when formal verification is impossible—by combining domain expertise, targeted testing, and iterative safety improvements through training.
This paper adapts System-Theoretic Process Analysis (STPA), a safety engineering method, to evaluate reinforcement learning systems in safety-critical applications like autonomous drones.