Statistically Assuring Safety of Control Systems using Ensembles of Safety Filters and Conformal Prediction

The introduction of a conformal prediction-based framework marks a significant advancement in the safety assurance of learning-enabled autonomous systems. Traditional methods like Hamilton-Jacobi (HJ) reachability analysis are crucial for verifying safety but are computationally intensive, particularly for high-dimensional systems. This new approach leverages reinforcement learning to approximate the HJ value function while addressing the inherent uncertainty in learned policies. By providing probabilistic safety guarantees, the conformal prediction framework ensures that control systems can avoid unsafe states, thus enhancing their reliability. The ensemble of independently trained HJ value functions acts as a safety filter, further solidifying the effectiveness of this method. This development is particularly relevant as the deployment of autonomous systems continues to grow, necessitating robust safety measures.