New AI control method guarantees safety even with imperfect models

Conformal prediction (CP) has been used to provide probabilistic bounds on errors between learned dynamics models and real systems, which can then be embedded into control Lyapunov (CLF) and barrier function (CBF) frameworks. However, a critical flaw arises: when the policy is deployed, the closed-loop trajectory distribution differs from the training data distribution, breaking the CP guarantees. This distribution shift invalidates the safety and stability assurances.

To address this, Omid Mirzaeedodangeh and colleagues propose an episodic framework that iteratively updates robust conformal CLF/CBF policies while preserving guarantees across episodes. They achieve this by (1) employing adversarially robust conformal prediction to handle worst-case deviations, and (2) quantifying a distribution shift budget derived from closed-loop trajectory sensitivity analysis. This budget controls how much model error can increase between policy updates, yielding both implicit and explicit update rules. The algorithm's convergence is proven and demonstrated on three case studies. These are the first results to provide stability/safety guarantees for robust conformal CBF/CLF policies.