Control Barrier--Value Functions under Partial Observability: Safety Guarantees via Conformal Prediction

What the math gives to ML

The paper offers a transferable safety mechanism for partially observed neural controllers: calibrate an estimator's state-error region with conformal prediction, then enforce a barrier or reachability certificate over the entire uncertainty region rather than only at the point estimate. The most practical neural-network construction is a real-time safety shield around an RNN, transformer, or reinforcement-learning policy, implemented as a minimally intervening quadratic program. Its strongest falsifiable signature is finite-sample coverage: under exchangeability, the true latent state should fall inside the calibrated error region with probability near 1-alpha, while safety violations should be controlled at approximately the same level when the certificate and dynamics model are valid.