Statistical Verification of Autonomous Systems using Surrogate Models and Conformal Inference

We propose a conformal-inference–based framework for statistical verification of cyber-physical systems (CPS) operating under parametric uncertainty. The method trains surrogate models on simulation data to predict satisfaction of Signal Temporal Logic (STL) specifications and uses conformal prediction to provide finite-sample probabilistic guarantees. A Gaussian-process–driven refinement procedure further yields localized assurance regions in the parameter space. Experiments on autonomous-vehicle and aerospace benchmarks demonstrate substantial reductions in simulation cost while preserving rigorous statistical soundness.

arXiv preprint arXiv:2004.00279