Reusable vision-based monitor certifies self-driving safety without retraining

In a new paper on arXiv, researchers propose a certified runtime monitoring method for autonomous systems that uses semantic latent representations to evaluate safety specifications from visual observations. The monitor is reusable: once trained and calibrated, it can certify any formula in a target fragment of past-time signal temporal logic (ptSTL) without needing to retrain for each new specification. The key innovation is the 'semantic basis'—a vector of robustness scores for temporal atoms that serves as a minimum prediction target within a class of monotone, 1-Lipschitz interfaces. A single conformal calibration pass then certifies the entire fragment, avoiding the union bound that would otherwise degrade guarantees.

On a pedestrian-crossroad benchmark, the semantic-basis monitor performs up to 4× tighter at long horizons, while a simpler 'rolling prediction monitor' does better at short horizons. Both approaches were validated on real-world Waymo driving data, achieving the theoretical conformal coverage guarantees in practice. This work bridges vision-based perception, formal methods, and uncertainty quantification—offering a path to safety-certified autonomy without the overhead of retraining for every new traffic rule or scenario.