Neural Safety Filters: Scaling Safety-Critical Control Without Surrendering Guarantees

The expanding reach of robotics and AI brings acute reliability challenges under limited information, rapidly changing conditions and close human interaction. Safety filters, runtime schemes that monitor a system's candidate actions and intervene to preempt catastrophic failures, have attracted wide interest, but the restrictive assumptions of model-based approaches and the weaker guarantees of data-driven alternatives have reinforced the perception of a fundamental safety-performance tradeoff. Jaime Fernández Fisac presents theoretical and experimental evidence against that view. Building on a unified understanding of safety filter mechanics, he introduces a general class of neural safety filters that extend the rigorous guarantees of safety-critical control beyond the reach of classical synthesis. Safety over a specified operational design domain (ODD) is cast as an adversarial reinforcement learning problem, co-training the system's controller against the worst environment realisations consistent with the ODD. These neural approximations enable certified safety filters in domains from legged locomotion to high-speed racing, including the synthesis of control barrier functions for general nonlinear systems with only black-box access to their dynamics. Wrapped around safety-agnostic planning and learning pipelines, the filters achieve near-optimal task performance in safety-critical ODDs while provably enforcing zero catastrophic failures. The talk then turns to human-centred systems, where safety requires steering interactions away from downstream harms, from anticipating black-swan road encounters to safeguarding novice human operators, and closes with early results on dynamic guardrails for generative AI assistants and open challenges for safety-critical intelligence beyond robotics.