Neural hybrid surrogates speed up safety falsification for hybrid systems
Learning a differentiable automaton model from data enables gradient-based falsification with fewer simulations.
Falsification of hybrid dynamical systems—which combine continuous dynamics with discrete mode transitions—remains a hard problem in safety verification. Existing surrogate-based methods rely on purely continuous dynamics and cannot handle mode-dependent behavior. To bridge this gap, Kötz and Åkesson introduce a neural hybrid surrogate that learns a differentiable hybrid automaton model directly from trajectory data. The model uses a neural hybrid automaton to encode latent modes and their associated vector fields. Transition guards are inferred from the data, yielding a fully differentiable surrogate that can be optimized with gradient-based optimal control.
Once trained, the surrogate is fed into a gradient-based optimal control loop that minimizes a smooth approximation of the safety specification—e.g., avoiding unsafe states. The resulting control solution is then tested on the original system to ensure soundness. Experiments on standard benchmarks show that the method consistently uncovers counterexamples on most specifications, achieving competitive or improved sample efficiency over tools like S-TaLiRo and Breach while using a reduced simulation budget. This work opens the door to more scalable falsification for complex cyber-physical systems.
- Uses neural hybrid automata to learn a differentiable surrogate model from trajectory data, capturing both continuous dynamics and discrete mode transitions.
- Gradient-based optimal control formulation directly minimizes safety specifications to identify counterexamples efficiently.
- Achieves competitive sample efficiency and reduced simulation budget compared to existing tools on benchmark falsification problems.
Why It Matters
Faster, more scalable safety verification for autonomous vehicles, robotics, and other cyber-physical systems.