SNN Model Reinterprets Safety Thresholds to Match Human Braking

Current autonomous driving systems rely on Surrogate Safety Measures (SSMs) with fixed thresholds to evaluate collision risk—but these often fail to capture how humans react to sustained borderline conditions or brief high-risk spikes. A new paper from researchers at (affiliations not explicitly stated in snippet) reimagines safety thresholds by modeling them as the spiking thresholds of leaky integrate-and-fire (LIF) neurons. The approach combines multiple SSM inputs into a spiking neural network (SNN), which is trained to emit spikes precisely aligned with human braking onsets. Training data came from a controlled car-following experiment using the 3D-CoAutoSim platform (built on CARLA/Unreal) with a 6-DOF motion platform to generate critical events.

The results demonstrate that the SNN's learned spiking activity qualitatively aligns with braking behavior across various scenarios, capturing reactions that threshold-based methods alone cannot explain. Notably, analysis across participants reveals that learned input thresholds remain relatively consistent, while the decay factors encode different temporal sensitivities for each SSM. This indicates that spiking dynamics can serve as a mechanism to reconcile objective safety metrics with subjective human perception of risk. By moving beyond rigid thresholds to a dynamic, brain-inspired model, the work opens a path toward autonomous driving systems that feel more natural and safer to human passengers.