Event

Driving VLA can perceive danger but fails to brake, study proposes training-free fix

1 reports · 1 sources

Matteo Cercola et al. found that reasoning-based vision-language-action (VLA) driving models exhibit a "perception-action binding gap": they can perceive and reason about safety-critical hazards but cannot translate that evidence into safety actions such as braking. Using matched simulation counterfactuals and log replay, the authors report that in-path hazards are indeed encoded in the visual representation, but longitudinal planning is dominated by the ego vehicle's speed history and responds weakly to hazard distance and the model's own reasoning traces. To fix this gap, the researchers inject hazard-conditioned energy into the frozen Alpamayo-1.5 flow-matching sampler without retraining or changing weights. The authors report that, given precise obstacle geometry, this guidance reduces the footprint intrusion rate from 58.1% to 14.5%, restores safe braking when stopping distance is sufficient, and completely eliminates collisions in their braking-related closed-loop tests.