A driving model can see a hazard yet fail to brake

When a driving model’s path reaches a hazard, what failed: did the model lack the information, could it not produce a safe response, or did the information fail to shape its action?

Those causes point to different explanations, and ordinary measures of average trajectory quality may not distinguish them.

The authors add a stopping constraint during trajectory generation to test whether direct hazard guidance can make a frozen driving model brake safely.

The authors tested whether Alpamayo-1.5’s weak response to stationary obstacles reflected missing visual information, inability to brake, or insufficient influence of hazard evidence on the plan.

Matched simulated scenes made obstacle presence the key contrast, while separate input ablations measured which channels moved the trajectory.

The visual representation distinguished obstacle scenes from empty roads, but close-range paths still intruded; removing ego history shifted longitudinal plans far more than blanking vision. When an exact stopping constraint guided generation, intrusion fell on seen sites, and guidance prevented contact in all 24 closed-loop scenes where braking was necessary.

Together, these results support a gap between hazard evidence and action, within the tested model and simulations.

The hazard scenes were synthetic, and footprint intrusion is a planning measure, not a real-world crash rate.

The closed-loop result covers 24 simulated scenes where braking was necessary and used an always-active, accurately specified stopping constraint; the paper does not show that a system can reliably decide when to activate it.