Camera-only end-to-end driving models are nearing deployment, but existing corruption-robustness benchmarks focus on detection or bird's-eye-view perception, not the planning output that actually drives the car. DriveDegrade, a new benchmark from Que and Yao, addresses this by injecting 16 corruption families at five severities on the fly inside the image loader, then evaluating open-loop planning on nuScenes and NAVSIM plus a CARLA closed-loop anchor.

The results show that mild degradation barely affects planning, and each corruption family has a clear threshold around mid severity where performance breaks. Blur, JPEG, and raindrops are the most damaging, while weather and bit errors are tolerated far into the range. The authors argue that a flat performance curve is ambiguous: a planner that genuinely reads its camera must lose accuracy when information is deleted, so they separate corruptions that degrade the image from those that remove it.

On these two axes, planners separate sharply, exposing what the authors call the ego-status shortcut. One released vision-language-action planner was flat on both axes, and blinding all six of its cameras cost it only 11.5 percent—indicating it was not truly relying on camera input. This distinction matters for deployment, where apparent robustness could mask a policy that ignores its sensors.