safety security
The Steering Failure Lived Between the Test Cases
Formal bounds covered disturbance combinations that a driving campaign could not enumerate and exposed failures without another simulation run.
Summary
Formal bounds covered disturbance combinations that a driving campaign could not enumerate and exposed failures without another simulation run.
Four small end-to-end steering networks were trained in CARLA on either clear conditions or a mix of clear, fog, night and low sun. Instead of driving every disturbance combination, bound propagation read the trained weights and calculated steering drift continuously between captured images. On one arterial route, 133 poses with ten intensities each would imply 10^133 combinations; the formal calculation finished in minutes on one GPU and found conditions that broke the clear-trained policy. The experiment is a simulated proof of concept, but it shows verification can examine spaces between discrete road tests.
Why it matters
Formal bounds covered disturbance combinations that a driving campaign could not enumerate and exposed failures without another simulation run.
Limits and context
No additional limitation was separately recorded.
Key claims
Formal bounds covered disturbance combinations that a driving campaign could not enumerate and exposed failures without another simulation run.
Evidence: source-2026-09-11-011
Sources
- arXiv preprint 2609.10951arXiv · primary research
Corrections
No corrections have been recorded for this story.