Lost in Fog: Sensor Perturbations Expose Reasoning Fragility in Driving VLAs
A new research paper titled "Lost in Fog" investigates the reasoning fragility of Vision-Language-Action (VLA) models in autonomous driving. The study subjected the Alpamayo R1 model to various sensor perturbations, including noise, extreme lighting, and fog, across nearly 2,000 scenarios. Researchers found that changes in the model's Chain-of-Causation (CoC) explanations directly correlated with significant increases in trajectory deviation, highlighting reasoning consistency as a critical safety indicator for VLA deployment. AI
IMPACT Reveals critical safety vulnerabilities in autonomous driving AI, motivating new runtime monitoring techniques for VLA systems.