When and How Severely: Scenario-Specific Safety Envelopes for Driving VLAs

The paper presents a safety certification framework for Vision-Language-Action (VLA) driving planners, specifically evaluating the Alpamayo R1 model, which has 10 billion parameters. It introduces a two-dimensional safety envelope approach, revealing that a single aggregate threshold can mask significant variations in failure severity across different scenarios. This finding underscores the necessity for a nuanced understanding of operational design domains (ODD) in safety assessments, which is crucial for practitioners developing reliable autonomous driving systems.