SafeVLA-Bench: A Benchmark for the Success-Safety Gap in Vision-Language-Action Models
The episode discusses SafeVLA-Bench, a framework that measures the success-safety gap in Vision-Language-Action models using metrics like Succ-But-Unsafe (SBU) and Violation Severity Index (VSI). The hosts discuss how this formal safety checking moves beyond binary success to expose dangerous behaviors. Future work focuses on integrating these formal specifications into training for policy robustness.