SafeVLA-Bench: A Benchmark for the Success-Safety Gap in Vision-Language-Action Models