Red-Teaming Vision-Language-Action Models via Quality Diversity Prompt Generation for Robust Robot Policies