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

Researchers developed Q-DIG, a red-teaming method that uses Quality Diversity techniques to identify diverse language instruction failures in Vision-Language-Action models for robotics. The approach generates adversarial prompts that expose vulnerabilities in robot behavior and improves task success rates when used for fine-tuning.

• →Q-DIG combines Quality Diversity techniques with Vision-Language Models to generate adversarial instructions that expose VLA robot vulnerabilities.
• →The method finds more diverse and meaningful failure modes compared to baseline approaches across multiple simulation benchmarks.
• →Fine-tuning VLA models on Q-DIG generated instructions improves task success rates on both seen and unseen instructions.
• →User studies show Q-DIG generates more natural and human-like prompts than baseline methods.
• →Real-world evaluations confirm simulation results, validating the approach's practical effectiveness for improving robot robustness.