Red-Teaming of Vision-Language-Action Models via Quality Diversity Environment Generation

Key points

• A long-standing goal in robotics is to develop generalist robots that can perform tasks across a wide range of settings, flexibly specified with images and text.1
• Vision-Language-Action (VLA) models offer a promising path toward this goal by training on large collections of robot demonstrations.1
• Q-DEG uses Quality Diversity (QD) optimization to find a set of high-failure environments, while promoting diversity using policy embeddings extracted from the VLA.1
• In simulation, we conduct experiments across three representative VLA models, and our results indicate that Q-DEG improves robustness to out-of-distribution (OOD) adversarial environments compared to the base VLA and VLAs fine-tuned using data collected from two red-teaming baselines.1

Sources (1)

• [1]Red-Teaming of Vision-Language-Action Models via Quality Diversity Environment Generation
  arXiv (AI, ML, NLP, CV, robotics, multi-agent) · Oct 10, 12:25 AM

  To address this, we propose Quality Diversity for Diverse Environment Generation (Q-DEG), a method for improving VLA robustness by generating adversarial environments and fine-tuning the VLA on its own successful rollouts.

  A long-standing goal in robotics is to develop generalist robots that can perform tasks across a wide range of settings, flexibly specified with images and text.

Extractive summary: sentences quoted from the sources.