Source note

Uncovering Linguistic Fragility in Vision-Language-Action Models via Diversity-Aware Red Teaming

Vision Language ActionRobot Red TeamingLinguistic RobustnessReinforcement LearningRobot Safety

This paper studies how fragile vision-language-action models are to harmless-looking instruction rewrites. It proposes DAERT, an RL-based red teaming method that searches for diverse paraphrases that keep the task meaning but make robot policies fail.

  • VLA robot policies can fail when a human gives the same task in different words, which is a safety risk for deployment.
  • Existing automated red teaming methods often find a small set of repetitive attacks because reward-maximizing RL collapses to one mode.
  • A useful attack generator must keep the original task meaning, stay executable, and still expose many different failure cases.
  • The paper trains a vision-language model attacker to rewrite a task instruction into a semantically equivalent but harder instruction, using simulator feedback from the target robot policy.
  • DAERT adds a diversity-aware RL objective inspired by ROVER. In simple terms, it rewards attacks that work while discouraging the policy from putting all probability on one rewrite pattern.
  • The method uses an implicit token-level actor-critic with a uniform-average successor value so generation stays spread across multiple valid continuations instead of collapsing to one phrase template.
  • A cascaded reward filters invalid attacks before simulator evaluation: format checks, semantic similarity to the original instruction, and a length cap. Only valid rewrites get credit for causing task failure.
  • Experiments use Qwen3-VL-4B as the attacker and test against two target VLAs, π0\pi_{0} and OpenVLA, on LIBERO, with transfer evaluation discussed for CALVIN and SimplerEnv.
  • On LIBERO against π0\pi_{0}, original instructions give 93.33% average task success, ERT drops this to 65.50%, GRPO to 20.45%, and DAERT to 5.85%.
  • On the same π0\pi_{0} setting, DAERT also improves diversity scores over GRPO: CLIP cosine-distance metric 12.23 vs. 7.05, and LLM-judge score 8.48 vs. 4.58.
  • Per-task-suite results for π0\pi_{0} + DAERT on LIBERO are 7.4% success on Spatial, 8.8% on Object, 3.0% on Goal, and 4.2% on Long. Lower is stronger because the attack causes more failures.
  • On OpenVLA finetuned on LIBERO, average success falls from 76.50% under original instructions to 32.15% with ERT, 17.00% with GRPO, and 6.25% with DAERT.
  • The paper claims DAERT beats previous methods by +59.7% in attack success rate and transfers across target VLAs and robotic domains.
  • A diagnostic test shows why the authors focus on π0\pi_{0} for language fragility: with a generic "no action" instruction, π0\pi_{0} gets 17.65% average success while π0.5\pi_{0.5} still gets 54.90%, suggesting π0.5\pi_{0.5} depends less on language.