Reproduce “TrapVLA: Trapping Vision-Language-Action Models in Configured Failure Modes” on OpenResearch
Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.
Universal app for Apple silicon and Intel. Requires macOS 11 or later.
About this paper
Jun-Hui Liu, Kun-Yu Lin, Yi-Lin Wei, Xu-Han Chen, Yinghao Li, Zhuohao Li, Yuan-Ming Li, Qing Zhang and 4 more
A new backdoor attack, "Configured Failure Trapping," enables adversaries to inject precise, attacker-specified failure modes into Vision-Language-Action (VLA) models using stealthy textual triggers, while maintaining high performance on benign tasks. The TrapVLA method, developed by researchers at Sun Yat-sen University and collaborators, overcomes the sparse action deviation problem by explicitly modeling trigger-induced action residuals, achieving high attack fidelity across various robotic manipulation tasks.
Key insights
- Enabling fine-grained, precise control over robotic failure modes represents a more sophisticated and insidious threat than merely causing arbitrary task disruption.
- Configured failure modes, designed to be 'behaviorally plausible' and induced by stealthy textual triggers, are significantly more difficult to detect than unconstrained or random failures.
- Explicitly supervising trigger-induced action residuals through Target Residual Steering (TRS) is crucial for reliably associating localized malicious behaviors with global triggers in VLA backdoor attacks, enhancing both attack effectiveness and clean-task preservation.
Summary by alphaXiv.
Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.