Skip to content

RefCloth: A Transferable Adversarial Tablecloth for Vision-Language-Action Models

RefCloth: Transferable Adversarial Tablecloths for Vision-Language-Action Models

arXiv: 其他 AI 安全论文· arXiv:2611.22736· Shaozhen Liu, Zhiwei Li, Jiayang Sun, Shuangjun Du, Yi Zhong, Shenrui Ma, Yaozu Liu, Qi Li, Zhenan Sun·Indexed by this site · 2 days ago· Originally publishedDate not givenAttacks· Its place in the scan record →
AI summary

The researchers propose the RefCloth framework, which optimizes a persistent adversarial tablecloth on a frozen source policy to attack vision-language-action (VLA) models, deployable zero-shot to unseen tasks and target models without modifying task objects or the camera. The method combines bounded texture parameterization, a shared surface UV transform, and a same-path clean reference to bridge the gap between differentiable and native renderers, keeping the tablecloth's attack effective under occlusion changes across static third-person and moving wrist dual views. In the native simulation of four LIBERO task suites, the same frozen texture transferred from OpenVLA-OFT to π0 and π0.5, reducing success rates by an average of 62.50 and 43.75 percentage points respectively compared with an unoptimized tablecloth, far exceeding amplitude- and variation-matched random controls. Dual-view auditing and closed-loop camera interventions show that the drop caused by the moving wrist view exceeds twice that of the main view alone and that the synergistic effect is strong; ablation experiments confirm the roles of UV diversification and path calibration.

Read originalarxiv.org