UniTexture: Cross-Task Universal Adversarial Textures for Vision-Language-Action Models
A single adversarial 3D texture cut VLA task success from 90.0% to 48.4% in the authors’ tests.
The paper introduces UniTexture, an attack that optimizes one textured object to steer robot policy actions across multiple tasks, instructions, states, and viewpoints. It targets Vision-Language-Action models by backpropagating from action outputs through a differentiable renderer into surface texture parameters. The authors evaluate it on OpenVLA and pi_0.5, reporting target-aligned action shifts plus cross-suite and cross-model transfer without re-optimization. Source: ArXiv · AI/CL/LG's note.
The paper introduces UniTexture, an attack that optimizes one textured object to steer robot policy actions across multiple tasks, instructions, states, and viewpoints. It targets Vision-Language-Action models by backpropagating from action outputs through a differentiable renderer into surface texture parameters. The authors evaluate it on OpenVLA and pi_0.5, reporting target-aligned action shifts plus cross-suite and cross-model transfer without re-optimization. Source: ArXiv · AI/CL/LG's note.
score 5