One Adversarial Texture Drops VLA Robot Success From 90.0% to 48.4% Across Multiple Tasks

arXiv 2608.13453·medium signalXBlueskyLinkedInCopy link

UniTexture backpropagates gradients from a Vision-Language-Action policy's action outputs to the surface texture parameters of a single 3D object through a differentiable renderer, jointly optimizing one shared texture over a distribution of tasks, instructions, states, and viewpoints. Evaluated on OpenVLA and pi-0.5, it cut mean task success from 90.0% under benign conditions to 48.4% under attack, induced target-aligned action shifts, and transferred across task suites and models without re-optimization. Unlike prior single-task attacks, this exposes a shared cross-task vulnerability — one physical object in the scene, no digital access required, and the generalist policy is the thing that makes it generalize.