One adversarial texture drops robot policy success from 90.0% to 48.4%.

UniTexture backpropagates gradients from a Vision-Language-Action policy's action outputs to the surface texture of a single 3D object through a differentiable renderer, optimizing one shared texture over a distribution of tasks, instructions, states and viewpoints. Tested on OpenVLA and pi-0.5, it transferred across task suites and models without re-optimization. arXiv 2608.13453 One physical object in the scene, no digital access required. The generalist policy is what makes the attack generalize.