UniTexture: Cross-Task Universal Adversarial Textures for Vision-Language-Action Models
UniTexture is a cross-task universal adversarial texture attack that uses a single textured 3D object to induce targeted deviations in VLA action predictions across multiple tasks. It backpropagates gradients from the policy's action outputs to surface texture parameters through a differentiable renderer, jointly optimizing the shared texture over a distribution of tasks, instructions, states, and viewpoints using a targeted action-space objective. The attack is evaluated on OpenVLA and π0.5 across diverse manipulation tasks and multiple evaluation settings.
Researchers introduced UniTexture, a cross-task universal adversarial texture attack targeting Vision-Language-Action (VLA) models. The method uses a single textured 3D object to cause targeted action deviations across multiple tasks, optimizing the texture via differentiable rendering and gradient backpropagation from action outputs. Evaluation on OpenVLA and π0.5 demonstrates the attack's effectiveness across diverse manipulation tasks and settings.
UniTexture leverages a differentiable renderer to backpropagate gradients from VLA action outputs to surface texture parameters, enabling joint optimization over task, instruction, state, and viewpoint distributions. This approach creates a universal adversarial texture that generalizes across tasks without per-task optimization, highlighting a cross-task vulnerability in multitask VLA policies.
The attack exposes a potential safety gap in deployed robotic systems using VLA models, as a single adversarial object could cause unsafe physical behaviors across multiple tasks. This may prompt robotics and AI safety teams to incorporate adversarial robustness testing into VLA development and deployment pipelines.
For robotics companies and AI safety vendors, UniTexture highlights a new class of security risks that could affect product liability and safety certification. It may create demand for adversarial robustness tools and testing services in the robotics and embodied AI market.
Future work may focus on developing defenses against universal adversarial textures, such as robust training or input sanitization, and extending the attack to other VLA architectures and real-world robotic platforms. The research may also influence safety benchmarks for embodied AI.