UniTexture: Universal Adversarial Texture Attack on VLA Models
A new adversarial texture attack called UniTexture has been developed by researchers, targeting Vision-Language-Action (VLA) models—versatile robotic systems that can execute various manipulation tasks based on language instructions. This innovative attack utilizes a singular textured 3D object to create specific disruptions in VLA action predictions across different tasks, taking advantage of weaknesses that existing single-task optimized attacks overlook. By employing a differentiable renderer, the method backpropagates gradients from the output actions of the policy to the texture parameters, optimizing the shared texture across various tasks, states, and viewpoints. Documented in arXiv paper 2608.13453, this research reveals potential safety hazards in embodied AI, emphasizing the necessity for stronger defenses in VLA models for safe real-world implementation.
Key facts
- UniTexture is a cross-task universal adversarial texture attack for VLA models.
- It uses a single textured 3D object to induce targeted deviations in action predictions.
- The attack backpropagates gradients from action outputs to texture parameters via a differentiable renderer.
- It jointly optimizes the texture over tasks, instructions, states, and viewpoints.
- Existing attacks are typically optimized for single tasks, leaving cross-task vulnerabilities unexplored.
- VLA models are generalist robotic policies that follow language instructions.
- Adversarial interference may cause unsafe physical behaviors in embodied agents.
- The paper is available on arXiv with ID 2608.13453.
Entities
Institutions
- arXiv