Vision language action models in robots face strong universal attacks
Exploiting Vulnerabilities: Universal Adversarial Attacks on Vision-Language-Action Models in Robotics
RoboticsCryptography and SecurityComputer Vision and Pattern Recognition
Summary
Robots that can see, understand language, and act based on that information are becoming more common. This paper shows that by placing a specially designed textured ball in a robot's view, the robot's ability to complete tasks drops significantly. The researchers created a way to confuse the robot at many levels, from planning movements to controlling actions. Their tests worked both in computer simulations and with real robots, showing the method lowers success rates a lot, especially in tricky tasks.
What this means in practice
- •For robotics security teams: Detect and defend against universal visual disruptions that impair robot task success to improve operational safety.
- •For industrial robot operators: Recognize risk factors from adversarial objects in a robot’s environment that could unintentionally reduce performance or cause failures.
Authors
Songhua Yang, Ziyu Liu, Yuanwei Liu, Xuetao Li, Xuanye Fei, He Huang, Zheng Wang, Miao Li
Abstract
Recently, Vision-Language-Action (VLA) models have revolutionized robotic manipulation by seamlessly integrating visual perception, language understanding, and action generation in an end-to-end learning framework. However, since these models are designed to interact directly with the physical world and humans, their security is critical, and even small vulnerabilities can lead to catastrophic failures. In this work, we propose the Universal Adversarial Object, a sphere with optimized surface texture that significantly degrades task success rates when placed within the robot's field of view. Specifically, our approach introduces a multi-level attack framework that jointly disrupts trajectory planning, task execution, and action control. We validate our method in both simulated and real-world robotic settings. Experimental results demonstrate that the adversarial object reduces the average task success rates by 31.2%-39.9% for two representative VLA models (Pi0 and RDT), with success rates dropping to near zero in complex scenarios. Index Terms--Vision-Language-Action models, adversarial attack, robotic security, universal adversarial object