VLSA: Vision-Language-Action Models with Plug-and-Play Safety Constraint Layer
Songqiao Hu, Zeyi Liu, Shuang Liu, Jun Cen, Zihan Meng, Xiao He
2025-12-16
Summary
This paper introduces a new system called AEGIS that makes robots safer when they're following instructions based on both vision (what they see) and language (what they're told to do).
What's the problem?
Robots are getting better at doing tasks when given instructions, but it's still hard to make them work safely in real-world environments. A big issue is preventing collisions – robots need to both complete the task *and* avoid bumping into things, which is a tricky balance.
What's the solution?
The researchers created AEGIS, which adds a 'safety layer' to existing robot control systems. This layer uses mathematical rules called control barrier functions to predict potential collisions and adjust the robot's movements to avoid them. Importantly, it doesn't make the robot forget how to follow instructions, it just adds a safety check.
Why it matters?
This work is important because it significantly improves robot safety without sacrificing their ability to perform tasks. They showed AEGIS avoids obstacles much more effectively than current methods and also increases the chances of successfully completing the task, paving the way for robots to be used more reliably in everyday situations.
Abstract
Vision-Language-Action (VLA) models have demonstrated remarkable capabilities in generalizing across diverse robotic manipulation tasks. However, deploying these models in unstructured environments remains challenging due to the critical need for simultaneous task compliance and safety assurance, particularly in preventing potential collisions during physical interactions. In this work, we introduce a Vision-Language-Safe Action (VLSA) architecture, named AEGIS, which contains a plug-and-play safety constraint (SC) layer formulated via control barrier functions. AEGIS integrates directly with existing VLA models to improve safety with theoretical guarantees, while maintaining their original instruction-following performance. To evaluate the efficacy of our architecture, we construct a comprehensive safety-critical benchmark SafeLIBERO, spanning distinct manipulation scenarios characterized by varying degrees of spatial complexity and obstacle intervention. Extensive experiments demonstrate the superiority of our method over state-of-the-art baselines. Notably, AEGIS achieves a 59.16% improvement in obstacle avoidance rate while substantially increasing the task execution success rate by 17.25%. To facilitate reproducibility and future research, we make our code, models, and the benchmark datasets publicly available at https://vlsa-aegis.github.io/.