VLSA: Vision-Language-Action Models with
Plug-and-Play Safety Constraint Layer

Figure 1: Functional architecture of VLA and VLSA models.

Overview

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.

Figure 2: Workflow of the AEGIS model.

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. Extensive experiments demonstrate that AEGIS achieves:

We summarize the main contributions as follows:

• We propose AEGIS, the first approach that integrates CBFs into VLA models to enforce explicit safety constraints. By introducing a plug-and-play SC layer, AEGIS effectively addresses the limitation of end-to-end models regarding the lack of safety guarantees. Our approach has the potential to enhance the safety of any existing VLA model without the need for retraining.
• We design a vision-language based safety assessment module and an action-driven safety-guaranteed control module, which bridge visual perception and semantic understanding with safety-guaranteed control. By leveraging a CBF-QP solver, the safety of the adjusted actions is theoretically guaranteed.
• We establish SafeLIBERO, a comprehensive safety-critical benchmark derived from the LIBERO dataset, encompassing 32 diverse scenarios and 1600 episodes. Extensive experiments demonstrate that AEGIS achieves superior performance over state-of-the-art baselines.

Demonstrations

We compare OpenVLA-OFT, $\pi_{0.5}$, and Ours across 32 scenarios on our constructed benchmark.

BibTeX

@misc{hu2025vlsavisionlanguageactionmodelsplugandplay,
      title={VLSA: Vision-Language-Action Models with Plug-and-Play Safety Constraint Layer},
      author={Songqiao Hu and Zeyi Liu and Shuang Liu and Jun Cen and Zihan Meng and Xiao He},
      year={2025},
      eprint={2512.11891},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2512.11891},
}