README.md

---
license: apache-2.0
pipeline_tag: robotics
---

# SafeVLA: Towards Safety Alignment of Vision-Language-Action Model via Constrained Learning

[**Project Page**](https://pku-safevla.github.io) | [**Paper**](https://arxiv.org/abs/2503.03480) | [**GitHub**](https://github.com/PKU-Alignment/SafeVLA)

SafeVLA is a family of vision-language-action models (VLAs) aligned for safety using an integrated safety approach (ISA). By modeling safety requirements and utilizing safe reinforcement learning within a constrained Markov decision process (CMDP) paradigm, SafeVLA explicitly integrates safety constraints into generalist robot policies. The model achieves an effective balance between task performance and safety violation mitigation, handling long-tail risks and extreme failure scenarios in mobile manipulation tasks.

## Citation

If you find our code or models useful in your work, please cite our paper:

```bibtex
@inproceedings{zhang25safevla,
    title={SafeVLA: Towards Safety Alignment of Vision-Language-Action Model via Constrained Learning},
    author={Borong Zhang and Yuhao Zhang and Jiaming Ji and Yingshan Lei and Josef Dai and Yuanpei Chen and Yaodong Yang},
    booktitle={Thirty-ninth Conference on Neural Information Processing Systems},
    year={2025},
    url={https://openreview.net/forum?id=dt940loCBT}
}
```