| 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} | |
| } | |
| ``` |