Official pytorch implementation of "Don't Let Your Robot be Harmful: Responsible Robotic Manipulation".

• [28/11] We released the Safety-as-policy on Github.

Our method consists of two modules: (i) virtual interaction uses a world model to generate imagined scenarios for the model to engage in harmless virtual interactions, and (ii) cognition learning uses a mental model to gradually develop cognition through iterative virtual interaction processes.

Heat food with microwaves.	Insert fork into block.	Push the phone aside.
Watering flower with watering can.	Light the package with a cigarette lighter.	Store the lighter properly.

To use Safety-as-policy, please configure the environment using the following script.

conda create -n sap python=3.10

conda activate sap
pip install -r requirements.txt

Finally, please manually configure the API key and endpoint. We recommend using Azure OpenAI Service, as it allows for the manual disabling of filters.

python main.py

Please download our dataset via this link.

If you find this work useful for your research and applications, please cite using this BibTeX:

@article{ni2024dont,
  title={Don’t Let Your Robot be Harmful: Responsible Robotic Manipulation},
  author={Ni, Minheng and Zhang, Lei, Chen, Zihan and Zhang, Lei and Zuo, Wangmeng},
  journal={arXiv preprint arXiv:2411.18289},
  year={2024}
}