Roboguard

Safety Guardrail for LLM-Enabled Robots.

Installation

Install the following dependencies

• spot LTL library
• SPINE

Then install this repo:

# if installing a local copy, skip this step
git clone git@github.com:ZacRavichandran/roboguard.git  

cd roboguard
python -m pip install -e .[examples]

Roboguard Example

The provided example demonstrates RoboGuard on a few malicious and benign plans.

You will need to configure an OpenAI API key (please see their documentation).

To run, converted the provided markdown file to a jupyter notebook

cd ./examples
jupytext --to ipynb example.md  -o example.ipynb

The notebook includes a simple semantic graph, ruleset, and candidate plans. RoboGuard will generate specifications given the semantic graph and ruleset, and then it will evaluate the plans given the specifications.

You can also generate plans with the provided LLM-enabled planner.

Running with an LLM-enabled planner

We evaluate RoboGuard when running on top of the SPINE LLM-enabled planner. While SPINE is agnostic to the specific robot platform, we use a Cleapath Jackal with the following configuration

• Intel Realsense
• Ouster OS1
• Nvidia RTX 4000
• Ryzen 4 3600

Citation

If you find this helpful, please cite

@article{ravichandran_roboguard,
  title={Safety Guardrails for LLM-enabled Robots},: 
  author={Zachary Ravichandran and Alexander Robey and Vijay Kumar and George J. Pappas and Hamed Hassani},
  year={2025},
  journal={arXiv preprint arXiv:2503.07885},
  url={https://arxiv.org/abs/2503.07885}
}