Graphormer-Guided Task Planning (GGTP)

🚀 Graphormer-Guided Task Planning is an AI-driven task planning framework that combines LLM-based decision-making with Graphormer-enhanced risk assessment. The overview video is avaliable as: https://www.bilibili.com/video/BV1HDQ7YWEYN

📦 Motivation

Robots may be great at moving around, but can they think before they act?
We introduce Graphormer-based task planning with LLM-driven safety perception,
helping domestic robots navigate complex environments without, well… accidentally microwaving the cat. We aim to enhance task planning in domestic robots by utilzing Graphormer-based and LLM-driven safety perception to dynamically adapt to complex environments.

🛠 Installation

Set 'OPENAI_API_KEY' as an environment variable with your own key.

Clone the repo and install dependencies:

git clone https://github.com/hwj20/GGTP.git
cd GGTP
conda env create -f environment.yml
conda activate GGTP
python main.py

This will run a sample task planning on ai2-thor floor1.

📺 Advertisement Break!

💡 Tired of fixing dependencies? Try our VSCode extension Auto Config Env to automatically install missing packages and debug your Python environment like a pro! 🚀

📌 Reproduce Expeirments

See as 'experiments/README_EXP.md'

📖 Citation

If you find our work helpful, please consider citing our paper:

@misc{huang2025graphormerguidedtaskplanningstatic,
      title={Graphormer-Guided Task Planning: Beyond Static Rules with LLM Safety Perception}, 
      author={Wanjing Huang and Tongjie Pan and Yalan Ye},
      year={2025},
      eprint={2503.06866},
      archivePrefix={arXiv},
      primaryClass={cs.RO},
      url={https://arxiv.org/abs/2503.06866}, 
}

📦 Dataset & Code Availability

Our full dataset and experimental results are available in this repository for reproducibility.
However, due to the large file size, please ensure you have sufficient storage and bandwidth before cloning (300 MB).

💡 For a quick overview, check out our Paper on Arxiv.