SG-Nav: Online 3D Scene Graph Prompting for LLM-based Zero-shot Object Navigation

Paper | Project Page | Video | 中文解读

SG-Nav: Online 3D Scene Graph Prompting for LLM-based Zero-shot Object Navigation
Hang Yin*, Xiuwei Xu* ${}^{†}$, Zhenyu Wu, Jie Zhou, Jiwen Lu${}^{‡}$

* Equal contribution $†$ Project leader $‡$ Corresponding author

We propose a zero-shot object-goal navigation framework by constructing an online 3D scene graph to prompt LLMs. Our method can be directly applied to different kinds of scenes and categories without training.

News

• [2025/02/27]: UniGoal, an extended version of SG-Nav which unifies different goal-oriented navigation tasks, is accepted to CVPR 2025!
• [2024/12/30]: We update the code and simplify the installation.
• [2024/09/26]: SG-Nav is accepted to NeurIPS 2024!

Demo

Scene1:

Scene2:

Demos are a little bit large; please wait a moment to load them. Welcome to the home page for more complete demos and detailed introductions.

Method

Method Pipeline:

Installation

Step 1 (Dataset)

Download Matterport3D scene dataset and object-goal navigation episodes dataset from here.

Set your scene dataset path SCENES_DIR and episode dataset path DATA_PATH in config file configs/challenge_objectnav2021.local.rgbd.yaml.

The structure of the dataset is outlined as follows:

MatterPort3D/
├── mp3d/
│   ├── 2azQ1b91cZZ/
│   │   └── 2azQ1b91cZZ.glb
│   ├── 8194nk5LbLH/
│   │   └── 8194nk5LbLH.glb
│   └── ...
└── objectnav/
    └── mp3d/
        └── v1/
            └── val/
                ├── content/
                │   ├── 2azQ1b91cZZ.json.gz
                │   ├── 8194nk5LbLH.json.gz
                │   └── ...
                └── val.json.gz

Step 2 (Environment)

Create conda environment with python==3.9.

conda create -n SG_Nav python==3.9

Step 3 (Simulator)

Install habitat-sim==0.2.4 and habitat-lab.

conda install habitat-sim==0.2.4 -c conda-forge -c aihabitat
pip install -e habitat-lab

Then replace the agent/agent.py in the installed habitat-sim package with tools/agent.py in our repository.

HABITAT_SIM_PATH=$(pip show habitat_sim | grep 'Location:' | awk '{print $2}')
cp tools/agent.py ${HABITAT_SIM_PATH}/habitat_sim/agent/

Step 4 (Package)

Install pytorch<=1.9, pytorch3d and faiss. Install other packages.

conda install -c pytorch faiss-gpu=1.8.0
pip install torch==1.9.1+cu111 torchvision==0.10.1+cu111 -f https://download.pytorch.org/whl/torch_stable.html
pip install -r requirements.txt
pip install "git+https://github.com/facebookresearch/pytorch3d.git"

Install Grounded SAM.

pip install -e segment_anything
pip install --no-build-isolation -e GroundingDINO
wget -O data/models/sam_vit_h_4b8939.pth https://dl.fbaipublicfiles.com/segment_anything/sam_vit_h_4b8939.pth
wget -O data/models/groundingdino_swint_ogc.pth https://github.com/IDEA-Research/GroundingDINO/releases/download/v0.1.0-alpha/groundingdino_swint_ogc.pth

Install GLIP model and download GLIP checkpoint.

cd GLIP
python setup.py build develop --user
mkdir MODEL
cd MODEL
wget https://huggingface.co/GLIPModel/GLIP/resolve/main/glip_large_model.pth
cd ../../

Install Ollama.

curl -fsSL https://ollama.com/install.sh | sh
ollama pull llama3.2-vision

Evaluation

Run SG-Nav:

python SG_Nav.py --visualize

Citation

@article{yin2024sgnav, 
      title={SG-Nav: Online 3D Scene Graph Prompting for LLM-based Zero-shot Object Navigation}, 
      author={Hang Yin and Xiuwei Xu and Zhenyu Wu and Jie Zhou and Jiwen Lu},
      journal={arXiv preprint arXiv:2410.08189},
      year={2024}
}