FloNa：Floor Plan Guided Embodied Visual Navigation

Jiaxin Li, Weiqi Huang, Zan Wang, Wei Liang, Huijun Di, Feng Liu

Welcome to FloNa! This repository is the official implementation of paper "FloNa: Floor Plan Guided Embodied Visual Navigation".

Table of Contents

• 1. Environment Setup
• 2. Dataset
• 3. Training
• 4. Testing in simulator

1. Environment Setup

a. Create a new conda environment:

git clone https://github.com/GauleeJX/flodiff.git
cd flodiff
conda env create -n flodiff -f environment.yaml
conda activate flodiff
pip install -e .

b. Install the diffusion_policy:

cd /path/to/flodiff
git clone https://github.com/real-stanford/diffusion_policy.git
pip install -e diffusion_policy/

2. Dataset

a. Download our pre-collected dataset:

We collected navigation episodes on Gibson static scenes for training and evaluation. The structure of the episode data is organized as follows:

|--<train>
|   |--<scene_0>
|   |-----floor_plan.png        # the floor plan of the scene
|   |-----<traj0>               # one episode in the scene
|   |----------traj_0.npy           # a .npy file saving ground truth 2d position and 2d orientation for each frame
|   |----------traj_0.txt           # a .txt file saving ground truth 2D position and orientation for each frame
|   |----------traj_0.png           # a .png file showing the trajectory on the floor plan
|   |----------obs_0.png            # frame 0
|   |----------obs_1.png            # frame 1
|   |----------...
|   |-----<traj1>
|   |----------traj_1.npy
|   |----------traj_1.txt
|   |----------traj_1.png
|   |----------obs_0.png
|   |----------obs_1.png
|   |----------...
|   |-----...
|   |--<scene_1>
|   |-----...
|--<test>
|   |...

You can download it from BaiduDisk. Please note that the dataset is approximately 500 GB, so make sure you have sufficient disk space.

b. Unpack this tar archive:

cat /path/to/dataset/dataset.tar_* > /path/to/dataset/dataset.tar
tar -xvf dataset.tar -C /path/to/flodiff/datasets
mkdir /path/to/flodiff/datasets/trav_maps
unzip trav_maps.zip -d /path/to/flodiff/datasets/trav_maps

3. Training

To train your own model, simply run the following command:

cd /path/to/flodiff/training
python train.py

The model is saved to /training/log/flona by default. If you need to change the save location, modify the flona.yaml configuration file. Note: This code is designed for training on a single GPU. If you wish to use multiple GPUs, you'll need to modify the code accordingly.

4. Testing in simulator

a. Install iGibson:

cd /path/to/flodiff
git clone https://github.com/StanfordVL/iGibson --recursive
cd iGibson
pip install -e . # This step takes about 4 minutes

b. Download scenes:

Please refer to this link to download the Gibson static scenes. We recommend saving the downloaded scenes in the following directory: /path_to_flodiff/iGibson/igibson/data/g_dataset.

c. Testing:

mkdir /path/to/flodiff/iGibson/igibson/scripts
mv /path/to/flodiff/test_flona_gtpos.py  /path/to/flodiff/iGibson/scripts/
cd /path/to/flodiff/iGibson
python -m igibson.scripts.test_flona_gtpos

For evaluating results, run:

python /path/to/flodiff/results/evaluate.py

🔗 Citation

If you find our work helpful, please cite:

@inproceedings{li2025flona,
    title={FloNa: Floor Plan Guided Embodied Visual Navigation},
    author={Li, Jiaxin and Huang, Weiqi and Wang, Zan and Liang, Wei and Di, Huijun and Liu, Feng},
    booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
    year={2025}
}

License

This project is licensed under the MIT licensed.

Acknowledgements

• iGibson: We adopt iGibson and the Gibson v2 dataset as our simulation environment and scene, leveraging their comprehensive capabilities for realistic and diverse scenario modeling.

• NoMad: Partial code is borrowed from NoMad.