OpenFMNav

OpenFMNav copied to clipboard

OpenFMNav: Towards Open-Set Zero-Shot Object Navigation via Vision-Language Foundation Models

NAACL 2024 Findings

This is the official repository of OpenFMNav: Towards Open-Set Zero-Shot Object Navigation via Vision-Language Foundation Models.

Setup

Dataset Preparation

Please follow HM3DSem to download the dataset and prepare the data. The data format should be:

data/
├── objectgoal_hm3d/
│   ├── train/
│   ├── val/
│   └── val_mini/
├── scene_datasets/
│   └── hm3d/
│       ├── minival/
│       └── val/
├── versioned_data/
├── matterport_category_mappings.tsv
└── object_norm_inv_perplexity.npy

Checkpoints

Please checkout Grounded-SAM to download groundingdino_swint_ogc.pth and sam_vit_h_4b8939.pth and put them into Grounded_SAM/.

Dependencies

1. Python & PyTorch

   This code is tested on Python 3.9.16 on Ubuntu 20.04, with PyTorch 1.11.0+cu113.

2. Habitat-Sim & Habitat-Lab

   # Habitat-Sim
   git clone https://github.com/facebookresearch/habitat-sim.git
   cd habitat-sim; git checkout tags/challenge-2022; 
   pip install -r requirements.txt; 
   python setup.py install --headless
   
   # Habitat-Lab
   git clone https://github.com/facebookresearch/habitat-lab.git
   cd habitat-lab; git checkout tags/challenge-2022; 
   pip install -e .
   

3. Grounded-SAM

   Please checkout Grounded-SAM to install the dependencies.

4. Others

   pip install -r requirements.txt
   

OpenAI API keys

You will need an OpenAI API key to use this repo. Please touch apikey.txt and paste your API key in the file.

Running

Example

An example command to run the pipeline:

CUDA_VISIBLE_DEVICES=0 python main.py --split val --eval 1 --auto_gpu_config 0 --prompt_type scoring \
-n 1 --num_eval_episodes 100 --text_threshold 0.55 --boundary_coeff 12 --start_episode 0 --tag_freq 100 \
--use_gtsem 0 --num_local_steps 20 --print_images 1 --exp_name test

Visualization

To make a demo video on your saved images, you can either use ffmpeg to make separate videos or use

python make_demo.py --exp_name test # add `--delete_img` to delete images after making video

to make batched videos.

Acknowledgements

This repo is heavily based on L3MVN. We thank the authors for their great work.

Citation

If you find this work helpful, please consider citing:

@article{kuang2024openfmnav,
  title={OpenFMNav: Towards Open-Set Zero-Shot Object Navigation via Vision-Language Foundation Models},
  author={Kuang, Yuxuan and Lin, Hai and Jiang, Meng},
  journal={arXiv preprint arXiv:2402.10670},
  year={2024}
}