Awesome

DISCO

This repository contains the official implementation associated with the paper:

<b>DISCO: Embodied Navigation and Interaction via Differentiable Scene Semantics and Dual-level Control</b>.

In European Conference on Computer Vision 2024 (ECCV 2024)

[arXiv]

Install

We use python3.8 in this work. You can setup the environment and install dependencies in requirements.txt via

conda create -n disco python==3.8
conda activate disco
pip install -r requirements.txt

Download Data

Next, download ALFRED DATA using the download_data.sh script. The Lite json version is ok in this repository.

cd data
sh download_data.sh json

It will finally in the following structure:

----data
   | ----json_2.1.0
      | ----train
      | ----valid_seen
      | ----valid_unseen
      | ----tests_seen
      | ----tests_unseen
   | ----splits
      | ----oct21.json

Download checkpoints

Then, download pretrained checkpoints from Google Drive. Unzip the file and put them into weights

Headless Machines

For headless machines, use the script startx.py to launch a headless render to enable the unity simulation.

python startx.py --gpu 0 --display [DISPLAY]

Running

Now, everything is prepared if you set up correctly. You can run DISCO using the folling commands

python run.py --n_proc 4 --split valid_seen   --x_display [DISPLAY] --name [EXP_NAME]
python run.py --n_proc 4 --split valid_unseen --x_display [DISPLAY] --name [EXP_NAME] 

The [DISPLAY] number should equal to the number when you start the X server

The result will be saved into ./logs/[EXP_NAME].

Evaluation

Last, you can evaluate the performances via

python evaluate.py --name [EXP_NAME] --split valid_unseen
python evaluate.py --name [EXP_NAME] --split valid_seen

Citation

If you find our work useful, please cite:

@misc{xu2024disco,
      title={DISCO: Embodied Navigation and Interaction via Differentiable Scene Semantics and Dual-level Control}, 
      author={Xinyu Xu and Shengcheng Luo and Yanchao Yang and Yong-Lu Li and Cewu Lu},
      year={2024},
      eprint={2407.14758},
      archivePrefix={arXiv},
      primaryClass={cs.CV},
      url={https://arxiv.org/abs/2407.14758}, 
}