Skip to content

Latest commit

 

History

History
112 lines (89 loc) · 3.6 KB

README.md

File metadata and controls

112 lines (89 loc) · 3.6 KB
Raw

SLING and OVRL-GD/DDPPO-GD

This repository includes the code that can be used to reproduce the results of our paper Last-Mile Embodied Visual Navigation.

teaser

This branch contains code to run a OVRL-GD and DDPPO-GD exploration with SLING. If you are interested in SLING + NRNS-GD/Straight-GD, please see the Experiment Variations section.

🗺 Table of Contents

🔬 Experiment Variations

To reproduce the results from NRNS-GD/Straight-GD with SLING, go to the nrns branch.

git checkout nrns

💿 Installation

To run this codebase please follow the given steps:

  1. Create a new conda env:
conda create -n habitat python=3.7 cmake=3.14.0
conda activate habitat
  1. Install Habitat-Sim version 0.2.1:
conda install habitat-sim=0.2.1 withbullet headless -c conda-forge -c aihabitat
  1. We provide a slightly modified version of Habitat-Lab, install it using:
cd habitat-lab
pip install -r requirements.txt
python setup.py develop --all # install habitat and habitat_baselines
  1. Add the src folder in the pythonpath. This folder contains a few files for the localnav policy used in SLING.
export PYTHONPATH=${PYTHONPATH}:/path/to/src/folder/
  1. Now download the checkpoint files required to run the evaluation over OVRL and DDPPO:
gdown 1OfJ3fMo7II1lNUZs743zalBMSV71JeDC
unzip checkpoint.zip

gdown 1KJ0XHLlg9CxgPBG7OAM1c8L9ZG4AKzxv
unzip data.zip

gdown 1bR4bH7-OrDqo7TItwBcg4BS0Ls08Z1vl
unzip models.zip
  1. Move these folders to the appropriate locations
mv checkpoint/ovrl_best_run/chkp/best_ovrl_ckpt.pth embodied_ssl/checkpoint/ovrl_best_run/chkp/
mv data embodied_ssl/
  1. We will also need the Gibson and MP3D scene datasets. Instructions to download the datasets are available here and here. Move the scene_datasets into the corresponding location.
mv gibson_train_val embodied_ssl/data/scene_datasets/
mv mp3d embodied_ssl/data/scene_datasets/
  1. To run the evaluation, first change the following variable in file sbatch_scripts/run_script.sh on line 5 to the actual path on your machine:
REPO_PATH=/path/to/code/embodied_ssl

🏃 Running DDPPO and OVRL with SLING

Finally run the following command to start the evaluation of SLING + OVRL-GD and SLING + DDPPO-GD on easy, medium and hard subsets of gibson-straight, gibson-curved, matterport-straight and matterport-curved.

./sbatch_scripts/run_ovrl.sh
./sbatch_scripts/run_ddppo.sh

You should get the following results for the gibson-curved x easy datasplit

Average episode reward: 2.4231
Average episode distance_to_goal: 1.2594
Average episode success: 0.6830
Average episode spl: 0.4567
Average episode softspl: 0.4091

📝 Citation

If you use this work, please cite:

@inproceedings{
  wasserman2022lastmile,
  title={Last-Mile Embodied Visual Navigation},
  author={Justin Wasserman and Karmesh Yadav and Girish Chowdhary and Abhinav Gupta and Unnat Jain},
  booktitle={6th Annual Conference on Robot Learning},
  year={2022},
}