RL-ViGen: A Reinforcement Learning Benchmark for Visual Generalization

This repository is a fork of the original RL-ViGen repository that allows easy installation of just the environments while requiring the strict minimum of dependencies (no PyTorch pinned version restriction, no Hydra, no Tensorboard, no Weights & Biases dependencies).
This is useful to quickly install and instantiate the environments for R&D purposes and quick prototyping.

Note : This repo is a WIP currently

RL-ViGen Authors' Website | RL-ViGen Paper

Portable Installation

The following will install the dependencies and environments required by RL-ViGen.
After the installation, one can import the environments in python and create them for train/eval purposes.

1. Manually download Gibson Full scene datasets : https://github.com/facebookresearch/habitat-sim/blob/main/DATASETS.md#gibson-and-3dscenegraph-datasets
   • This step is manual because you must sign a form to download the data (out of our control).
2. Extract the gibson scene dataset archive to data/datasets/gibson_full/
3. Create a conda environment

   conda create -f environment.yaml -y

   • This will create a conda env named rl-vigen-portable, alternatively, take a look at the environment.yaml file and add the dependencies to your existing conda environment.
4. Activate the conda environment

   conda activate rl-vigen-portable

5. Download external datasets & setup Carla sim

   bash setup.sh

6. Add wrappers and carla to your PYTHONPATH, run the following from the root of the repository :

   export PYTHONPATH=$(pwd):third_party/CARLA_0.9.12/PythonAPI/carla:$PYTHONPATH

At the end, your data/ folder should look like this :

data/
  |
  -- datasets/
        |
        -- gibson_full/
        -- pointnav/

Documentation

See doc/ folder for documentation on how to create environments, a README with minimum working examples is created for each environment.

Important

Many code changes were done so that the environments no longer depend on a Hydra config to be created. Instead, the functions that allow the creation of the environments expect a python dictionary with the config for the environment.
An easy way to use this is to use Hydra in your own repository and simply convert your hydra DictConfig object to a python dict using OmegaConf.to_container(my_hydra_dict_config).
This was done to make you the owner of the Hydra config, you can setup Hydra the way you want and simply pass a dict now.

Code Structure

• envs: various RL-ViGen benchmark environments. In addtion, each sub-folder contains specific README.md for the introduction of the environment.
• third_party: submodules from third parties. We won't frequently change the code in this folder.
• wrappers: includes the wrappers for each environment.

If you find RL-ViGen useful in your research, please consider citing the authors work as follows:

Citations

@article{yuan2023rl,
  title={RL-ViGen: A Reinforcement Learning Benchmark for Visual Generalization},
  author={Yuan, Zhecheng and Yang, Sizhe and Hua, Pu and Chang, Can and Hu, Kaizhe and Wang, Xiaolong and Xu, Huazhe},
  journal={arXiv preprint arXiv:2307.10224},
  year={2023}
}