gemcollector / TLDA

TLDA

Codebase for “Don’t Touch What Matters: Task-Aware Lipschitz Data Augmentation for Visual Reinforcement Learning”

Enviroment Setup

The following commands can install all depenedcies and libararies for this work:

conda env create -f conda.yaml

Your installed conda environment can be activated as follows:

conda activate tlda

Usage

Running the code as the following steps:

1. install the experimental enviroment:

   setup/install_envs.sh

2. The bash script is ready for running the code. Or you can run the following command in your terminal:

   python3 src/train.py \
     --algorithm tlda \
     --seed 1
   

   This should give you an output of the form like:

   Working directory: logs/cheetah_run/tlda/1
   Evaluating: logs/cheetah_run/tlda/1
   | eval | S: 0 | ER: 26.2285 | ERTEST: 25.3730
   | train | E: 1 | S: 250 | D: 50.8 s | R: 0.0000 | ALOSS: 0.0000 | CLOSS: 0.0000 | AUXLOSS: 0.0000
   | train | E: 2 | S: 500 | D: 0.2 s | R: 14.5838 | ALOSS: 0.0000 | CLOSS: 0.0000 | AUXLOSS: 0.0000
   

3. For the data augmentation, you can download the Place dataset as follows:

   wget http://data.csail.mit.edu/places/places365/places365standard_easyformat.tar

Acknowledgement

Our code implementation is based on DMC-GB. We would like to thank the researchers to open their source code and give detailed instructions.

Citation

If you use our repo in your own work, you can cite our paper as follows:

@article{yuan2022don,
  title={Don't Touch What Matters: Task-Aware Lipschitz Data Augmentation for Visual Reinforcement Learning},
  author={Yuan, Zhecheng and Ma, Guozheng and Mu, Yao and Xia, Bo and Yuan, Bo and Wang, Xueqian and Luo, Ping and Xu, Huazhe},
  journal={arXiv preprint arXiv:2202.09982},
  year={2022}
}