This repository contains the Pytorch implementation of Online Deicsion Transformer by Qinqing Zheng, Amy Zhang, and Aditya Grover.

If you use this code for your research, please cite us as:

@article{zheng2022online,
  title={Online decision transformer},
  author={Zheng, Qinqing and Zhang, Amy and Grover, Aditya},
  journal={arXiv preprint arXiv:2202.05607},
  year={2022}
}

conda env create -f conda_env.yml
source activate odt

If you encounter the libstdc++.so.6: version 'GLIBCXX_3.4.xx' not found error, the following command might help:

export LD_LIBRARY_PATH=$LD_LIBRARY_PATH:<path-to-your-conda-env>/lib

I have also found that tensorboard wants protobuf version to be 3.20.x, and this helped

# you might need to uninstall dm-control
pip3 install --upgrade protobuf==3.20.0 

To train an ODT agent for hopper with the medium-v2 dataset:

python main.py

This will produce the exp folder, where all the outputs are going to be logged including tensorboard blobs. One can attach a tensorboard to monitor training by running:

tensorboard --logdir exp

The majority of online-dt is licensed under CC-BY-NC, however portions of the project are available under separate license terms:

• D4RL dataset - Creative Commons Attribution 4.0 License (CC-BY)
• D4RL code, transformers, Lamb - Apache 2.0 License
• stable-baselines3, Gym, decision-transformer - MIT License