Offline Reinforcement Learning with Wasserstein Regularization via Optimal Transport Maps

This repository contains the official implementation of the paper Offline Reinforcement Learning with Wasserstein Regularization via Optimal Transport Maps, which introduces Q-DOT.

Overview

Offline reinforcement learning (RL) aims to learn an optimal policy from a static dataset, making it particularly valuable in scenarios where data collection is costly, such as robotics. A major challenge in offline RL is distributional shift, where the learned policy deviates from the dataset distribution, potentially leading to unreliable out-of-distribution actions. To mitigate this issue, regularization techniques have been employed. While many existing methods utilize density ratio-based measures, such as the f-divergence, for regularization, we propose an approach that utilizes the Wasserstein distance, which is robust to out-of-distribution data and captures the similarity between actions. Our method employs input-convex neural networks (ICNNs) to model optimal transport maps, enabling the computation of the Wasserstein distance in a discriminator-free manner, thereby avoiding adversarial training and ensuring stable learning. Our approach demonstrates comparable or superior performance to widely used existing methods on the D4RL benchmark dataset.

Run

Locomotion

python train_offline.py --env_name=halfcheetah-medium-expert-v2 --config=configs/mujoco_config.py  --temp=20

AntMaze

python train_offline.py --env_name=antmaze-large-play-v0 --config=configs/antmaze_config.py --eval_episodes=100 --eval_interval=100000 --temp=20

Kitchen

python train_offline.py --env_name=pen-human-v0 --config=configs/kitchen_config.py --temp=400

Citation

@inproceedings{
  omura2025qdot,
  title={Offline Reinforcement Learning with Wasserstein Regularization via Optimal Transport Maps},
  author={Motoki Omura and Yusuke Mukuta and Kazuki Ota and Takayuki Osa and Tatsuya Harada},
  booktitle={Reinforcement Learning Conference},
  year={2025},
  url={https://openreview.net/forum?id=RxdQBYKjtE}
}

Acknowledgements

Our work builds upon the excellent implementations provided by IQL and DVL, and we would like to express our sincere gratitude to them.