Offline Reinforcement Learning with Domain-Unlabeled Data

Official implementation of Offline Reinforcement Learning with Domain-Unlabeled Data accepted at RLC 2025. The offline RL implementations are borrowed from JAX-CORL.

How to install

git clone https://github.com/nissymori/PUORL.gitn

pip install -r requirements.txt

How to reproduce

1. Download data

python download.py

2. Remove prefix

./remove_prefix.sh

3. Train classifier at sh/classifier

./pu.sh

4. Train agent at sh/offline/...

./body_mass.sh

Wandb results for reported result

• td3bc
• iql
• classifier

Citeation

@article{nishimori2025offline,
    title={Offline Reinforcement Learning with Domain-Unlabeled Data},
    author={Nishimori, Soichiro and Cai, Xin-Qiang and Ackermann, Johannes and Sugiyama, Masashi},
    journal={Reinforcement Learning Journal},
    volume={6},
    pages={1773--1793},
    year={2025}
}