MangoBench: A Benchmark for Multi-Agent Goal-Conditioned Offline Reinforcement Learning

Official implementation of MangoBench (CVPR 2026).

MangoBench is the first fully cooperative multi-goal benchmark for offline MARL, covering 47 tasks across locomotion and bimanual manipulation. See the project page for videos and more details about the environments, tasks, and baseline algorithms.

Note: The manipulation environment code is available at mangobench-manipulation.

Installation

conda create -n mangobench python=3.10
conda activate mangobench

Follow OGBench to set up the base environment, then install additional dependencies:

cd impls
pip install -r requirements.txt

Running

bash ./impls/hyperparameters_multi.sh

Citation

If you find this work useful, please cite:

@inproceedings{Wang2026MangoBench,
  title={MangoBench: A Benchmark for Multi-Agent Goal-Conditioned Offline Reinforcement Learning},
  author={Wang, Yi and Zhong, Ningze and Fu, Zhiheng and Wang, Longguang and Zhang, Ye and Guo, Yulan},
  booktitle={IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR)},
  year={2026}
}

License

This project is licensed under the MIT License - see the LICENSE file for details.