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 locomotion environment code is available at mangobench-locomotion.

Installation

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

cd robofactory
pip install -r requirements_ogbench.txt

Data

Follow RoboFactory to generate datasets and process data.

Note: The training code automatically converts processed imitation learning datasets into standard RL datasets.

Running

bash ./policy/OGCRL/train_eval_acp.sh

Note: The first training run requires saving goals for evaluation. Add save_goal=True and save_goal_path=policy/OGCRL/ogcrl/goal to your training command.

Example:

python policy/OGCRL/train.py \
    --config-name=robot_gc.yaml \
    task.name=LiftBarrier-rf \
    task.dataset.zarr_path=data/zarr_data/LiftBarrier-rf_Agent0_150.zarr \
    training.debug=False \
    training.seed=100 \
    training.device=cuda:0 \
    exp_name=LiftBarrier-rf-robot_gc-train \
    logging.mode=online \
    env_name=robofactory \
    agent=hiql \
    agent.high_alpha=3.0 \
    agent.low_alpha=3.0 \
    train_steps=15000 \
    log_interval=1000 \
    save_interval=15000 \
    agent.batch_size=256 \
    agent.encoder=impala_small \
    agent.low_actor_rep_grad=True \
    agent.p_aug=0.5 \
    agent.subgoal_steps=10 \
    observation=visual \
    save_dir=expacp \
    save_goal=True \
    save_goal_path=policy/OGCRL/ogcrl/goal

Evaluate

If you want to reproduce the results in our paper, you can download our checkpoints in huggingface and evaluate it.

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.