DexMimicGen (Reproduction)

This is a reproduction of DexMimicGen's bimanual data-generation pipeline, not the original authors' release. It reconstructs the data-generation flow and validates it end-to-end on all seven tasks that ship a source dataset (threading, three_piece_assembly, lift_tray, box_cleanup, drawer_cleanup, coffee, and pouring, across the Panda, PandaDex, and GR1 embodiments); transport and can_sort_random ship no source data in this repo and remain static-only. See REPRODUCTION.md for the full reproduction process, root-cause fixes, and success-rate analysis. All credit for the method belongs to the original authors (links below).

Reproduction last updated: 2026-10-17.

This repository contains the official release of simulation environments and datasets for the ICRA 2025 paper "DexMimicGen: Automated Data Generation for Bimanual Dexterous Manipulation via Imitation Learning".

Website: https://dexmimicgen.github.io

Paper: https://arxiv.org/abs/2410.24185

For business inquiries, please submit this form: NVIDIA Research Licensing

---

Getting Started

To use this repository, you need to first install the latest robosuite. For more information, please refer to robosuite.

git clone https://github.com/ARISE-Initiative/robosuite
pip install -e robosuite

Then git clone this repository and install.

git clone https://github.com/NVlabs/dexmimicgen.git
cd dexmimicgen
pip install -e .

After installation, you can run the following command to test the environments.

python scripts/demo_random_action.py --env TwoArmThreading --render

Note: If you are on a headless machine, you can run without the --render flag.

Environments

For detailed information about the environments, please refer to environments.md.

Datasets

You can download the datasets from HuggingFace.

You can also run the script to download the datasets.

python scripts/download_hf_dataset.py --path /path/to/save/datasets

By default, the datasets will be saved to ./datasets.

And then, you can playback one demo in the dataset by running:

python scripts/playback_datasets.py --dataset xxxxx.hdf5 --n 1

Launch Training with robomimic

We provide config and training code to reproduce the BC-RNN result in our paper.

First, you need to install robomimic

git clone https://github.com/ARISE-Initiative/robomimic.git -b dexmimicgen
cd robomimic
pip install -e .

Then you need to generate the config file for the training.

cd dexmimicgen
python scripts/generate_training_config.py --dataset_dir /path/to/datasets --config_dir /path/to/save/config --output_dir /path/to/save/output

By default, it will try to find the datasets in ./datasets, and save the config and output in ./datasets/train_configs/bcrnn_action_dict and ./datasets/train_results/bcrnn_action_dict respectively.

After that, you can run the training script.

cd robomimic
python scripts/train.py --config /path/to/config

License

The code is released under the NVIDIA Source Code License and the datasets are released under CC-BY 4.0.

Citation

Please cite the DexMimicGen paper if you use this code in your work:

@inproceedings{jiang2024dexmimicen,
      title     = {DexMimicGen: Automated Data Generation for Bimanual Dexterous Manipulation via Imitation Learning},
      author    = {Jiang, Zhenyu and Xie, Yuqi and Lin, Kevin and Xu, Zhenjia and Wan, Weikang and Mandlekar, Ajay and Fan, Linxi and Zhu, Yuke},
      booktitle = {2025 IEEE International Conference on Robotics and Automation (ICRA)},
      year      = {2025}
}