ManipTrans: Efficient Dexterous Bimanual Manipulation Transfer via Residual Learning

Accepted to CVPR 2025

Kailin Li, Puhao Li, Tengyu Liu, Yuyang Li, Siyuan Huang

---

📑 Table of Contents

1. Installation
2. Prerequisites
3. Usage
4. 🤗 Extending to New Hand-Object Datasets 🤗
5. 🤗 Extending to New Dexterous Hand URDF Files 🤗
6. Check out Our Paper
7. Acknowledgement
8. License

---

🛠️ Installation

Steps:

1. Clone the repository and initialize submodules:

   git clone https://github.com/ManipTrans/ManipTrans.git
   git submodule init && git submodule update

2. Create a virtual environment named maniptrans with Python 3.8. Note that IsaacGym only supports Python versions up to 3.8.

   conda create -y -n maniptrans python=3.8
   conda activate maniptrans
   pip install torch==1.13.1+cu117 torchvision==0.14.1+cu117 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu117

3. Download IsaacGym Preview 4 from the official website and follow the installation instructions in the documentation. Test the installation by running an example script, such as joint_monkey.py, located in the python/examples directory.
4. Install additional dependencies.

   pip install git+https://github.com/ZhengyiLuo/smplx.git
   pip install git+https://github.com/KailinLi/bps_torch.git
   pip install fvcore~=0.1.5
   pip install --no-index --no-cache-dir pytorch3d==0.7.3 -f https://dl.fbaipublicfiles.com/pytorch3d/packaging/wheels/py38_cu117_pyt1131/download.html
   pip install -r requirements.txt
   pip install numpy==1.23.5 # downgrade numpy to 1.23.5 to avoid compatibility issues

---

📋 Prerequisites

Steps:

Demo data for Grab dataset

1. For fair comparisons with QuasiSim, we directly use the demo data from the official Grab dataset repository. Download it from the official link and extract it into the data/grab_demo directory.

2. Copy the demo URDF file assets/obj_urdf_example.urdf to data/grab_demo/102/102_obj.urdf.

3. The directory structure should look like this:

   data
   └── grab_demo
       └── 102
           ├── 102_obj.npy
           ├── 102_obj.obj
           ├── 102_obj.urdf
           ├── 102_sv_dict.npy
           └── 102_sv_dict_st_0_ed_108.npy
   

OakInk-V2 dataset

1. Download the OakInk-V2 dataset from its official website and extract it into the data/OakInk-v2 directory. (You may skip downloading images; only annotated motion data is required.)

2. For each object mesh in data/OakInk-v2/object_preview/align_ds, generate the COACD file by running:

   python maniptrans_envs/lib/utils/coacd_process.py -i data/OakInk-v2/object_preview/align_ds/xx/xx.obj -o data/OakInk-v2/coacd_object_preview/align_ds/xx/xx.obj --max-convex-hull 32 --seed 1 -mi 2000 -md 5 -t 0.07

3. For each generated COACD file in data/OakInk-v2/coacd_object_preview/align_ds, create a corresponding URDF file based on assets/obj_urdf_example.urdf.

4. Download the body_upper_idx.pt file from the official website and place it in the data/smplx_extra directory.

5. The directory structure should look like this:

   data
   ├── smplx_extra
   │   └── body_upper_idx.pt
   └── OakInk-v2
       ├── anno_preview
       ├── coacd_object_preview
       ├── data
       ├── object_preview
       └── program
   

MISC

1. Download MANO model files from the official website and extract them into the data/mano_v1_2 directory.
2. Download the SMPL-X model files from the official website and extract them into the data/body_utils/body_models directory.

Imitator Checkpoints

1. Download the pre-trained imitator checkpoints imitator_ckp from the official website and place them in the assets directory.

▶️ Usage

✋ Training Single-Hand Policies

Note: Due to licensing restrictions, we cannot provide some dexterous hand URDF files (e.g., XHand and Inspire FTP). If you have the required authorization from the respective robot companies, please contact us for configuration files.

1. Preprocessing (Optional but recommended)
   Processes the trajectory to obtain a series of non-colliding, near-object states for reference state initialization (RSI), enhancing training stability and efficiency.

   # for Inspire Hand
   python main/dataset/mano2dexhand.py --data_idx g0 --dexhand inspire --headless --iter 2000

   For other hands:

   # for Shadow Hand
   python main/dataset/mano2dexhand.py --data_idx g0 --dexhand shadow --headless --iter 3000
   # for Arti-Mano
   python main/dataset/mano2dexhand.py --data_idx g0 --dexhand artimano --headless --iter 2000
   # for Allegro Hand
   python main/dataset/mano2dexhand.py --data_idx g0 --dexhand allegro --headless --iter 4000
   # for XHand
   python main/dataset/mano2dexhand.py --data_idx g0 --dexhand xhand --headless --iter 3000
   # for Inspire FTP Hand
   python main/dataset/mano2dexhand.py --data_idx g0 --dexhand inspireftp --headless --iter 4000

2. Training

   Use the following command to train single-hand policies:

   # for Inspire Hand
   python main/rl/train.py task=ResDexHand dexhand=inspire side=RH headless=true num_envs=4096 learning_rate=2e-4 test=false randomStateInit=true rh_base_model_checkpoint=assets/imitator_rh_inspire.pth lh_base_model_checkpoint=assets/imitator_lh_inspire.pth dataIndices=[g0] early_stop_epochs=100 actionsMovingAverage=0.4 experiment=cross_g0_inspire

   For other hands:

   # for Shadow Hand (with imitator in PID-controlled mode)
   python main/rl/train.py task=ResDexHand dexhand=shadow side=RH headless=true num_envs=4096 learning_rate=2e-4 test=false randomStateInit=true rh_base_model_checkpoint=assets/imitator_pid_rh_shadow.pth lh_base_model_checkpoint=assets/imitator_pid_lh_shadow.pth dataIndices=[g0] early_stop_epochs=100 actionsMovingAverage=0.4 usePIDControl=True experiment=cross_g0_shadow_pid
   # for Arti-Mano (with imitator in PID-controlled mode)
   python main/rl/train.py task=ResDexHand dexhand=artimano side=RH headless=true num_envs=4096 learning_rate=2e-4 test=false randomStateInit=true rh_base_model_checkpoint=assets/imitator_pid_rh_artimano.pth lh_base_model_checkpoint=assets/imitator_pid_lh_artimano.pth dataIndices=[g0] early_stop_epochs=100 actionsMovingAverage=0.4 usePIDControl=True experiment=cross_g0_artimano_pid
   # for Allegro Hand (with imitator in PID-controlled mode)
   python main/rl/train.py task=ResDexHand dexhand=allegro side=RH headless=true num_envs=4096 learning_rate=2e-4 test=false randomStateInit=true rh_base_model_checkpoint=assets/imitator_pid_rh_allegro.pth lh_base_model_checkpoint=assets/imitator_pid_lh_allegro.pth dataIndices=[g0] early_stop_epochs=100 actionsMovingAverage=0.4 usePIDControl=True experiment=cross_g0_allegro_pid
   # for XHand (with imitator in PID-controlled mode)
   python main/rl/train.py task=ResDexHand dexhand=xhand side=RH headless=true num_envs=4096 learning_rate=2e-4 test=false randomStateInit=true rh_base_model_checkpoint=assets/imitator_pid_rh_xhand.pth lh_base_model_checkpoint=assets/imitator_pid_lh_xhand.pth dataIndices=[g0] early_stop_epochs=100 actionsMovingAverage=0.4 usePIDControl=True experiment=cross_g0_xhand_pid
   # for Inspire FTP Hand (with imitator in PID-controlled mode)
   python main/rl/train.py task=ResDexHand dexhand=inspireftp side=RH headless=true num_envs=4096 learning_rate=2e-4 test=false randomStateInit=true rh_base_model_checkpoint=assets/imitator_pid_rh_inspireftp.pth lh_base_model_checkpoint=assets/imitator_pid_lh_inspireftp.pth dataIndices=[g0] early_stop_epochs=100 actionsMovingAverage=0.4 usePIDControl=True experiment=cross_g0_inspireftp_pid
   
   # for Shadow Hand (with imitator in 6D-Force mode)
   python main/rl/train.py task=ResDexHand dexhand=shadow side=RH headless=true num_envs=4096 learning_rate=2e-4 test=false randomStateInit=true rh_base_model_checkpoint=assets/imitator_rh_shadow.pth lh_base_model_checkpoint=assets/imitator_lh_shadow.pth dataIndices=[g0] early_stop_epochs=100 actionsMovingAverage=0.4 experiment=cross_g0_shadow
   # for Arti-Mano (with imitator in 6D-Force mode)
   python main/rl/train.py task=ResDexHand dexhand=artimano side=RH headless=true num_envs=4096 learning_rate=2e-4 test=false randomStateInit=true rh_base_model_checkpoint=assets/imitator_rh_artimano.pth lh_base_model_checkpoint=assets/imitator_lh_artimano.pth dataIndices=[g0] early_stop_epochs=100 actionsMovingAverage=0.4 experiment=cross_g0_artimano
   # for Allegro Hand (with imitator in 6D-Force mode)
   python main/rl/train.py task=ResDexHand dexhand=allegro side=RH headless=true num_envs=4096 learning_rate=2e-4 test=false randomStateInit=true rh_base_model_checkpoint=assets/imitator_rh_allegro.pth lh_base_model_checkpoint=assets/imitator_lh_allegro.pth dataIndices=[g0] early_stop_epochs=100 actionsMovingAverage=0.4 experiment=cross_g0_allegro
   # for XHand (with imitator in 6D-Force mode)
   python main/rl/train.py task=ResDexHand dexhand=xhand side=RH headless=true num_envs=4096 learning_rate=2e-4 test=false randomStateInit=true rh_base_model_checkpoint=assets/imitator_rh_xhand.pth lh_base_model_checkpoint=assets/imitator_lh_xhand.pth dataIndices=[g0] early_stop_epochs=100 actionsMovingAverage=0.4 experiment=cross_g0_xhand
   # for Inspire FTP Hand (with imitator in 6D-Force mode)
   python main/rl/train.py task=ResDexHand dexhand=inspireftp side=RH headless=true num_envs=4096 learning_rate=2e-4 test=false randomStateInit=true rh_base_model_checkpoint=assets/imitator_rh_inspireftp.pth lh_base_model_checkpoint=assets/imitator_lh_inspireftp.pth dataIndices=[g0] early_stop_epochs=100 actionsMovingAverage=0.4 experiment=cross_g0_inspireftp

   The early_stop_epochs is dependent on the complexity of the task. For simple tasks, you can set it to 100, while for more complex tasks (e.g. cap the pen), you may largely increase it.

3. Test

   After training, test the model using the following command:

   # for Inspire Hand
   python main/rl/train.py task=ResDexHand dexhand=inspire side=RH headless=false num_envs=4 learning_rate=2e-4 test=true randomStateInit=false rh_base_model_checkpoint=assets/imitator_rh_inspire.pth lh_base_model_checkpoint=assets/imitator_lh_inspire.pth dataIndices=[g0] actionsMovingAverage=0.4 checkpoint=runs/cross_g0_inspire__xxxxxx/nn/cross_g0_inspire.pth

   For other hands:

   # for Shadow Hand (with imitator in PID-controlled mode)
   python main/rl/train.py task=ResDexHand dexhand=shadow side=RH headless=false num_envs=4 learning_rate=2e-4 test=true randomStateInit=false rh_base_model_checkpoint=assets/imitator_pid_rh_shadow.pth lh_base_model_checkpoint=assets/imitator_pid_lh_shadow.pth dataIndices=[g0] actionsMovingAverage=0.4 usePIDControl=True checkpoint=runs/cross_g0_shadow_pid__xxxxxx/nn/cross_g0_shadow_pid.pth
   # for Arti-Mano (with imitator in PID-controlled mode)
   python main/rl/train.py task=ResDexHand dexhand=artimano side=RH headless=false num_envs=4 learning_rate=2e-4 test=true randomStateInit=false rh_base_model_checkpoint=assets/imitator_pid_rh_artimano.pth lh_base_model_checkpoint=assets/imitator_pid_lh_artimano.pth dataIndices=[g0] actionsMovingAverage=0.4 usePIDControl=True checkpoint=runs/cross_g0_artimano_pid__xxxxxx/nn/cross_g0_artimano_pid.pth
   # for Allegro Hand (with imitator in PID-controlled mode)
   python main/rl/train.py task=ResDexHand dexhand=allegro side=RH headless=false num_envs=4 learning_rate=2e-4 test=true randomStateInit=false rh_base_model_checkpoint=assets/imitator_pid_rh_allegro.pth lh_base_model_checkpoint=assets/imitator_pid_lh_allegro.pth dataIndices=[g0] actionsMovingAverage=0.4 usePIDControl=True checkpoint=runs/cross_g0_allegro_pid__xxxxxx/nn/cross_g0_allegro_pid.pth
   # for XHand (with imitator in PID-controlled mode)
   python main/rl/train.py task=ResDexHand dexhand=xhand side=RH headless=false num_envs=4 learning_rate=2e-4 test=true randomStateInit=false rh_base_model_checkpoint=assets/imitator_pid_rh_xhand.pth lh_base_model_checkpoint=assets/imitator_pid_lh_xhand.pth dataIndices=[g0] actionsMovingAverage=0.4 usePIDControl=True checkpoint=runs/cross_g0_xhand_pid__xxxxxx/nn/cross_g0_xhand_pid.pth
   # for Inspire FTP Hand (with imitator in PID-controlled mode)
   python main/rl/train.py task=ResDexHand dexhand=inspireftp side=RH headless=false num_envs=4 learning_rate=2e-4 test=true randomStateInit=false rh_base_model_checkpoint=assets/imitator_pid_rh_inspireftp.pth lh_base_model_checkpoint=assets/imitator_pid_lh_inspireftp.pth dataIndices=[g0] actionsMovingAverage=0.4 usePIDControl=True checkpoint=runs/cross_g0_inspireftp_pid__xxxxxx/nn/cross_g0_inspireftp_pid.pth
   
   # for Shadow Hand (with imitator in 6D-Force mode)
   python main/rl/train.py task=ResDexHand dexhand=shadow side=RH headless=false num_envs=4 learning_rate=2e-4 test=true randomStateInit=false rh_base_model_checkpoint=assets/imitator_rh_shadow.pth lh_base_model_checkpoint=assets/imitator_lh_shadow.pth dataIndices=[g0] actionsMovingAverage=0.4 checkpoint=runs/cross_g0_shadow__xxxxxx/nn/cross_g0_shadow.pth
   # for Arti-Mano (with imitator in 6D-Force mode)
   python main/rl/train.py task=ResDexHand dexhand=artimano side=RH headless=false num_envs=4 learning_rate=2e-4 test=true randomStateInit=false rh_base_model_checkpoint=assets/imitator_rh_artimano.pth lh_base_model_checkpoint=assets/imitator_lh_artimano.pth dataIndices=[g0] actionsMovingAverage=0.4 checkpoint=runs/cross_g0_artimano__xxxxxx/nn/cross_g0_artimano.pth
   # for Allegro Hand (with imitator in 6D-Force mode)
   python main/rl/train.py task=ResDexHand dexhand=allegro side=RH headless=false num_envs=4 learning_rate=2e-4 test=true randomStateInit=false rh_base_model_checkpoint=assets/imitator_rh_allegro.pth lh_base_model_checkpoint=assets/imitator_lh_allegro.pth dataIndices=[g0] actionsMovingAverage=0.4 checkpoint=runs/cross_g0_allegro__xxxxxx/nn/cross_g0_allegro.pth
   # for XHand (with imitator in 6D-Force mode)
   python main/rl/train.py task=ResDexHand dexhand=xhand side=RH headless=false num_envs=4 learning_rate=2e-4 test=true randomStateInit=false rh_base_model_checkpoint=assets/imitator_rh_xhand.pth lh_base_model_checkpoint=assets/imitator_lh_xhand.pth dataIndices=[g0] actionsMovingAverage=0.4 checkpoint=runs/cross_g0_xhand__xxxxxx/nn/cross_g0_xhand.pth
   # for Inspire FTP Hand (with imitator in 6D-Force mode)
   python main/rl/train.py task=ResDexHand dexhand=inspireftp side=RH headless=false num_envs=4 learning_rate=2e-4 test=true randomStateInit=false rh_base_model_checkpoint=assets/imitator_rh_inspireftp.pth lh_base_model_checkpoint=assets/imitator_lh_inspireftp.pth dataIndices=[g0] actionsMovingAverage=0.4 checkpoint=runs/cross_g0_inspireftp__xxxxxx/nn/cross_g0_inspireftp.pth

🤲 Training BiManual Policies

1. Preprocessing

   Preprocess data for both hands:

   # for Inspire Hand
   python main/dataset/mano2dexhand.py --data_idx 476@0 --side right --dexhand inspire --headless --iter 7000
   python main/dataset/mano2dexhand.py --data_idx 476@0 --side left --dexhand inspire --headless --iter 7000
   # for other hands, just replace `inspire` with the corresponding hand name

2. Training Train bi-manual policies:

   python main/rl/train.py task=ResDexHand dexhand=inspire side=BiH headless=true num_envs=4096 learning_rate=2e-4 test=false randomStateInit=true dataIndices=[476@0] rh_base_model_checkpoint=assets/imitator_rh_inspire.pth lh_base_model_checkpoint=assets/imitator_lh_inspire.pth early_stop_epochs=1000 actionsMovingAverage=0.4 experiment=cross_476@0_inspire

   Similar to single-hand training, the early_stop_epochs parameter can be adjusted based on the task complexity.

3. Test Test the bi-manual policy:

   python main/rl/train.py task=ResDexHand dexhand=inspire side=BiH headless=false num_envs=4 learning_rate=2e-4 test=true randomStateInit=false dataIndices=[476@0] rh_base_model_checkpoint=assets/imitator_rh_inspire.pth lh_base_model_checkpoint=assets/imitator_lh_inspire.pth actionsMovingAverage=0.4 checkpoint=runs/cross_476@0_inspire__xxxxxx/nn/cross_476@0_inspire.pth

---

🤗 Extending to New Hand-Object Datasets 🤗

We highly encourage researchers to transfer their hand-object datasets to dexterous hands using ManipTrans, contributing to the growth and development of the embodied AI community.

To facilitate this, we provide a straightforward interface to help you quickly adapt new datasets. Please refer to the examples in main/dataset/grab_dataset_dexhand.py and main/dataset/oakink2_dataset_dexhand_rh.py. Follow these steps:

1. Ensure your dataset's dataloader inherits from the ManipData class.
2. Implement the __getitem__ method in your dataloader. Ensure the returned dictionary includes the following keys:
   • 'data_path'
   • 'obj_id'
   • 'obj_verts'
   • 'obj_urdf_path'
   • 'obj_trajectory'
   • 'wrist_pos'
   • 'wrist_rot'
   • 'mano_joints'
3. Add the handling logic for your dataset's index in the dataset_type function of main/dataset/factory.py.

With these steps completed, you should be able to run ManipTrans on your dataset 🤗. If you encounter any issues, feel free to contact us for assistance.

---

🤗 Extending to New Dexterous Hand URDF Files 🤗

We warmly welcome contributions from hardware engineers in the industry to adapt URDF files of their self-designed dexterous hands to ManipTrans. Such contributions are critical for advancing real-world dexterous manipulation applications.

To assist with this, we provide a simple interface that allows you to quickly adapt your dexterous hand URDF file. Please refer to the configuration files located in maniptrans_envs/lib/envs/dexhands and follow the steps below to create your own configuration:

1. Create a new configuration class that inherits from DexHand.
2. Define the following essential variables in your configuration class:
   • body_names and dof_names: These should ideally follow the order from IsaacGym's definition.
   • hand2dex_mapping: This mapping is critical as it defines the correspondence between human hand keypoints and the joints of the dexterous hand.
   • contact_body_names: The names of the body parts (e.g., fingertips) that are expected to interact with objects.
   • weight_idx: A mapping of joint indices for different training weight levels:
     • level_1_joints: Critical joints, excluding the fingertips.
     • level_2_joints: Other less-critical joints.
   • bone_links: The finger connection sequence, used primarily for visualization purposes.
   • relative_rotation: The rotation of the dexterous hand’s wrist relative to the MANO wrist, which needs manual adjustment.
   • relative_translation: The translation of the dexterous hand’s wrist relative to the MANO wrist. If the wrist position in the URDF is at the origin, set relative_translation to 0.

After completing this configuration, you should be able to use ManipTrans with your custom-designed dexterous hand 🤗. If you encounter any issues, feel free to contact us for support. We also encourage you to share your URDF files with the community to help advance research and development in dexterous manipulation.

📄 Check out Our Paper

Our paper is posted on CVPR25. If you find our work useful, please consider citing us!

  @inproceedings{li2025maniptrans,
    title={ManipTrans: Efficient Dexterous Bimanual Manipulation Transfer via Residual Learning},
    author={Li, Kailin and Li, Puhao and Liu, Tengyu and Li, Yuyang and Huang, Siyuan},
    booktitle=CVPR,
    year={2025}
  }

🙏 Acknowledgement

We thank OakInk V2 for the dataloader and Transic for the training pipeline used in this work.

📜 License

This codebase is released under the GPL v3 License.