DemoFunGrasp: Universal Dexterous Functional Grasping via Demonstration-Editing Reinforcement Learning

The official pytorch implementation of: DemoFunGrasp: Universal Dexterous Functional Grasping via Demonstration-Editing Reinforcement Learning (CVPR 2026)

DemoFunGrasp is a reinforcement learning framework for universal dexterous functional grasping. The learned policy generalizes to unseen combinations of objects and functional grasping conditions, and achieves zero-shot sim-to-real transfer. For the same object, the policy can produce diverse grasps by adjusting the grasping style and affordance.

Installation

Reinforcement Learning Environment Setup

It is recommended to use a conda environment to manage dependencies:

conda create -n demofungrasp python=3.8.19
conda activate demofungrasp

# For CUDA 11.8; you can change the version according to your GPU
pip install torch==2.3.0+cu118 torchvision==0.18.0+cu118 torchaudio==2.3.0 --index-url https://download.pytorch.org/whl/cu118

Install Isaac Gym and IsaacGymEnvs:

# Download IsaacGym_Preview_4_Package.tar.gz from the NVIDIA website
tar -zxvf IsaacGym_Preview_4_Package.tar.gz
cd ./isaacgym/python
pip install -e .
cd ../../

git clone https://github.com/isaac-sim/IsaacGymEnvs.git
cd IsaacGymEnvs/
pip install -e .

Install other required Python packages:

pip install -r requirements.txt

Download Object Dataset and Textures

Download the object assets and textures from here and unzip them into the directory ./assets/.

Run the Pretrained Policy

We provide checkpoints in the checkpoint directory for both ShadowHand and Inspire Hand. You can run:

bash script/train_inspire_bash.sh
# For shadow hand, run:
bash scripts/train_shadow_bash.sh

You can modify num_envs to fit your GPU memory and multiObjectList to select a subset of objects.

---

Training and Testing

Training

After modifying num_envs=5000, setting test=False, and changing +run_name to your desired experiment name, you can train your own policy by running:

bash script/train_inspire_bash.sh

Use TensorBoard to monitor training progress.

Testing the State-Based Policy

You can test the trained policy and record demonstrations by running:

bash scripts/state_based_demo.sh

Collecting Vision Dataset

You can collect RGB observation datasets in various formats. To collect the Lerobot format, run:

bash scripts/collect_vision_dataset.sh

After collecting a sufficient number of trajectories (e.g., 30k), you can use any policy to imitate them.

Using Other Object Datasets

To train and test on other object datasets, you must preprocess the .stl or .obj files to generate point clouds and additional information. Run the following scripts:

# Generate point clouds and bounding boxes
python dataset_processor/generate_object_dataset_pcl.py

# Generate object list
python dataset_processor/generate_urdfs_from_meshes.py

---

Acknowledgments

This repository is built upon:

• IsaacGymEnvs
• UniDexGrasp
• DemoGrasp
• Lerobot

---

Citation

If you find this work useful, please consider citing:

@article{mao2025universal,
  title={Universal Dexterous Functional Grasping via Demonstration-Editing Reinforcement Learning},
  author={Mao, Chuan and Yuan, Haoqi and Huang, Ziye and Xu, Chaoyi and Ma, Kai and Lu, Zongqing},
  journal={arXiv preprint arXiv:2512.13380},
  year={2025}
}