Official code for DemoGrasp: Universal Dexterous Grasping from a Single Demonstration (ICLR 2026).

DemoGrasp is a framework for learning universal dexterous grasping policies via reinforcement learning (RL) augmented with a single demonstration. It achieves state-of-the-art performance across diverse robotic hand embodiments and transfers effectively to real robots, demonstrating strong generalization.

Requirements

Create a conda environment with python 3.8.19, then install packages:

conda create -n demograsp python=3.8.19
conda activate demograsp
pip install torch==2.3.0+cu121 torchvision==0.18.0+cu121 torchaudio==2.3.0 --index-url https://download.pytorch.org/whl/cu121
pip install numpy==1.23.5 sapien==3.0.0b0 chumpy==0.70 addict==2.4.0 yapf==0.43.0 h5py sorcery==0.2.2 pynvml==11.5.0 ipdb==0.13.13 matplotlib==3.5.1 networkx==3.1 tqdm==4.66.4 datasets==3.1.0 einops==0.4.1 jsonlines

Install IsaacGym Preview 4 and IsaacGymEnvs.
Download assets and checkpoints from this link.
- Unzip robots.zip, union_ycb_unidex.zip, and textures.zip into the assets/ folder.
- Unzip ckpt.zip into the project root.
Your project structure should look like:

DemoGrasp/
├── assets/
│   ├── textures/
│   ├── union_ycb_unidex/
│   ├── franka/
│   ├── inspire_tac/
│   ├── ...
├── ckpt/
│   ├── inspire.pt
│   ├── ...

QuickStart

Try our object-centric demonstration replay:

python run_rl_grasp.py task=grasp num_envs=16 task.env.asset.multiObjectList="union_ycb_unidex/example.yaml" +debug=test_demo_replay

You can see the Inspire hand grasping different objects with the same position, orientation, and hand pose.

Play the trained RL policies

For Inspire hand, run:

python run_rl_grasp.py \
    task=grasp \
    train=PPOOneStep \
    test=True \
    num_envs=175 \
    task.env.observationType="eefpose+objinitpose+objpcl" \
    task.env.armController=pose \
    task.env.asset.multiObjectList="union_ycb_unidex/union_ycb_debugset.yaml" \
    task.env.randomizeTrackingReference=True \
    task.env.randomizeGraspPose=True \
    task.env.trackingReferenceFile=tasks/grasp_ref_inspire.pkl \
    task.env.episodeLength=50 \
    task.env.enablePointCloud=True \
    train.params.is_vision=True \
    checkpoint='ckpt/inspire.pt'

Use headless=True on a headless machine. To evaluate on unseen object categories in DexGraspNet, set task.env.asset.multiObjectList="union_ycb_unidex/test_set_unseen_cat.yaml".
Scripts for all hands are listed in play_policy.sh

Training

For Inspire hand, run:

python -u run_rl_grasp.py \
    task=grasp \
    train=PPOOneStep \
    hand=fr3_inspire_tac \
    num_envs=7000 \
    task.env.armController=pose \
    task.env.trackingReferenceFile=tasks/grasp_ref_inspire.pkl \
    task.env.trackingReferenceLiftTimestep=13 \
    task.env.asset.multiObjectList="union_ycb_unidex/union_ycb_debugset.yaml" \
    task.env.randomizeTrackingReference=True \
    task.env.randomizeGraspPose=True \
    task.env.resetDofPosRandomInterval=0.2 \
    task.env.observationType="eefpose+objinitpose+objpcl" \
    task.env.episodeLength=40 \
    task.env.enablePointCloud=True \
    train.params.is_vision=True \
    task.env.enableRobotTableCollision=False

Checkpoints and tensorboard logs are saved in runs_ppo/.

Use headless=True on a headless machine. To train on DexGraspNet benchmark (3200 objects), set task.env.asset.multiObjectList="union_ycb_unidex/train_set.yaml" headless=True num_envs=6400.
Scripts for all hands are listed in train.sh. Policies are expected to achieve high success rates after 2000 training iterations.

Collect robot dataset with image observations

A headed machine with at least 20GB of GPU memory is required.
For Inspire hand, run:

python run_rl_grasp.py \
    task=grasp \
    train=PPOOneStep \
    task.env.asset.multiObjectList="union_ycb_unidex/union_ycb_debugset.yaml" \
    task.env.observationType="eefpose+objinitpose+objpcl" \
    task.env.armController=pose \
    task.env.randomizeTrackingReference=True \
    task.env.randomizeGraspPose=True \
    task.env.trackingReferenceFile=tasks/grasp_ref_inspire.pkl \
    task.env.enablePointCloud=True \
    train.params.is_vision=True \
    task.env.render.enable=True \
    task.env.render.camera_ids="[1,2]" \
    task.env.render.data_type="rgb" \
    task.env.render.randomize=True \
    task.env.render.randomization_params.camera_pos="[-0.02,0.02]" \
    task.env.render.randomization_params.camera_quat="[-0.02,0.02]" \
    task.env.render.randomization_params.table_xyz="[0.05,0.02,0]" \
    task.env.episodeLength=40 \
    checkpoint='ckpt/inspire.pt' \
    +debug=collect_real_dataset \
    num_envs=175 \
    +num_episodes=30000

Set +num_episodes to the total number of successful trajectories to collect. The dataset will be saved in data/datasets/$robot_name_$datetime in LeRobot v2.0 format.

You can use any imitation learning method to train an RGB-based policy on this dataset. To achieve sim-to-real deployment on your robot, you should revise the robot configurations, camera intrinsics and extrinsics, and table sizes to match your hardware settings.

Acknowledgments

Our codebase builds upon the following projects:

Citation

If you find our work useful, please consider citing us!

@article{yuan2025demograsp,
  title={DemoGrasp: Universal Dexterous Grasping from a Single Demonstration},
  author={Yuan, Haoqi and Huang, Ziye and Wang, Ye and Mao, Chuan and Xu, Chaoyi and Lu, Zongqing},
  journal={arXiv preprint arXiv:2509.22149},
  year={2025}
}