Public repository for the Lightning Grasp system. paper
Thousands of dexterous grasps in seconds.
Apr 23, 2026: We added mimic joint support.
- Operating System: Ubuntu 22.04 or 24.04
- GPU: NVIDIA GPU with Pascal, Volta/Turing, Ampere, or Ada Lovelace architecture
- CUDA 12
Follow these steps to set up the environment and dependencies.
Choose one of the following setup methods.
We provide pre-configured Conda environments for different Python/CUDA combinations. For most users with modern NVIDIA GPUs (RTX 20-series and newer), we recommend:
conda env create -f conda_env/conda_py39.yml
conda activate lygra
Available environments:
- conda_py39.yml: Python 3.9 + PyTorch 2.8.0 (recommended for newer GPUs)
- conda_py38.yml: Python 3.8 + PyTorch 2.4.1 (for legacy systems)
Not sure if you are tired of setting up a new heavy-weight environment for every new thing.
If you prefer a minimal setup and already have Python 3.8/3.9 and pytorch installed:
pip install open3d==0.19.0 urdfpy==0.0.22 trimesh==4.9.0
pip install networkx==3.2.1
Done. (Do not merge these pip install into one line)
This release includes pre-compiled CUDA kernel binaries. The CUDA C++ source code will be published in a future release.
Supported Python Version: 3.8, 3.9, 3.10. See the release page for details. For other Python versions, please open an issue and I'll be happy to add support.
Setup Steps:
- Download the appropriate compiled binaries for your Python version from the Releases page
- Extract the downloaded files into the
lygra/cpp/build/directory - Verify the file structure matches:
lygra/cpp/build/lbvh/lbvh.so
lygra/cpp/build/geometry/geometry.so
Download hand and object assts from the Releases page, and put them under the ./assets folder. Verify the file structure matches:
./assets/hand/...
./assets/object/...
Run a single forward pass demo with visualization:
python demo.py --visualize
This will visualize results on Allegro + YCB Apple setup.
To use other hands and objects, run:
python demo.py --visualize --robot shadow --object_mesh_path ./assets/object/ycb/042_adjustable_wrench/textured.obj
Ensure your object mesh is scaled to meter units before processing. ;)
The supported hands are shown in lygrasp/robot/__init__.py. We provide several examples in the repository. Please refer to the next section to setup your own hands.
The system includes several tunable parameters to optimize performance for your hardware and use case:
--n_contact: Controls the number of active contacts to search during grasp optimization--batch_size_outer&--batch_size_inner: Adjust these values to maximize GPU utilization- For GPUs with large memory (≥ 12GB), increase both batch sizes for better performance
- Start with default values and scale up until memory limits are reached
- Some typical setups: (128, 256), (192, 256), (256, 256), (256, 512).
There are several examples in lygra/robot/ folder. You can refer to lygra/robot/allegro.py for an tutorial. Basically, you simply need to setup a config object that specifies the contact field rules (i.e. which patches to use defined by allowed normals), canonical object space (i.e. where to initialize the object), and some URDF metadata. That's it!
Setting Up Collision
You may need to use get_white_list_pairs() to disable specific self-collision checks within the hand, please refer to lygra/robot/base.py for more information. If Lightning Grasp fails to synthesize grasps on a new model, it is likely that the collision filtering is not configured correctly.
Mimic Joints and Joint Order
You can define mimic joints in your URDF using the <mimic> tag. Lightning grasp will automatically detect these joints. Ensure that get_active_joints() in the corresponding python configuration class includes mimic joints. The generated joint positions q follow the ordering defined in get_active_joints(), but the returned values exclude mimic joints. For example, if the active joints are [j0, j1, j2, j3] and j1 and j3 are mimic joints, then Lightning Grasp will return q ordered as [j0, j2].
However, we observe that synthesis speed for hands with mimic joints is currently slow (~10–200 SPS on A100/RTX 4090). This feature is still experimental and may require additional tuning.
Known Limitations
This released version has the following limitations.
- Do not use huge objects for now. I am working on a version that clamps the object mesh and resample the point cloud after placement.
Known Issues
We observe that the current implementation does not properly release GPU memory with torch==2.11.0. You will see CUDA OOM error if you run the main loop repeatedly. Downgrading to torch==2.7.0 resolves this issue.
Comments
For the grasp synthesis with grippers, consider using specialized approaches for higher efficiency (i.e. directly sample antipodal points on a ray.).
- If you encounter "Package X not found" errors, simply run
pip install X. - If you find any error message containing
networkx, just run
pip install networkx==3.2.1
For other problems, leave an issue or email me zhaohengyin@cs.berkeley.edu (I can be quite busy, but I’ll do my best to respond as soon as I can).
CC-By-NC 4.0.
This software and its generated data are licensed for academic and research use only. The license does not grant rights for commercial application, including but not limited to:
- Integration into commercial systems
- Commercial services using this software/data
A separate commercial license is required for any business use. Please contact the author for licensing terms.
If this work helps your research, a citation would be greatly appreciated!
@article{yin2025lightninggrasp,
title = {Lightning Grasp: High Performance Procedural Grasp Synthesis with Contact Fields},
author = {Yin, Zhao-Heng and Abbeel, Pieter},
journal = {arXiv preprint arXiv:2511.07418},
year = {2025},
url = {https://arxiv.org/abs/2511.07418}
}