nvidia/PhysicalAI-Robotics-GraspGen

GraspGen: Scaling Sim2Real Grasping

GraspGen is a large-scale simulated grasp dataset for multiple robot embodiments and grippers.

We release over 57 million grasps, computed for a subset of 8515 objects from the Objaverse XL (LVIS) dataset. These grasps are specific to three grippers: Franka Panda, the Robotiq-2f-140 industrial gripper, and a single-contact suction gripper (30mm radius).

Dataset Format

The dataset is released in the WebDataset format. The folder structure of the dataset is as follows:

grasp_data/
    franka/shard_{0-7}.tar
    robotiq2f140/shard_{0-7}.tar
    suction/shard_{0-7}.tar
splits/
    franka/{train/valid}_scenes.json
    robotiq2f140/{train/valid}_scenes.json
    suction/{train/valid}_scenes.json

We release test-train splits along with the grasp dataset. The splits are made randomly based on object instances.

Each json file in the shard has the following data in a python dictionary. Note that num_grasps=2000 per object.

‘object’/
    ‘scale’ # This is the scale of the asset, float
‘grasps’/
    ‘object_in_gripper’ # boolean mask indicating grasp success, [num_grasps X 1]
    ‘transforms’ # Pose of the gripper in homogenous matrices, [num_grasps X 4 X 4]

The coordinate frame convention for the three grippers are provided below:

Visualizing the dataset

We have provided some minimal, standalone scripts for visualizing this dataset. See the header of the visualize_dataset.py for installation instructions.

Before running any of the visualization scripts, remember to start meshcat-server in a separate terminal:

meshcat-server

To visualize a single object from the dataset, alongside its grasps:

cd scripts/ && python visualize_dataset.py --dataset_path /path/to/dataset --object_uuid {object_uuid} --object_file /path/to/mesh --gripper_name {choose from: franka, suction, robotiq2f140}

To sequentially visualize a list of objects with its grasps:

cd scripts/ && python visualize_dataset.py --dataset_path /path/to/dataset --uuid_list {path to a splits.json file} --uuid_object_paths_file {path to a json file mapping uuid to absolute path of meshes} --gripper_name {choose from: franka, suction, robotiq2f140}

Objaverse dataset

Please download the Objaverse XL (LVIS) objects separately. See the helper script download_objaverse.py for instructions and usage. Note that running this script autogenerates a file that maps from UUID to the asset mesh path, which you can pass in as input uuid_object_paths_file to the visualize_dataset.py script.

License

License Copyright © 2025, NVIDIA Corporation & affiliates. All rights reserved.

Both the dataset and visualization code is released under a CC-BY 4.0 License.

For business inquiries, please submit the form NVIDIA Research Licensing.

Contact

Please reach out to Adithya Murali ([email protected]) and Clemens Eppner ([email protected]) for further enquiries.