---
license: cc-by-nc-4.0
pretty_name: GraspXL
size_categories:
- 10M<n<100M
tags:
- grasping
- dexterous-hand
- motion-generation
- objaverse
- physics-simulation
- robotics
- 3d
---

# GraspXL

GraspXL is a large-scale dataset of physically plausible dexterous grasping
motion sequences over Objaverse objects. It contains generated grasping motions
for different hand models, together with processed object meshes used by the
recorded sequences.

The main archives cover diverse hand-object interactions where hands approach
objects from different directions. The tabletop archives provide a complementary
setting where objects are placed on a table and hands approach from above, which
is useful for applications focused on tabletop grasping scenes.

## Resources

- [Paper](https://arxiv.org/abs/2403.19649)
- [Project page](https://eth-ait.github.io/graspxl/)
- [Code](https://github.com/zdchan/GraspXL)
- [Visualizer](https://github.com/zdchan/GraspXL_visualization)

## Highlights

- 10M+ diverse grasping motions for 500k+ objects and multiple dexterous hand
  models.
- Motions generated with physics simulation to improve physical plausibility.
- Frame-level object and hand poses for each motion sequence.
- Both diverse-approach and tabletop grasping settings.

## Potential Applications

- Generating large-scale pseudo 3D RGBD grasping motions with texture generation methods, which can support downstream applications such as
  training pose estimation or mesh reconstruction models (e.g., [HGGT](https://lym29.github.io/HGGT/)).
- Simulating general human hand motions for human-robot interaction.
- Serving as expert demonstrations for imitation learning and generative models (e.g., [PAM](https://gasaiyu.github.io/PAM.github.io/)).
- Add precise finger motions for full-body grasping motion generation.
- Achieve zero-shot text-to-motion generation with off-the-shelf text-to-mesh
  generation methods.

## Archive Overview

The dataset is stored as `.zip` archives. To make downloading easier, recorded
motion sequences are split across several archives. For many objects, each
sequence archive contains several motion sequences, so users can choose which
archives to download.

| Archive | Content |
| --- | --- |
| `object_dataset.zip` | Processed object meshes used by the recorded sequences. |
| `objaverse_urdf.zip` | URDF files of the Objaverse objects used by the simulator. |
| `allegro_dataset_1.zip`, `allegro_dataset_2.zip` | Allegro Hand motions with diverse approach directions. |
| `mano_dataset_1.zip.part00`, `mano_dataset_1.zip.part01` | Split parts of `mano_dataset_1.zip`, containing MANO motions with diverse approach directions. |
| `mano_dataset_2.zip.part00`, `mano_dataset_2.zip.part01` | Split parts of `mano_dataset_2.zip`, containing additional MANO motions with diverse approach directions. |
| `mano_dataset_3.zip.part00`, `mano_dataset_3.zip.part01` | Split parts of `mano_dataset_3.zip`, containing additional MANO motions with diverse approach directions. |
| `leap_dataset_1.zip` | LEAP hand motions for a subset of the objects. |
| `allegro_tabletop.zip` | Allegro Hand tabletop grasping motions. |
| `mano_tabletop.zip` | MANO tabletop grasping motions. |
| `sharpa_tabletop.zip` | Sharpa Wave Hand tabletop grasping motions. |

## Object Data

`object_dataset.zip` contains processed object meshes. The meshes are scaled and
decimated before use.

```text
object_dataset.zip
|-- small
|   |-- <object_id>
|   |   `-- <object_id>.obj
|   `-- ...
|-- medium
|   |-- <object_id>
|   |   `-- <object_id>.obj
|   `-- ...
`-- large
    |-- <object_id>
    |   `-- <object_id>.obj
    `-- ...
```

`small`, `medium`, and `large` contain object meshes with different scales used
by the recorded sequences. Please check the paper for more details.

`objaverse_urdf.zip` contains the corresponding URDF files for the Objaverse
objects. These URDF files are used by the simulation code to generate motions.

## Motion Archives

All motion archives use the same high-level layout:

```text
<motion_archive>.zip
|-- small
|   |-- <object_id>
|   |   |-- <hand_prefix>_0.npy
|   |   |-- <hand_prefix>_1.npy
|   |   `-- ...
|   `-- ...
|-- medium
|   |-- <object_id>
|   |   |-- <hand_prefix>_0.npy
|   |   `-- ...
|   `-- ...
`-- large
    |-- <object_id>
    |   |-- <hand_prefix>_0.npy
    |   `-- ...
    `-- ...
```

Not every object has the same number of recorded sequences.

### Diverse-Approach Motions

These archives contain motions where hands approach objects from diverse
directions to generate broad hand-object interaction coverage.

| Archive | Hand model | File prefix | Hand pose shape |
| --- | --- | --- | --- |
| `allegro_dataset_1.zip`, `allegro_dataset_2.zip` | Allegro | `allegro_*.npy` | `(frame_num, 22)` |
| `mano_dataset_1.zip`, `mano_dataset_2.zip`, `mano_dataset_3.zip` | MANO | `mano_*.npy` | `(frame_num, 45)` |
| `leap_dataset_1.zip` | LEAP | `leap_*.npy` | `(frame_num, 22)` |

The MANO archives are stored as split files to keep each repository file under
the hosting file-size limit. Reconstruct them before extracting:

```bash
cat mano_dataset_1.zip.part00 mano_dataset_1.zip.part01 > mano_dataset_1.zip
cat mano_dataset_2.zip.part00 mano_dataset_2.zip.part01 > mano_dataset_2.zip
cat mano_dataset_3.zip.part00 mano_dataset_3.zip.part01 > mano_dataset_3.zip
```

### Tabletop Motions

These archives contain motions in a tabletop setting. Objects are placed on a
table and the hand approaches from above. The tabletop setting uses part of the
method from RobustDexGrasp and is intended for applications that need tabletop
grasping data.

| Archive | Hand model | File prefix | Hand pose shape |
| --- | --- | --- | --- |
| `allegro_tabletop.zip` | Allegro | `allegro_*.npy` | `(frame_num, 22)` |
| `mano_tabletop.zip` | MANO | `mano_*.npy` | `(frame_num, 45)` |
| `sharpa_tabletop.zip` | Sharpa Wave | `sharpa_*.npy` | `(frame_num, 28)` |

## Sequence Format


### Common Fields

Each `.npy` file contains a single motion sequence:

```python
data = np.load("<sequence>.npy", allow_pickle=True).item()
```

The sequence dictionary contains a `right_hand` entry and one object entry keyed
by `<object_id>`.

- `data["right_hand"]["trans"]`: NumPy array of shape `(frame_num, 3)`. Wrist
  position sequence.
- `data["right_hand"]["rot"]`: NumPy array of shape `(frame_num, 3)`. Wrist
  orientation sequence in axis-angle representation.
- `data["right_hand"]["pose"]`: Hand-model-specific pose sequence. See the
  archive tables above for the pose dimensionality.
- `data[object_id]["trans"]`: NumPy array of shape `(frame_num, 3)`. Object
  position sequence.
- `data[object_id]["rot"]`: NumPy array of shape `(frame_num, 3)`. Object
  orientation sequence in axis-angle representation.


### Pose Conventions

- Diverse-approach archives also include `data[object_id]["angle"]`, which is
a reserved field and is not used.
- The actual wrist pose is always represented by
  `data["right_hand"]["trans"]` and `data["right_hand"]["rot"]`. For non-MANO robotic dexterous hands, the first 6 dimensions of `data["right_hand"]["pose"]` are virtual wrist joint values not used for the actual wrist pose. The remaining dimensions are the hand joint angles.
- For MANO, `data["right_hand"]["rot"]` and `data["right_hand"]["pose"]`
  concatenate to the original MANO pose parameter. The diverse-approach MANO
  archives use `flat_hand_mean=False`, while the tabletop MANO archive uses
  `flat_hand_mean=True` with an additional `wrist_bias` applied to `data["right_hand"]["trans"]`.

## Data Loading

For more detailed data loading and visualization examples, please refer to the
[visualizer repository](https://github.com/zdchan/GraspXL_visualization).

```python
import numpy as np

data = np.load("allegro_0.npy", allow_pickle=True).item()
object_id = next(key for key in data.keys() if key != "right_hand")

right_hand_trans = data["right_hand"]["trans"]
right_hand_rot = data["right_hand"]["rot"]
right_hand_pose = data["right_hand"]["pose"]

object_trans = data[object_id]["trans"]
object_rot = data[object_id]["rot"]
```

## Citation

If you use GraspXL, please cite:

```bibtex
@inProceedings{zhang2024graspxl,
  title={{GraspXL}: Generating Grasping Motions for Diverse Objects at Scale},
  author={Zhang, Hui and Christen, Sammy and Fan, Zicong and Hilliges, Otmar and Song, Jie},
  booktitle={European Conference on Computer Vision (ECCV)},
  year={2024}
}
```

The tabletop setting uses part of the method from RobustDexGrasp. If you find
this setting useful, please consider citing:

```bibtex
@inproceedings{zhang2025RobustDexGrasp,
  title={{RobustDexGrasp}: Robust Dexterous Grasping of General Objects},
  author={Zhang, Hui and Wu, Zijian and Huang, Linyi and Christen, Sammy and Song, Jie},
  booktitle={Conference on Robot Learning (CoRL)},
  year={2025}
}
```

## License

This dataset is released under the
[Creative Commons Attribution-NonCommercial 4.0 International License](https://creativecommons.org/licenses/by-nc/4.0/).