---
license: cc-by-nc-4.0
language:
- en
pretty_name: HUGS Object Assets
tags:
- robotics
- dexterous-grasping
- point-cloud
- 3d
- mesh
size_categories:
- 100K<n<1M
---

# HUGS Object Assets

This dataset provides the DGN_2k object assets used by HUGS: meshes and collision assets reprocessed with MeshProcess, scene configurations, original train/test splits, and rendered partial point clouds.

**Release status: complete.** Both archives have been uploaded and their remote SHA-256 metadata matches the locally verified files. The archive content is available at dataset revision `b6df8d4721a05a73f657afd15d935da081a3e1e5`.

These are locally reprocessed assets, not a byte-identical mirror of [JiayiChenPKU/BODex](https://huggingface.co/datasets/JiayiChenPKU/BODex). Raw mesh sources are available from that upstream dataset and are not redistributed in this release.

## Repository layout

```text
object_assets/
├── checksums.sha256
└── DGN_2k/
    ├── DGN_2k_processed_scene_cfg_valid_split.tar.gz
    └── DGN_2k_three_realsense_d435_random4096.tar.gz
```

The first archive contains `processed_data/`, `scene_cfg/`, and `valid_split/`. The second contains `vision_data/three_realsense_d435_random4096/`. **Neither archive contains an outer `DGN_2k/` directory.** Both must be extracted into the same `object_assets/DGN_2k/` destination.

The existing mesh, NPY, and JSON files, object/scene IDs, scales, splits, and directory relationships are preserved. This release does not include raw meshes, single-view-only vision directories, logs, backups, grasps, Human Prior data, or checkpoints.

## Download and extract

From a directory where the downloaded repository will be stored:

```bash
hf download MingruiYu/HUGS --repo-type dataset --local-dir HUGS
cd HUGS
(cd object_assets && sha256sum -c checksums.sha256) &&
mkdir -p object_assets/DGN_2k &&
tar -xzf object_assets/DGN_2k/DGN_2k_processed_scene_cfg_valid_split.tar.gz \
  -C object_assets/DGN_2k &&
tar -xzf object_assets/DGN_2k/DGN_2k_three_realsense_d435_random4096.tar.gz \
  -C object_assets/DGN_2k
```

The resulting data layout is:

```text
object_assets/DGN_2k/
├── processed_data/<object_id>/
│   ├── mesh/
│   ├── urdf/
│   │   ├── coacd.urdf
│   │   ├── coacd.xml
│   │   └── meshes/
│   ├── info/
│   └── complete_point_cloud.npy
├── scene_cfg/<object_id>/
│   ├── floating/*.npy
│   └── tabletop_ur10e/*.npy
├── valid_split/
│   ├── all.json
│   ├── train.json
│   └── test.json
└── vision_data/three_realsense_d435_random4096/
    ├── cam_in.npy
    └── <object_id>/<scene_type>/<scene_stem>/
        ├── partial_pc_00.npy ... partial_pc_03.npy
        └── cam_ex_00.npy ... cam_ex_03.npy
```

Only the first archive is required for workflows that consume scene configurations and complete object point clouds. Download the second archive for partial point clouds. The checksum command above assumes both archives are present.

For existing AnyScaleGrasp consumers, use `${ANYSCALEGRASP_DATA_ROOT}/object/DGN_2k` as the extraction destination instead: create that directory, then replace both `-C` arguments with that full path. Extracting into `object/` would be incorrect for these archives. No `--strip-components` option is needed.

## Contents and coverage

Sizes below are uncompressed decimal bytes.

| Component | Files | Bytes |
| --- | ---: | ---: |
| `processed_data/` | 76,688 | 2,455,769,539 |
| `scene_cfg/` | 185,088 | 260,923,296 |
| `valid_split/` | 3 | 209,144 |
| `vision_data/three_realsense_d435_random4096/` | 1,480,705 | 16,273,005,574 |
| **Total** | **1,742,484** | **18,989,907,553** |

- `processed_data/` includes 2,397 object directories, including partially processed objects. Not every directory has every output shown above.
- The original valid split contains 2,331 objects: 1,864 train and 467 test, with no overlap.
- Scene configurations exist for 2,323 objects. Eight objects in the split have no scene configurations (six train, two test). Splits and incomplete directories are preserved, not silently filtered or repaired.
- There are 185,088 scenes: 27,876 floating and 157,212 tabletop.
- Each scene has four partial point clouds and four extrinsic files, totaling 740,352 clouds and 740,352 extrinsic files, plus the shared `cam_in.npy`.

## Data conventions

| Item | Convention |
| --- | --- |
| Scale anchors | `0.020, 0.028, 0.042, 0.060, 0.081, 0.106, 0.133, 0.162, 0.194, 0.227, 0.263, 0.300` |
| Scale source | Read the actual `scale` array inside the scene NPY dictionary. A filename containing `scale008` can represent `0.081`. |
| Scene pose | Translation in meters, quaternion order WXYZ. Mesh/XML/URDF/info paths are resolved relative to the scene file's parent directory. |
| Complete point clouds | Sampled files have shape `(4096, 3)`, float16, in normalized object coordinates. Apply the scene scale and pose for world coordinates. |
| Partial point clouds | Float16 XYZ points in scene world coordinates. 4096 is an upper bound; some clouds contain fewer points. |
| Three-view directory | Tabletop samples fuse three camera views; floating samples remain single-view. Four samples are stored per scene. |
| Extrinsics | Sampled floating matrices have shape `(4, 4)` and tabletop matrices `(3, 4, 4)`, float32. MeshProcess uses row-vector camera-to-world transforms. |
| `cam_in.npy` | A 4×4 row-vector OpenGL projection matrix, not a conventional 3×3 intrinsic matrix. |
| Scene NPY | Serialized dictionaries loaded with `numpy.load(path, allow_pickle=True).item()`. Load pickle-containing files only from a trusted release. |

The checked-in renderer configuration uses 1280×720 images, `fx=910.944763183594`, `fy=910.721984863281`, `cx=642.13818359375`, `cy=367.6796875`, near plane 0.1 and far plane 10.0. These values are consistent with the released `cam_in.npy`; the complete historical generation environment has not been recovered.

## Provenance and limitations

Local processing logs dated May 13, 2026 record full MeshProcess processing; the selected rendering log is dated May 20, 2026. A previous comparison of eight simplified meshes against a local BODex backup found different SHA-256 hashes. The upstream dataset link therefore documents provenance, not a substitute for these exact reprocessed assets.

The MeshProcess checkout inspected for format interpretation was `022b41f2ebb32b3de3ea78b2723f180eaef2118e`. This is an inspection reference, **not a confirmed historical generating commit**. Historical seeds and the complete producer environment remain unconfirmed. No full numerical quality audit or simulation benchmark is claimed.

## Archive verification

| Archive | Compressed bytes | SHA-256 |
| --- | ---: | --- |
| `DGN_2k_processed_scene_cfg_valid_split.tar.gz` | 872,815,299 | `32204e6557ca2c7f35b01f05633341b4d226f475502b19c6742a98d63d66d63e` |
| `DGN_2k_three_realsense_d435_random4096.tar.gz` | 14,386,467,307 | `bf658b10af71ceeb0d32d43126940da9415e607917e47b22eacf5d0fb00d5db5` |

Both archives were compared against all source files with GNU tar, and their member paths, file counts, and uncompressed sizes were checked. Selected files were extracted into a fresh DGN_2k directory; split sets, two scene configurations, relative mesh/XML/URDF/info references, complete/partial point clouds, and camera matrices passed a loading smoke check. This does not constitute a full numerical quality audit or a physics/simulation benchmark.

## Attribution and license

Source dataset: [JiayiChenPKU/BODex](https://huggingface.co/datasets/JiayiChenPKU/BODex), inspected revision `7f25e59b111ef3405aea242199ec3e07f2965c13`. Its dataset card declares **CC BY-NC 4.0**. This derived release retains that noncommercial license declaration and upstream attribution. Original mesh-source terms continue to apply where applicable; see [THIRD_PARTY_NOTICES.md](THIRD_PARTY_NOTICES.md).

Please cite the upstream work when using these assets:

> BODex: Scalable and Efficient Robotic Dexterous Grasp Synthesis Using Bilevel Optimization. ICRA 2025. arXiv:2412.16490.

Upstream [project](https://pku-epic.github.io/BODex/) and [MeshProcess code](https://github.com/JYChen18/MeshProcess).