HUGS-Human

Human grasp data with hand poses, object geometry, and an interactive viewer.

Highlights | Quick start | Data format | License and notices

HUGS-Human provides reconstructed human grasps for the HUGS Human Prior pipeline. Download formatted records for learning, or add the source meshes to inspect hands and objects together in the browser.

Highlights

• 1,849 grasps across 304 objects, with single-hand and two-hand interactions.
• Five grasp types, covering two-finger, three-finger, and full-hand grasps.
• Object-disjoint splits: 243 training objects and 61 test objects.
• Interactive visualization with MANO hands, textured objects, same-object grids, and random batches.

Downloads

Package	Contents	Size
Formatted records	1,849 grasps, canonical point clouds, scene configs, splits, and metadata	178 MB
Source inputs	1,851 reconstructed scenes, textured object meshes, fitted poses, hand detections, and calibration	6.93 GB

Formatted records are sufficient for workflows using poses and point clouds. Visualization requires both packages and separately obtained MANO models. Each package includes a SHA-256 checksum and a manifest with its file inventory and provenance.

Quick start

Download the repository and extract both packages:

python -m pip install --upgrade huggingface_hub
hf download MingruiYu/HUGS-Human --repo-type dataset --local-dir HUGS-Human
cd HUGS-Human

(cd source/v1 && sha256sum -c checksums.sha256)
(cd formatted/v1 && sha256sum -c checksums.sha256)

export HUGS_DATA_ROOT="$PWD/data"
mkdir -p "$HUGS_DATA_ROOT/OurHumanGrasp" "$HUGS_DATA_ROOT/OurHumanGraspFormat"
tar -xzf source/v1/v1_source_processing_inputs.tar.gz -C "$HUGS_DATA_ROOT/OurHumanGrasp"
tar -xzf formatted/v1/v1_training_core.tar.gz -C "$HUGS_DATA_ROOT/OurHumanGraspFormat"

Visualize

Use a Python 3.10 environment for the legacy MANO dependencies:

python -m pip install --upgrade pip wheel 'setuptools<81' 'numpy>=1.23.5,<1.24'
python -m pip install --no-build-isolation -r requirements-visualize.txt

Obtain the licensed models from MANO, then point --mano-root to the directory containing MANO_LEFT.pkl and MANO_RIGHT.pkl:

python visualize.py \
  --formatted-root "$HUGS_DATA_ROOT/OurHumanGraspFormat" \
  --source-root "$HUGS_DATA_ROOT/OurHumanGrasp" \
  --mano-root /path/to/mano/models \
  --host 127.0.0.1 --port 9067

Open http://127.0.0.1:9067. The viewer runs on CPU and supports single grasps, paginated same-object grids grouped by pose_index, and random batches of up to 12 grasps. Toggle hands, objects, the table, labels, and contact markers independently. Left hands are blue; right hands are coral.

For a remote server, forward the port with ssh -N -L 9067:127.0.0.1:9067 user@server. Use python visualize.py --help for additional options.

Data format

The extracted formatted directory contains:

OurHumanGraspFormat/
  grasp/<object_id>/seq_<index>.npy
  object/
    vision_data/
    scene_cfg/
    valid_split/
    canonical_registration.json
  metadata.csv

Grasp files are NumPy dictionaries containing object pose, relative scale, source identifiers, canonical registration, and active-hand MANO parameters. Translations are in meters; rotations are axis-angle. Object point clouds contain 4,096 points in the canonical object frame.

import numpy as np

grasp = np.load("data/OurHumanGraspFormat/grasp/obj_443/seq_0.npy", allow_pickle=True).item()
object_info = grasp["object"]
hand_info = grasp["hand"]

Load only trusted records and MANO models: both use Python pickle. The archives are file-based datasets, not Hugging Face load_dataset tables.

Notes

• Source inputs are reconstructed processing assets; raw RGB/depth captures and preview images are excluded.
• Two source candidates, obj_436_seq_1 and obj_500_seq_0, have no formatted grasp record.
• Formatted canonical meshes are not included. The viewer uses source meshes with the stored registration; source and canonical meshes have different coordinate frames.
• The stored hand scale converts MANO output to meters. Contact markers represent finger annotations, not measured mesh intersections.
• Source reprocessing is not guaranteed to reproduce the formatted files. See the package manifests for provenance and exclusions.

License and notices

No dataset license is currently specified; publication does not grant commercial-use or redistribution rights. Review THIRD_PARTY_NOTICES.md and obtain any permissions required for your use. MANO models are not distributed here and remain subject to their own terms.