1M-HUGs

Grasp dataset with three splits.

split	location	samples	format
train	train/*.tar	1,279,142	tar shards of <name>.pkl
val	val/<scene>/<object>/*.pkl	600	raw tree + sim_assets/
test	test/<scene>/<object>/*.pkl	300	raw tree + sim_assets/

train is split into 256 plain tar shards (5,000 pkls each) purely so it transfers as ~256 files instead of 1.28M tiny ones. Each shard is an ordinary tarball — extract it with stdlib tarfile or tar xf, no special library. val/test are small and ship as raw directory trees (they also carry a sim_assets/ dir per object).

Quick start

Download, then untar train into a flat grasp_data/ directory:

huggingface-cli download kevinywu/1m-hugs --repo-type dataset --local-dir 1m-hugs
cd 1m-hugs && mkdir -p grasp_data
for t in train/*.tar; do tar -xf "$t" -C grasp_data/; done   # 1.28M pkls, ~70 GB

val/test download already unpacked — use them directly.

Prefer not to untar? The shards are plain tarballs, so you can read a pkl straight out of one with stdlib tarfile:

import tarfile, pickle
with tarfile.open("train/00000.tar") as tar:
    for m in tar:
        sample = pickle.load(tar.extractfile(m))  # dict with image/depth/mask bytes

Sample schema

Each .pkl unpickles to a dict:

key	type	train	val/test	notes
object_name	str	✓	✓	object identity
frame_index	int	✓	✓	frame number
grasp_index	int	✓	✓	grasp id
camera	dict	✓	✓	camera intrinsics/extrinsics
camera_original	dict	✓	✓	pre-crop camera params
grasp	dict	✓	None	grasp label (withheld in val/test)
image	bytes	✓	✓	encoded RGB image
depth	bytes	✓	✓	encoded depth
object_mask	bytes	✓	✓	encoded segmentation mask
T_world_camera	array		✓	world→camera transform
condition_point	array		✓	2D query pixel [u, v]

train carries a populated grasp dict (MANO-style: pose, pose_6d, shape, landmarks_3d/2d, T_camera_wrist, mesh_vertices, t). In val/test grasp is None (the label to predict) and each object additionally ships a sim_assets/ dir (object.urdf, object.xml, visual.obj, collision/, properties.json).