TROV2Dataset

TROV2Dataset is a collection of object-level Zarr datasets for robotic grasping research. The training data contains 25,228 object Zarr directories packed into 101 uncompressed tar shards (object_zarrs_00000.tar through object_zarrs_00100.tar). The tar files total 6,806,567,178,240 bytes (6.19 TiB). All 101 shard paths are present in this repository. Validation scenes, object point clouds, and object meshes are available in three additional archives.

The Zarr directories were selected from a locally prepared TRO-1B object-Zarr collection. The selection excludes the union of objects in the hand-specific evaluation lists: 5,938 unique evaluation objects were excluded from 31,166 source objects. The 200-object challenge-validation union was contained in that excluded set. The Zarr training subset remains evaluation-excluded; the validation scenes are distributed separately below.

Repository layout

Path	Description
tar_shards/object_zarrs_*.tar	The 101 archives. Each contains top-level <object_name>.zarr/ directories.
manifest.tsv	One row per shard: index, number of objects, archive path, object-list path, and archive size in bytes.
shard_lists/objects_shard_*	Exact Zarr directory names assigned to each shard, one name per line.
object_zarrs.txt	All 25,228 Zarr directory names, one per line.
sha256sums.txt	SHA-256 checksums for all tar archives, with paths relative to the dataset root.
tar_shards/TRO-val-dataset.tar.zst	Validation scenes, with the top-level TRO-val-dataset/ directory inside.
tar_shards/object_pc.tar.zst	Object point clouds, with the top-level object_pc/ directory inside.
tar_shards/object_urdf.tar.zst	Object geometry files, with the top-level object_urdf/ directory inside.
supplementary_sha256sums.txt	SHA-256 checksums for the three supplementary archives.

Shards 00000–00099 each contain 250 object directories; shard 00100 contains 228. The archives are tar files, not individual Hub dataset rows, so the Dataset Viewer does not provide a row-level preview.

Download and extract

Install the Hugging Face CLI, then download the index files and one example shard:

hf download Barrybarry-Smith/TROV2Dataset \
  manifest.tsv object_zarrs.txt sha256sums.txt \
  --repo-type dataset --local-dir ./TROV2Dataset

hf download Barrybarry-Smith/TROV2Dataset \
  tar_shards/object_zarrs_00000.tar \
  --repo-type dataset --local-dir ./TROV2Dataset

cd TROV2Dataset
sha256sum --check --ignore-missing sha256sums.txt
mkdir -p object_zarrs
tar -xf tar_shards/object_zarrs_00000.tar -C object_zarrs

To locate a particular object, find its directory name in object_zarrs.txt, then search shard_lists/ for that exact name. The suffix of the matching list file gives the tar-shard index. For example:

grep -lFx 'YOUR_OBJECT_NAME.zarr' shard_lists/objects_shard_*

To download the full release, use hf download with --include 'tar_shards/*.tar' and a local directory with enough space for roughly 6.19 TiB of archives, plus any extracted copies. Run sha256sum --check sha256sums.txt from the downloaded dataset root after all shards are present.

Validation scenes and geometry

The supplementary archives contain 8,200 validation scene .pt files across hand-specific folders, 53,675 object point cloud .pt files, and 107,498 object geometry files. Download and verify them separately from the Zarr shards:

hf download Barrybarry-Smith/TROV2Dataset \
  tar_shards/TRO-val-dataset.tar.zst \
  tar_shards/object_pc.tar.zst \
  tar_shards/object_urdf.tar.zst \
  supplementary_sha256sums.txt \
  --repo-type dataset --local-dir ./TROV2Dataset

cd TROV2Dataset
sha256sum --check supplementary_sha256sums.txt
tar -I zstd -xf tar_shards/TRO-val-dataset.tar.zst
tar -I zstd -xf tar_shards/object_pc.tar.zst
tar -I zstd -xf tar_shards/object_urdf.tar.zst

Each archive preserves the original top-level directory name, so extraction recreates TRO-val-dataset/, object_pc/, and object_urdf/. The TROGraspV2 validation and visualization scripts use these directories; update their local data-root paths after downloading.

Data format

Each archive contains Zarr v2 directories. An object directory has root attributes such as object_name, dataset_names, hand_names, and total_samples. The sampled object inspected for this card has data/pc, data/joint, and data/joint_mask arrays, plus lookup and sample-metadata groups. Array shapes and available source names can vary by object; inspect each directory's .zattrs and .zarray metadata before using it.

For example, after extraction:

import zarr

group = zarr.open_group("object_zarrs/YOUR_OBJECT_NAME.zarr", mode="r")
print(dict(group.attrs))
print(group["data/pc"].shape, group["data/joint"].shape)

Source and terms of use

The packaged Zarr directories come from the locally prepared TRO-1B object-Zarr collection. Object metadata records contributing dataset names; an inspected object lists unidexgrasp and objaversexl. This card does not establish that those names are exhaustive for the full release.

No license or redistribution terms have been declared for this repository. Check the terms of the contributing sources before reusing or redistributing their content. A public download link does not itself grant a license.