OmniDex

OmniDex: Scaling Dexterous Hand Grasping to Diverse Cluttered Scenes.

Upload in progress. This public repository is receiving the full STD data and its object/support assets. Files become available progressively. The release is complete only when std/1.0.0/release.json reports uploaded_verified_public. File counts and total sizes are being measured; this page does not claim a complete release yet.

Layout

assets/r1/
  visual/{train,test}/
  collision/{train,test}/
  supports/
std/1.0.0/
  train/{observations,annotations}/
  test/{observations,annotations}/
  grasp_seeds/{train,test}/
  manifests/
  splits/
  README.md
  croissant.json

Large components are packaged into numbered archives, generally around 5 GiB of source files per shard. Observations use .tar; other components use .tar.zst. A scene or an object directory is kept within one archive. Each archive has a per-file size and SHA256 sidecar. Original annotation values and component-relative paths are preserved.

Mapping to the original data

Original directory	Published component
train/output, test/output	std/1.0.0/{train,test}/observations
train/scene_best, test/scene_best	std/1.0.0/{train,test}/annotations
train/raw_lift, test/raw_lift	std/1.0.0/grasp_seeds/{train,test}
train/norm_ds, test/norm_ds	assets/r1/visual/{train,test}
train/raw, test/raw	assets/r1/collision/{train,test}
table_best	assets/r1/supports

All original annotation groups are retained. Canonical scenes use exact split/group/scene identities. Extra annotation groups are indexed separately rather than silently merged into train/test observations.

Download after the release completes

Install Python huggingface_hub and the zstd command, then download the helper:

python download_dataset.py --dataset std --release 1.0.0 --with-assets --restore-original-layout --output ./OmniDex

To use HF Mirror, add --endpoint https://hf-mirror.com. The helper checks archive SHA256 hashes and restores OmniDex/data/train/output, scene_best, norm_ds, raw, raw_lift and test equivalents. Stored absolute paths in original annotation files need path remapping; the helper includes resolve_original_path for this purpose. It does not execute serialized annotation payloads.

Versions and additional datasets

STD lives under std/1.0.0 and references shared assets under assets/r1. Additional datasets or versions can use independent configuration files and release directories. See the uploader guide and configuration template. No unverified v2/v3/v4 source paths are assumed.

Source inventory and known data issues

The completed source inventory contains about 6.814 TB of retained original files before archive overhead/compression. This is a source size, not uploaded progress.

One source scene, train/scene_02_01/scene_00069782, lacks the standard file realsense/rgb/0000.png and instead contains realsense/rgb/0000_0001.png. Both source names and data are preserved without assuming that these files are equivalent. This scene is excluded only from the default complete-scene train list; it remains in the archives. The scene index and known-source-anomaly manifest explicitly record its status. Source train scene count is 982731; the default complete-scene train list has 982730 scenes. Test has 163987 complete scenes. Auxiliary annotation groups are separately preserved.

Release status and terms

Transfer completion, metadata validation, and license documentation are separate statuses. Croissant metadata and complete manifests will be generated after verified transfer. Schema validation, author review, and upstream asset license/provenance documentation remain pending. Original upstream asset terms continue to apply; this repository does not assign a blanket new license to third-party models or textures. See third-party notices.

_transport_tests/ contains synthetic transport test files, not benchmark samples.