OakInk2 tracking references for dex-track

A compact derivative of OakInk-v2 for human-hand / rigid-object pose tracking. Prepared with dex-track and the pinned FlashDexRetarget preprocessing code.

This release contains 627 recordings, 2840 annotated stages, shared object meshes and three prepared single-task packs. It does not include RGB video, SMPL-X body parameters or MANO model weights.

Contents

• reference/*.npz: lossless compact arrays extracted from the original annotation recordings: frame_ids, fps, rh__pose_coeffs, lh__pose_coeffs, per-hand tsl / betas, object_ids and object_pose.
• reference/manifest.json: recording coverage and provenance.
• reference/object_repair.tar.gz: shared repaired object meshes/descriptions.
• reference/program.tar.gz: original stage boundaries, action/hand-object labels and language descriptions.
• annotations/stage_objects.json: complete stage inventory with required, candidate and background/context object evidence, exclusions and uncertainty.
• annotations/rigid_candidates.json: 981 preparation candidates. Another 1269 stages are excluded by the current independent-rigid-body scope; 590 require review.
• annotations/audit_v2_summary.json: audit coverage and decision summary.
• training/whiteboard, training/pour_milk, training/flask: packed motion plus the scene's object assets. These are example tracking packs, not the complete eligible corpus.
• release.json and SHA256SUMS: release scope and integrity.

Coordinates and time

Compact references preserve the original OakInk2 120 Hz, y-up coordinates. MANO pose_coeffs hold wxyz quaternions, root followed by 15 local joint rotations. The compact arrays are not robot-joint coordinates. tsl is the dataset's MANO wrist translation; model shape is represented by betas. Object poses have axes [frame, object, row, column] and are homogeneous 4×4 transforms.

Prepared training packs are 60 Hz, converted to the simulator's z-up scene. motion.pt['meta']['fps'] is authoritative. The dex-track loader independently interpolates each clip to 50 Hz; physics runs at 200 Hz. The hand's retargeted robot joints are used only for reset. Policies track MANO and object references.

Example	Prepared frames	Scene objects
whiteboard	786	2
pour_milk	1414	2
flask	1270	3

Readiness and limitations

Full-corpus training admission is not complete. The 981 entries are preparation candidates, not a validated physics benchmark. Full-corpus collision, reset retarget, support/contact checks and final admission are ongoing. Object annotations combine source labels/descriptions with automated geometric evidence; they are not asserted to be complete human ground truth. Keep each record's unresolved issues and provenance.

The simulator currently supports independent rigid objects, up to eight per clip, with one fixed associated object per wrist. Soft-body/material-removal and coupled articulated mechanisms are excluded or marked for review. Pouring means container pose tracking, not fluid simulation; whiteboard means hand/eraser/board tracking, not simulated ink removal. The existing examples do not establish successful learned policies or equivalence of Isaac and MuJoCo dynamics.

Download and reproduce

Use scripts/download_tracking_data.py from dex-track with --part all, --part reference, --part annotations, or --part training. Pin a Hugging Face commit with --revision for reproducibility. The script safely extracts the two reference archives. Verify downloaded content with:

sha256sum --check --ignore-missing SHA256SUMS

The code repository includes original-data download, lossless compacting, stage object consolidation, rigid-scope auditing, MANO forward kinematics, collision preparation, reset-only retargeting and packing scripts. MANO model files must be obtained separately under their license and are never redistributed here.

Attribution and license

Source: OakInk2 project, official toolkit, and official dataset distribution. Original OakInk2 creators retain attribution. This derivative preserves the source distribution's Creative Commons Attribution-ShareAlike 4.0 license. Modifications are compact storage, derived stage-object/eligibility metadata, simulation coordinate/time conversion and robot reset/collision preparation. Cite the original OakInk2 work when using these data and describe any rigid-only selection or simulation approximations in resulting evaluations.