DexYCB → VITRA Stage-1 (dexycb-vitra-streaming-v1)

Build status: SUPERSEDED — DO NOT USE FOR TRAINING.

This incomplete v1 build was stopped after its element-wise MANO rotation matrix closure threshold produced false-positive quarantines for geometrically valid sequences. It is retained only as an audit trail. Use LeoJiangOR/dexycb-vitra-streaming-v2, which measures rotation closure by SO(3) geodesic angle while retaining the independent 0.1 mm 3D-joint and 0.25 px 2D-projection hard gates.

This is a deterministic, quality-audited derivative of the official DexYCB dataset, licensed CC BY-NC 4.0. The conversion pins the official NVlabs/dex-ycb-toolkit split implementation at commit 64551b001d360ad83bc383157a559ec248fb9100.

What is published

Each accepted fixed allocentric camera view contains an RGB MP4 and one VITRA Stage-1 episode with camera intrinsics/extrinsics, MANO rotations, wrist translation, 21 world/camera joints, object transforms, visibility masks, and a causal templated command. MANO model files are not redistributed.

The canonical action trajectory comes from sequence-level world MANO. Every camera is retained only when world→camera MANO, official camera MANO, 3D joints, 2D projections, and object transforms close numerically. Training indices use only current frames where the active hand is visible and a valid t→t+1 action exists; the complete trajectory remains in the episode file.

The default dexycb_train/val/test split is official setup S0: 800/40/160 physical sequences before quality quarantine. All views of a physical action stay together. S1/S2/S3 membership is preserved as metadata. Because multiple views share one action, Build/sampling_groups/ should be used for hierarchical sequence→camera sampling when equal physical-action weighting is desired.

Language policy

DexYCB has no recorded natural-language command. text and text_rephrase are causal templates generated only from the official pick-up protocol, ycb_grasp_ind, and active hand side. TEMPURA-Qwen2.5-VL is run on the best visible completed clip only as a non-authoritative semantic audit. Its output is future-observing and is never inserted into training text.

Streaming safety

One official subject archive is held on disk at a time. Each physical sequence is converted, checked with an all-frame RGB detector audit, smoke-tested through the VITRA loader, uploaded atomically, and verified at the immutable Hugging Face commit before its raw and converted local files are removed. The local SQLite WAL ledger pins source hashes and conversion-code hashes.

Until a top-level COMPLETE file exists, do not treat this repository as a finished training release. DexYCB contains eight fixed allocentric views; it supplements but does not replace egocentric long-horizon VLA data.

Attribution

Please cite the original DexYCB CVPR 2021 paper and comply with its noncommercial license. See the official project page for citation and terms.