Data Card: DATA-RAW-HF-PUBLIC-20261008-V1 / reorientation-raw-stage1-public-v1
| Field | Value |
|---|---|
| Data Task ID | DATA-RAW-HF-PUBLIC-20261008-V1 |
| Dataset ID | reorientation-raw-stage1-public-v1 |
| Status | COVERAGE_GAP: licensed hand/model dependencies intentionally omitted; source calibration provisional |
| Purpose | Public archive of capture inputs, frozen spatial calibration and raw-to-Stage1 code |
| Source mode | derived_view |
| Source lineage | raw_data/manifests/raw_to_stage1_v1.json SHA256 e998266c13f0926cd28fffa3fcc4d74eabbaeb981c2a22057652d021a411061d |
| Episodes / splits | 26 paired capture episodes; diagnostic-only, no policy train/test split |
| Sampling | Original raw files unchanged. Producer selects longest continuous simultaneously valid hand/object interval at original MANUS timestamps |
| Observation | Privileged human joint capture, Vicon hand/object rigid poses, object geometry and calibration |
| Action / label | No executed actions, reward, policy rollout or state-transition labels |
| Normalization | No learned statistics; MANUS cm→m, Vicon mm→m; frozen axis mapping |
| Producer | Frozen pipeline/stage1_v1; main 878e1a25d4c69b97c66905139a78561184f29893; Sharpa 83fb2afb5d111afb74418c239c01d7b1a6bb8fb4 |
| Artifact / manifest | UPLOAD_MANIFEST.json SHA256 c3de79ac3854b06c51cfa7a692c5867a1e27fa2d1c1b4414c760580bb964eec3; member paths, sizes and hashes materialized |
| Claim boundary | Public subset is not independently runnable until licensed dependencies are restored; no physics or training acceptance |
Source and Provenance
| Artifact | Identity / value | Evidence |
|---|---|---|
| Spatial offset | af948e4cdd200dfe52f14afa8d91b913b33b4a95d9ffc4079b113aed497734bd | raw_data/calibration/raw_to_stage1_v1/fixed_offset.json |
| Frozen source manifest | e998266c13f0926cd28fffa3fcc4d74eabbaeb981c2a22057652d021a411061d | raw_data/manifests/raw_to_stage1_v1.json |
| Published projection | 560 source members; 11 licensed-resource files omitted | UPLOAD_MANIFEST.json |
| Projection transforms | Symlinks become regular files with identical byte hashes; no numeric fields changed | included_files and excluded_files |
Schema
| Field | Shape / type | Semantics |
|---|---|---|
| MANUS CSV | N rows, original header-defined columns | Frame/time and 25 joint positions (cm), global extrinsic xyz Euler rotations (degrees); None/IMU/Tracker files preserve original columns |
| Vicon CSV | N rows, native multi-body header | Frame/subframe and per-rigid-body translation (mm) and rotation vectors; invalid observations retained |
| Calibration JSON | 4×4 float matrices, configuration objects | T_A_from_B maps B to A; translation metres; fixed offset and object extrinsics |
| VSK/XML and marker layout | Original structured files | Marker labels and rigid-body definitions, not per-frame dynamics |
| Object OBJ | Vertex and face records | Original rigid mesh geometry, interpreted in the corresponding frozen object calibration |
| frame_validity.npz | frames int64[N]; timestamps float64[N]; hand_valid/object_valid/valid bool[N]; scalar time offset and sync gate | Full original MANUS time axis and interpolation validity, no mesh payload |
| Manifests / reports | JSON objects and per-file SHA256 | Relative active paths, source identities, episode membership and selected valid frame ranges |
| Stage1 output specification | Documented, not uploaded generated data | Full array schema in raw_data/README.md; no MANO-derived output mesh NPZ is distributed here |
| Historical archive files | Original mrec/xcp/system and photos | Supporting source evidence; opaque capture formats remain unchanged, not substitute dynamic input CSVs |
Temporal Alignment
| Item | Contract |
|---|---|
| Timing | t_vicon = t_manus + per-recording dt; no universal temporal offset |
| Sequence | Raw capture at t → fixed rig/shape → same-t Vicon pose → objectMeshA Stage1; not physics state_t+1 |
| Missing samples | Maximum 50 ms interpolation gap; select longest contiguous valid interval; original invalid frames retained |
| Sync gates | 16 PASS / 10 HOLD; PASS is estimator consistency, not hardware ground truth |
Splits, Sampling, and Leakage
| Split | Episodes / rows | Purpose |
|---|---|---|
| diagnostic-only | marker_001, marker_002, marker_003, marker_004, marker_005, marker_006, marker_007, marker_008, marker_009, marker_010, marker_011, marker_012, scissors_001, scissors_002, scissors_new_001, scissors_new_002, scissors_new_003, scissors_new_004, scissors_new_006, scissors_new_007, scissors_new_008, scissors_new_009, scissors_new_010, scissors_new_011, scissors_new_012, scissors_new_013 | Frozen capture inspection and reproducibility; no held-out policy performance claim |
| coverage-gap | scissors_new_005 lacks corresponding Vicon CSV | Explicitly skipped, not fabricated |
Shared spatial calibration was developed on earlier captures. The 26 episodes are not independent heldout evidence. No windows cross episode boundaries and no normalization statistics are fitted.
Normalization
| Field | Contract |
|---|---|
| Units / axes | F=diag(-1,1,1); positions Fp, rotations FRF; canonical rotations output as wxyz |
| Learned normalization | None; no train/validation statistics |
| Labels | No action labels; numerical raw inputs unchanged by this projection |
Validation and Readiness
| Gate | Result | Evidence |
|---|---|---|
| Local identity | Selected source members verified against frozen hashes | UPLOAD_MANIFEST.json |
| Local reproduction before exclusions | marker_004: 772 frames; scissors_new_006: 1079; 23 arrays each matched exactly, including relocation | raw_data/manifests/freeze_validation_v1.json |
| Runnable archive closure | COVERAGE_GAP until omitted hand_shape files restored at exact hashes | excluded_files |
| Remote transfer | Checked by publication job against uploaded bytes at pinned commit; receipt retained locally | HF commit and local upload verification receipt |
| Training / physics readiness | NOT ACCEPTED | Provisional offset and retained HOLDs |