Lifted dataset smoke-test report
Date: 2026-09-21. Branch: chaoyi/new_dataset.
Success-only upload/download/render: 16/16 passed.
Dataset: https://huggingface.co/datasets/retarget/retarget_full
Verified data revision: 081bb1bd5825c07e7af7645e6663a9824ed6df28. All 408 checksummed files matched after a clean download; all 16 fresh videos are byte-identical to their published previews. The public verification report records rendering and Viser inspection of the downloaded data.
Verification evidence
- Converted four source clips into the current
retarget_exampledirectory layout, with one saved rollout for each of four robot hands (16 combinations). - 408 hashed files, 109,581,007 bytes before the checksum index (about 104.5 MiB).
- Every scene resolves its mesh dependencies within the converted dataset.
- All runs passed current
process_config()andload_data(), shape/time-grid/finite checks, and comparison of recomputed metrics to the source receipts. - All 16 fresh videos used local
setup_renderer()/render_image()with EGL. - All 16 trials were loaded and scrubbed in a real Chromium browser using Viser; no browser JavaScript errors. Playback, trial switching, and visibility toggles passed.
- Three regression tests passed: GLB node scale/ground alignment, preservation of hand/object offsets during synthetic lift, and success-only upload filtering. New Python files pass Ruff; the existing Viser module retains its 37 pre-existing lint findings with no new findings.
Per-example results
| Dataset | Robot | Frames | Lift (m) | Final error (m) | Load / render / Viser |
|---|---|---|---|---|---|
| dexycb_lifted | allegro | 175 | 0.3362 | 0.0072 | Pass |
| dexycb_lifted | xhand | 175 | 0.3333 | 0.0038 | Pass |
| dexycb_lifted | inspire | 175 | 0.3321 | 0.0026 | Pass |
| dexycb_lifted | sharpa | 175 | 0.3332 | 0.0072 | Pass |
| hrdexdb_24f | allegro | 175 | 0.2811 | 0.0167 | Pass |
| hrdexdb_24f | xhand | 175 | 0.2835 | 0.0199 | Pass |
| hrdexdb_24f | inspire | 175 | 0.2819 | 0.0178 | Pass |
| hrdexdb_24f | sharpa | 175 | 0.2709 | 0.0081 | Pass |
| oakink_lifted | allegro | 150 | 0.2506 | 0.0058 | Pass |
| oakink_lifted | xhand | 150 | 0.2542 | 0.0091 | Pass |
| oakink_lifted | inspire | 150 | 0.2491 | 0.0044 | Pass |
| oakink_lifted | sharpa | 150 | 0.2359 | 0.0034 | Pass |
| hot3d_v2 | allegro | 225 | 0.2437 | 0.0101 | Pass |
| hot3d_v2 | xhand | 225 | 0.2553 | 0.0169 | Pass |
| hot3d_v2 | inspire | 225 | 0.2580 | 0.0185 | Pass |
| hot3d_v2 | sharpa | 225 | 0.2560 | 0.0206 | Pass |
Inspect locally
- Gallery:
http://localhost:8082while the local gallery server is running. - Viser:
http://localhost:8080while the inspector is running. - Gallery files:
example_video_data/retarget_full_roundtrip/after_download/index.html. - Verification data:
example_video_data/retarget_full_roundtrip/after_download/verification.json. - Browser snapshots:
example_video_data/retarget_full_roundtrip/viser/. - Downloaded data:
example_datasets/retarget_full_download/.
These artifacts are local, Git-ignored outputs; use the guide to reproduce them.
Limits
The samples were chosen for successful outcomes, not representative success rates. The conversion preserves recorded robot NPZ files and reconstructs missing object collision meshes. This validates saved-state playback and current-loader compatibility, not original collision-mesh equivalence or control-replay dynamics. Only successful scored trajectories were uploaded; failed and weak-reference runs are excluded.