FlappyHimself/reorientation-artifacts

Reorientation artifacts

Current Stage1/Stage2 selection: 35 trajectories = 20 Phone + 10 Marker + 5 scissors. Phone data, checkpoints, BC artifacts and original references are preserved unchanged. Marker/scissors Stage1/Stage2 payloads contain only the 15 user-selected demonstrations listed below. Stage3 visualization supplements are indexed separately.

MANO / Retargeting / actual Rollout comparisons

24-video index · Manifest: FullDR with preceding nominal gravity curriculum; one fixedseed0 perexperiment. New3×3 fitted-MANO/reference/measured comparison, all30Hz frames retained. Original metrics unchanged; fitted MANO is explicitly a display approximation.

Historical selected WUJI2 FullDR checkpoint archive

Four unchanged final checkpoints and their original configs/context/resume records are now archived. This is a checkpoint-only archive: exact training H5 inputs differ from the existing Stage2 release, so evaluation dependencies remain incomplete. No evaluation was started. Reward >1 is a training-metric filter, not a policy-success claim.

Trajectory	FullDR run / checkpoint	Keypoint reward	Readiness
marker_005	rawbvfp3 / model_9998.pt	1.047750	Missing matching training inputs
marker_009	jdfgpdaq / model_9998.pt	1.008895	Missing matching training inputs
marker_011	i277x4dj / model_9998.pt	1.004174	Missing matching training inputs
scissors_new_006	58i6zoqd / model_9998.pt	1.070471	Missing matching training inputs

See archive manifest. Original Phone and Stage1/Stage2 files, including their immutable manifests and download inventories, are unchanged.

Selected Marker and scissors

• Marker: marker_003 through marker_012.
• Scissors: scissors_new_001, scissors_new_006, scissors_new_007, scissors_new_011, scissors_new_013.
• 15 Stage1 NPZs and the exact human-reviewed full videos; 15 Sharpa and 15 WUJI2 Stage2 H5s, previews, configs, asset snapshots and QA reports.
• Stage2 kinematic QA: Sharpa 5 PASS / 10 HOLD; WUJI2 1 PASS / 14 HOLD. PASS here is kinematic QA only, not physics or teacher acceptance.
• Fresh sync audit flags remain for marker_011 and scissors_new_001; human acceptance is recorded separately.

See the 15-row delivery table, CSV, Data Card, release manifest and catalog.

objects/<marker|scissors>/trajectories/<take>/
  manifest.json
  stage1/selected15-20261008-v1/          # exact Stage2-parent NPZ, reviewed video, sync audit, frame mapping
  sharpa/stage2/selected15-20261008-v1/   # trajectory.h5, QA, preview, portable handoff
  wuji2/stage2/selected15-20261008-v1/    # trajectory.h5, QA, preview, portable handoff
provenance/selected15-20261008-v1/         # selected manifest, portable assets/configs, download tooling

Download only these 15 Stage1/Stage2 trajectories

For the frozen Stage1/Stage2 release below, set REV=7e6bef8f22540f90732f8053ccf578a44b367725. The script requires an immutable revision and verifies every downloaded SHA256.

python -m pip install huggingface_hub
REV=7e6bef8f22540f90732f8053ccf578a44b367725
hf download FlappyHimself/reorientation-artifacts --repo-type dataset --revision "$REV" \
  provenance/selected15-20261008-v1/download_selected15.py --local-dir selected15
python selected15/provenance/selected15-20261008-v1/download_selected15.py --revision "$REV" --output selected15
python selected15/provenance/selected15-20261008-v1/materialize.py --root selected15 --output selected15-runtime-assets

The download inventory contains only selected trajectories and their shared dependencies. Existing full hand calibration/license packages are not newly copied into this delivery. Historical absolute paths inside H5/source_metadata are provenance; consume repository-relative handoff.json paths. materialize.py creates a separate asset/config copy with local paths and records original/rebased hashes. Original data bytes are never modified.

Stage1 uses meters, wxyz quaternions and objectMeshA axes. Stage2 uses meters/radians, wxyz quaternions and Vicon world axes translated to the first object-mesh origin; timestamps and frame ordering are preserved. See the Data Card for exact conversion. The NPZ contains the longest jointly valid continuous interval; the reviewed video covers the full source timeline and may be longer. This is documented per trajectory.

Preserved Phone coverage

• Phone: 20 preserved Stage1 payloads with freshly generated full Stage1 previews, 20 Sharpa Stage2 references, ten selected teacher checkpoints, ten existing BC archives and the DATA046 derived view. Per-take legacy sync verdicts remain unchanged; legacy calibration source-file hashes are retained where the original files are not bundled.
• Missing evidence stays missing: phone_15 has no located selected teacher. Ten per-phone GRU student registrations have no bundled checkpoint. Existing UR10e evaluation evidence is not a substitute for missing checkpoint bytes. No stand-alone frozen teacher config is fabricated from a card.
• BC normalization and split members remain inside their original archives; dataset_layout.json states their locations. phone_07 and phone_18 BC producers differ from selected teachers, and DATA018's historical readiness conflict is preserved. Stage2's native trajectory.h5 is an alias of canonical_retarget.h5, not an additional demonstration. UR10e bound releases retain explicit pointers to their complete original bundles.

History and boundaries

The earlier full inventory and superseded Stage1 release remain available at revision d4e2483b88ce. The selected15 Stage1/Stage2 release removed the 11 unselected Marker/scissors data payloads and their dedicated sync files. New Stage3 visualization-only entries do not restore those Stage1/Stage2 payloads. The external raw-capture repository is unchanged.

Object mass 0.2 kg and diagonal inertia 1e-4 in diagnostic assets are placeholders, not measured physical properties. Structural checks and visual/kinematic QA do not establish physical calibration, training readiness or policy success. Original licenses remain applicable; no additional downstream rights are granted.

Selected FullDR checkpoint evaluations (2026-10-11)

Versioned evaluation report: marker_009 seeds0/1/2 completed, final RMS8.21mm, full-hand geometric hold0/3. Other3 selected models await original code; their9 requested cases are NOT_RUN. Exact training H5/geometry dependencies are now supplied separately from existing Stage2. Videos show original MANUS skeleton above measured policy rollout; formal MANO evidence is unavailable.