FR3 cube stacking — 3dgs — full dataset + V9 diagnostic decisions
All 10,000 episodes / 1,537,140 frames are included, including HOLD and REJECT. No episode selection or reindexing was performed. train is a loader label, not training approval or an independent evaluation split.
| ACCEPT | HOLD | REJECT |
|---|---|---|
| 9866 | 10 | 124 |
Source: AIDASLab/fr3-3cube-10k-pde, revision 39fa3063c2163985af71037eb27d00db5be5cd95. Binary data, video, and episode metadata are preserved from the exact locally verified filtering input. publication.json records derivations, original hashes, and portable provenance. No new license is granted here; consult the source owner for applicable rights and terms.
Read decisions
Join filtering/episode_decisions.parquet to native rows by episode_index. native_gate_decision is ACCEPT/HOLD/REJECT; diagnostic_status is PASS/REVIEW/FAIL. filtering/checks.jsonl.gz contains all 23 per-episode check slots, including SKIP/NOT_REACHED and raw evidence. filtering/policy.yaml freezes V9 authority rules. Filtering used action.eef_delta_metric; native columns are not renamed or silently substituted.
Limitations — read before use
This is native diagnostic output, not a production canonical/training release. ACCEPT does not establish task success, physical validity, or that all 23 checks passed. training_promotion_allowed remains false. VLM and independent success verification were not run. Seed lineage remains unknown. Collision-solver and video–state projection checks are excluded and explicitly unverified. Missing signals remain SKIP/NOT_REACHED, not PASS. NT has inferred pseudo-actions and no measured state/FK authority.
The same V9 statistical-authority policy applies to all four datasets. Statistical extreme_value REVIEW is advisory, not a physical rejection boundary; raw evidence is retained. Other REVIEW/FAIL results remain effective. Heuristic motion thresholds are not proof of physical faults.
filtering/historical_frame_diagnostics.parquet contains historical diagnostic masks, not recomputed V9 masks and not approved training loss masks. Local path references are replaced with opaque local-artifact-sha256: IDs. Nonfinite diagnostic sentinels use tagged JSON objects; source action arrays were not modified. Historical provenance may contain uploaded: false, describing the original diagnostic run, not this publication.
Integrity and access
SHA256SUMS.json covers every published artifact except itself and Hub-managed .gitattributes. The repository is public with manually approved gated data access. Approval requests are handled by the repository owner.
- Downloads last month
- 15