Upstream report
Title
lerobot/droid_1.0.1: task_category and date duplicate unrelated semantic fields on every
referenced frame
Summary
At dataset revision 0eabc778f959c54b8c5aa3626cc1128d2d2e54d4, I verified all 86 Parquet
files referenced by episode metadata: 27,630,375 rows and 95,658 episodes.
- task_category is byte-for-byte equal to building on every row.
- date is byte-for-byte equal to collector_id on every row.
- zero date values match YYYY-MM-DD.
The task-category cause is visible in the pinned conversion script: the destination task_category
is populated from source building instead of source task_category. I am keeping the date root cause
separate because I have not pinned where that substitution occurred.
Independent source check
I joined a generation-pinned 100-episode slice back to official DROID raw metadata through the
official annotation index and raw manifest. Source objects linked to multiple LeRobot episodes were
quarantined first. All 100 selected identity checks passed: UUID, instruction tuple, collector,
building, exact boolean success/outcome, episode length, and finite six-element extrinsics for all
three cameras. Both published fields disagreed with their raw source on all 100 episodes.
The artifacts omit verbatim source-object names and directly copied identity fields. They retain
episode indexes and public-source fingerprints, so they are linkable and are not presented as
anonymous or broadly PII-free.
Public evidence
The release includes SHA-256 manifests, exact row counts, a sparse repair sidecar, and an independent
exact-generation replay. Its CI runs 15 tests, Ruff, checksum verification, and a package build.
Requested action
- Mark task_category and date invalid for the current dataset revision.
- Fix the converter field mapping and add a semantic collision/domain check.
- Rebuild these columns from source metadata, or remove them until a complete source-backed rebuild
is available.
- Update the v3.0 tag after the corrected dataset revision is published.
This report is unrelated to and does not duplicate the known orphan Parquets in discussion 9 or
the repaired videos in discussion 8. It also does not claim that robot states, actions, videos, or
success labels are corrupted, and it does not claim a measured policy-quality improvement.
Upstream report
Title
lerobot/droid_1.0.1: task_category and date duplicate unrelated semantic fields on every
referenced frame
Summary
At dataset revision 0eabc778f959c54b8c5aa3626cc1128d2d2e54d4, I verified all 86 Parquet
files referenced by episode metadata: 27,630,375 rows and 95,658 episodes.
The task-category cause is visible in the pinned conversion script: the destination task_category
is populated from source building instead of source task_category. I am keeping the date root cause
separate because I have not pinned where that substitution occurred.
Independent source check
I joined a generation-pinned 100-episode slice back to official DROID raw metadata through the
official annotation index and raw manifest. Source objects linked to multiple LeRobot episodes were
quarantined first. All 100 selected identity checks passed: UUID, instruction tuple, collector,
building, exact boolean success/outcome, episode length, and finite six-element extrinsics for all
three cameras. Both published fields disagreed with their raw source on all 100 episodes.
The artifacts omit verbatim source-object names and directly copied identity fields. They retain
episode indexes and public-source fingerprints, so they are linkable and are not presented as
anonymous or broadly PII-free.
Public evidence
https://github.com/sawhney17/droid-semantic-repair/releases/tag/v0.1.0
https://huggingface.co/datasets/lerobot/droid_1.0.1/tree/0eabc778f959c54b8c5aa3626cc1128d2d2e54d4
https://github.com/huggingface/lerobot/blob/22bd7a2/examples/port_datasets/port_droid.py#L250-L256
The release includes SHA-256 manifests, exact row counts, a sparse repair sidecar, and an independent
exact-generation replay. Its CI runs 15 tests, Ruff, checksum verification, and a package build.
Requested action
is available.
This report is unrelated to and does not duplicate the known orphan Parquets in discussion 9 or
the repaired videos in discussion 8. It also does not claim that robot states, actions, videos, or
success labels are corrupted, and it does not claim a measured policy-quality improvement.