DATASET CHANGELOG
What changed in physical AI datasets
A source-backed log of dataset catalog updates, commercial use interpretation changes, and new physical AI data discovery signals.
DIRECT ANSWER
This dated, source-anchored ledger records what changed in public robotics datasets: before and after values, changed field, change class, source version and time, retrieval hash, buyer impact, reviewer, and checked date. Use it to track robotics dataset changes, dataset version changes, and any dataset license change. It includes the LeRobot v3 migration and OpenVLA's warning that the OXE Bridge component is outdated. Subscribe by RSS or pull one complete change per NDJSON line; a rights change is always routed to human or needs-review status.
Canonical ledger
Entries are newest first. IDs and revisions are immutable; a correction is appended as a new record and keeps the original record addressable. Download the same ordered records as NDJSON or subscribe to the RSS feed.
| Record | Before → after | Evidence | Buyer impact and review |
|---|---|---|---|
LeRobot v3 storage and migration guidance recordedchange-lerobot-v3-migration-2026-07-22-r1 · revision 1schema · storage format and migration path LeRobot documents a v3 dataset layout using shared Parquet and MP4 shards plus relational metadata, with a migration path from v2.1. | Before: LeRobotDataset v2.1 layout2.1 schema-versionAfter: LeRobotDataset v3.0 layout 3.0 schema-version | LeRobotDataset v3.0 Primary docs source Locator: Directory layout (simplified); Migrate v2.1 → v3.0 Checked: 2026-07-22 sha256:4dee918a5898405a94140466f1d53f13f44a506d4d372e267d5dae5594cd10feConfidence: high
| Pin the repository revision, inspect v3 metadata, and run the migration guidance before assuming an older loader can consume the dataset. Review: human LeRobot datasets profile · Readiness doctor · License checker |
OpenVLA flags the OXE Bridge component as outdatedchange-oxe-bridge-outdated-2026-07-22-r1 · revision 1version · BridgeData V2 component source OpenVLA reports that the BridgeData V2 component in Open X-Embodiment was out of date as of 2023-12-20 and directs users to the official bridge_orig source. | Before: OXE BridgeData V2 componentoxe-bridge-v2After: Official BridgeData V2 bridge_orig bridge_orig | OpenVLA README observed 2026-07-22 Primary project source Locator: README.md#vla-pretraining-datasets Checked: 2026-07-22 sha256:71e4be636dfc677ecb33bb005b6c0b2586f80640cf44af2968a48ee0c08ca212Confidence: high
| Do not treat the OXE Bridge component as the current BridgeData V2 source without comparing the pinned inputs and loader path. Review: human Open X-Embodiment profile · Readiness doctor · License checker |
Commercial-use and consent-risk fields addedchange-droid-risk-fields-2026-05-01-r1 · revision 1catalog · catalog risk fields Dataset profiles now separate source metadata from conservative commercial-use and consent-risk interpretation fields. | Before: Source metadatasource-metadataAfter: Separate commercial-use and consent-risk interpretation fields separate-risk-fields | DROID project page observed 2026-05-01 Primary project source Locator: Project page and linked dataset documentation Checked: 2026-05-01 sha256:5c8bbd4ad4647d2fb9de6b963db66f860370d90ca4c3b76f0d4a669c1909efb3Confidence: medium
| Treat these fields as review prompts; they are not legal certification and do not establish downstream rights or consent. Review: human |
Egocentric video facet launchedchange-ego4d-facet-2026-05-01-r1 · revision 1catalog · modality facet Ego4D, EPIC-KITCHENS, HOI4D, and other first-person video datasets are grouped for comparison. | Before: No egocentric-video facetabsentAfter: Egocentric-video facet egocentric-video | Ego4D project page observed 2026-05-01 Primary project source Locator: Project overview Checked: 2026-05-01 sha256:844aee20f895d4a3fff4e927ab7b2ef9e9c40e9a8ae4ba3034625bb9f9527440Confidence: medium
| Use the facet for discovery, then verify task, environment, provenance, rights, and transfer fit on each profile. Review: human |
Hugging Face robotics dataset watchlist addedchange-hf-robotics-watchlist-2026-05-01-r1 · revision 1catalog · catalog discovery coverage The catalog now watches public Hugging Face robotics and embodied-AI repositories as discovery candidates. | Before: Curated dataset profilescuratedAfter: Hugging Face robotics watchlist hf-robotics-watchlist | Hugging Face dataset search observed 2026-05-01 Primary dataset source Locator: Dataset search results for robotics Checked: 2026-05-01 sha256:25e32ef86c618e66a5644af5b800b3b9ea31a1486c5926275c77476fe0ba5e6cConfidence: medium
| Watchlist presence is a discovery signal, not evidence that a dataset is suitable, licensed, consent-cleared, or loader-ready. Review: human Open X-Embodiment profile · Readiness doctor · License checker |
How the ledger is maintained
Each diff cites the affected dataset's own page, repository, or documentation with an exact locator, retrieval hash, and checked date. HTML, NDJSON, and RSS use the same newest-first records. Source-reported values stay distinct from TrueLabel analysis: the change class, normalized diff, and buyer-impact note are editorial structuring.
Limitations
- The ledger makes cited diffs checkable; it cannot prove every upstream change has been captured.
- Human review status records who must judge a rights change; it cannot prove that judgment is correct.
- Subscription value is longitudinal and cannot be established from one snapshot.
TRUELABEL ROUTING
Replace a stale slice with cleared, current data
Inspect the affected profile, rerun the readiness and license checks, then scope a replacement or supplement for the broken source.
KEEP DIGGING
Use this record as part of a broader dataset review
A dataset record is only useful when it connects into the rest of the buyer workflow. The next review step is usually not another summary; it is a fit check, rights triage, source comparison, or custom request spec that names the missing proof.
For physical AI teams, the hard question is whether the public source can support a specific model objective under real deployment constraints. That requires adjacent dataset records, tools, comparisons, and sourcing paths, plus external references that a reviewer can open and challenge.
Use the links below to keep the review grounded. Start broad when discovery is incomplete, move into profile and comparison pages when the candidate source is known, and switch to custom collection when the blocker is rights, consent, geography, robot embodiment, or target environment coverage.
Where to go next
- Robot training data marketplaceRobotics datasets
- Physical AI dataset catalogUse the catalog to compare source-backed dataset profiles by modality, task, rights signal, consent risk, and deployment fit.
- Hugging Face robotics indexScan the broader robotics dataset surface before narrowing into promoted profiles, comparisons, and custom collection specs.
- Dataset fit checkerScore whether a public source is enough for the model, rights path, modalities, and target environment.
- License risk checkerSeparate source license language from contributor consent, redistribution, private-space risk, and model-use assumptions.
- Data spec generatorTurn a public-source gap into a scoped capture request with sample QA, metadata, and delivery requirements.
- Vendor alternatives hubCompare data providers when the answer is not another public dataset but a better sourcing or capture route.
- Data annotation companiesUse the company index to separate annotation vendors, data engines, marketplaces, and specialist capture teams.
Other places to verify the claims
- Scale AI physical AI data engineMarket context for why physical AI systems need custom, enriched, real-world data beyond generic labeling workflows.
- LeRobot documentationRobotics dataset and tooling context for Hugging Face based collection, sharing, conversion, and training workflows.
- Open X-EmbodimentA cross-embodiment robotics dataset reference for comparing trajectory scale, robot diversity, and VLA training assumptions.
- DROID datasetA large in-the-wild robot manipulation dataset reference for real-world trajectory capture and deployment transfer risk.