Open robot dataset census
Answer in brief
On 8 November 2026, the public LeRobot dataset index listed 84,331 datasets. In a random sample of 978 public datasets, 0.3 percent carry no licence tag, 3.9 percent record no robot type and 3.2 percent have no image stream. Median size is 10 episodes.
Content reviewed 8 November 2026. How this is assessed.
Vorrick and this reference are run by the same operator. Vorrick is excluded from this page, and this page ranks nothing.
What does this census of open robot datasets measure?
This census reads metadata from the public LeRobot dataset index. It reports shares across 978 public datasets chosen at random with seed 20261108, read on 8 November 2026. Only summary figures are published, and no dataset is named.
The listing held 84,331 datasets. Of 1,000 sampled, 7 were gated or private and skipped, 15 had no metadata file and 0 failed. Nothing beyond the metadata file was read, and no episode data was downloaded. Download the numbers as census.csv or census.json, and the script as fetch.py.
What licence do open robot datasets on this index use?
In this sample of 978, 0.3 percent carry no licence tag, per the public LeRobot dataset index. A tag is what the uploader wrote, and it is not evidence of the rights held in the data.
Licence tags in 978 sampled datasets. A dataset can carry more than one tag.
| Licence tag | Datasets | Share |
|---|---|---|
| apache-2.0 | 960 | 98.2% |
| cc-by-4.0 | 7 | 0.7% |
| mit | 3 | 0.3% |
| No licence tag | 3 | 0.3% |
| cc-by-sa-4.0 | 2 | 0.2% |
| cc-by-nc-nd-4.0 | 1 | 0.1% |
| cc-by-nc-4.0 | 1 | 0.1% |
| other | 1 | 0.1% |
Do open robot datasets record what robot was used?
3.9 percent of sampled datasets record no robot type, per the public LeRobot dataset index, and the rest use 171 distinct values. Every other recorded value, named commercial models included, is grouped here as other recorded value.
Robot type field. An open design from an academic or community project is named. Everything else is grouped.
| Robot type group | Datasets | Share |
|---|---|---|
| Other recorded value | 910 | 93.0% |
| Not recorded | 38 | 3.9% |
| Koch low-cost arm (open community design) | 17 | 1.7% |
| ALOHA (open academic design) | 13 | 1.3% |
How many episodes do open robot datasets have?
The median sampled dataset has 10 episodes (quartiles 2 and 50), and 71.1 percent have fewer than 50, per the public LeRobot dataset index.
What frame rate do open robot datasets use?
The most common stated frame rate in this sample is 30 fps, in 77.7 percent of datasets, per the public LeRobot dataset index.
Frames per second in the metadata of 978 sampled datasets. Values written 30.0 (2) and 50.0 (1) are counted with 30 and 50.
| Frames per second | Datasets | Share |
|---|---|---|
| 30 | 760 | 77.7% |
| 15 | 61 | 6.2% |
| 20 | 52 | 5.3% |
| 10 | 52 | 5.3% |
| 50 | 23 | 2.4% |
| 25 | 14 | 1.4% |
| 21 | 4 | 0.4% |
| 5 | 4 | 0.4% |
| 120 | 1 | 0.1% |
| 109 | 1 | 0.1% |
| 100 | 1 | 0.1% |
| 60 | 1 | 0.1% |
| 40 | 1 | 0.1% |
| 33 | 1 | 0.1% |
| 31 | 1 | 0.1% |
| 2 | 1 | 0.1% |
How many cameras do open robot datasets record?
3.2 percent of sampled datasets have no image or video stream in their metadata, and the most common camera count is 2, per the public LeRobot dataset index.
Image and video streams declared per dataset.
| Streams | Datasets | Share |
|---|---|---|
| 0 | 31 | 3.2% |
| 1 | 192 | 19.6% |
| 2 | 498 | 50.9% |
| 3 | 197 | 20.1% |
| 4 or more | 60 | 6.1% |
Which format version do open robot datasets use?
The most common format version in this sample is v3.0, in 73.7 percent of datasets, per the public LeRobot dataset index.
Format version in each sampled dataset's metadata.
| Format version | Datasets | Share |
|---|---|---|
| v3.0 | 721 | 73.7% |
| v2.1 | 230 | 23.5% |
| v2.0 | 27 | 2.8% |
Do licences on re-uploaded robot datasets match the original?
In 32 repositories identified by hand as re-uploads of BridgeData V2, DROID, FMB, whose project pages state CC BY 4.0, 29 carry a different licence tag, 1 carries none and 2 match, per the public LeRobot dataset index.
Each pair was read by hand on the repository name, and on the card where the card says more; most cards are template text, so identification rests mainly on the name. Sample size is small and re-uploads can copy each other, so this is not a rate for all re-uploads. A tag is what the uploader set. It does not change the terms of the original data, and this page draws no legal conclusion. The upload code in the conversion tool defaults the card to apache-2.0, read on 8 November 2026, which may explain some of the mismatch. Check the upstream licence in the dataset spec ledger.
What can this census of open robot datasets not show?
It cannot separate original recordings from re-uploads, so one source dataset can appear several times. It cannot say which datasets are widely used, since it samples uniformly and takes no account of downloads, so popularity bias runs the other way from a list of popular datasets.
It covers public repositories only. Gated and private datasets are counted as skipped and their contents are unknown. Metadata fields are self-reported, and a missing field may be missing on the page but true of the data. This sample is one read on one date, and the index changes daily.
Sources for this entry
- Conversion tool source, card default: Read for the default licence value on the dataset card.
- The public LeRobot dataset index: Enumerated to 84,331 tagged datasets on 8 November 2026, then each sampled dataset's meta/info.json was read once, anonymously, at no more than 2 requests a second.
Continue this reading path
- Best Robot Training Data category guide
- How to choose: a robot training data provider
- Resource directory
Cite this entry
Best Robot Training Data. "Open robot dataset census." Content reviewed 8 November 2026. https://bestrobottrainingdata.com/open-robot-dataset-census/