Overview GitHub Hugging Face arXiv Orgs New

What the public robotics data actually looks like

Thousands of datasets and repositories are published as the foundation of robot learning. Very few people have looked at them together. We index both the hardware and the software, so we can ask questions neither half answers alone.

Three in particular: which of this is still maintained, what you are legally allowed to build on, and which robot each dataset was actually recorded on.

Where we look

Hugging Face datasets

Updated weekly · 1,467 tracked

5.3M monthly downloads 576 robots identified

Every robotics dataset above 100 monthly downloads, with its license read, the robot it was recorded on where the card says, and whether that robot can still be bought.

Summary

GitHub repositories

Updated monthly · 2,410 tracked

49% stale or archived 28% no usable license

Robotics code with 50 or more stars, across 78 topics. Categorised, and checked for whether anyone has touched it in the last year.

Browse

The full dataset list

Filterable · 1,467 rows

Everything behind the summary, searchable and filterable by license, by robot, by publisher. Shareable as a link with the filter applied.

Open

Open X-Embodiment, arXiv, Zenodo

Not yet automated

The aggregated collections, the papers that name what trained on what, and the DOI repositories where academic releases outlive the labs that made them.

In progress

What we have found so far

The most-used robot in machine learning cannot be bought. Franka appears in more datasets than any other platform. The Panda was withdrawn in favour of the FR3, Franka Emika filed for insolvency in 2023, and the official repository states the model is unsupported and reported issues will not be fixed. Not one dataset we harvested names the current model.

Half of public robotics code is no longer maintained. 49% of the repositories we track have had no commit in over a year. That includes reference implementations the field cites constantly — diffusion policy, mobile ALOHA, LIBERO.

Roughly a third of releases have no license you can act on. 28% of repositories and a similar share of datasets either declare nothing or pick an option that points nowhere. Open X-Embodiment, the field's reference collection, publishes 22 columns of metadata about its constituent datasets and not one of them is a license.

How this is built, and what it does not claim

These are defined samples, not censuses. Hugging Face alone carries 80,198 datasets under the robotics tag, most of them automatically generated by people trying the tooling. Every figure here describes a stated boundary — a download floor, a star floor, a topic list — rather than the whole field.
Licenses are read, not verified. We record what the publisher stated and when we read it. Nothing here is legal advice, and unclear means we could not establish the terms, not that the terms are bad.
Robot identification is partial. It resolves only where a tag, dataset card or README names the hardware. Where it does not, the field is blank rather than guessed.
Every claim carries a date. Licenses change, repositories get archived, vendors stop trading. Anything older than twelve months should be treated as stale, including ours.
The underlying data is open. Versioned JSON, CC BY 4.0, at github.com/robotindex/robotindex-data. Corrections with a source are welcome and will be credited.

The Robotics Report

Once a month: what changed in the public robotics data. New datasets worth knowing about, projects that went quiet, licenses that moved, and hardware that stopped being available. No marketing, no digest of other people's news.

One email a month. Unsubscribe in one click. We do not share the list, and we do not sell placement in the index — the verification is the whole point of it.