Can you use that VLA dataset commercially?
The robotics datasets everyone trains on are described as open, and open is a statement about access rather than about permission. A team downloads a well-known collection, trains a policy, and only when the model approaches a release does anyone ask what the terms permit. At that point the answer is expensive whichever way it goes. The terms are usually findable, and they are often not where you would look, and for the large collections the license on the front page is frequently not the license governing the part you used. What follows is where to look and what to record. It is not legal advice.
Key takeaways
- Open answers whether you can download a dataset. It does not answer whether you can train on it, and it does not answer whether you can ship the resulting model.
- For collections, the headline license often does not govern the part you are using. Open X-Embodiment places software under Apache 2.0 and other materials under CC-BY 4.0, and directs users to per-dataset terms for its constituents.
- The license is frequently not on the page you would check. On 16 September 2026 the DROID project page carried no license statement, while a copy hosted on HuggingFace declared CC-BY 4.0 in its card metadata.
- A mirror's license metadata is set by whoever uploaded that copy, which makes it evidence about the uploader rather than a statement from the publisher.
- Record the artifact, the URL, the stated terms, and the date you checked. That record is what a later review asks for and cannot be reconstructed afterward.
Open does not mean commercially usable
Three separate questions hide inside the word open, and a dataset can answer them differently.
Can you obtain it, which is about access and is usually the only question a download page addresses. Can you train on it, which is about permitted use and depends on the license attached to the data rather than to the code. Can you distribute what you trained, which is about derivative works and is the question that surfaces latest and costs the most to get wrong.
These are worth separating because they fail at different moments. Access is settled the day you download. Permitted use becomes relevant the day you train, which is usually months before anyone reviews it. Derivative distribution surfaces at release, when the cost of a wrong answer is highest and the option to retrain on something else is most expensive.
A permissive answer to the first implies nothing about the other two. Many research datasets are freely downloadable specifically because they were released for research, and that intent sometimes appears in terms and sometimes only in a paper's acknowledgements. Neither situation makes the question go away when a model is ready to ship.
Where the license really lives
Four places are worth checking, and they disagree more often than you would expect.
The paper
Papers describe datasets and rarely license them. A dataset paper may not mention terms at all, and where it does the statement is usually a summary rather than the operative text. Treat it as a pointer.
The project page
The obvious place, and not a reliable one. Checked on 16 September 2026, the DROID project site presented the dataset's composition, results, and access instructions with no license statement anywhere on the page. That absence is itself worth recording, because a dataset with no stated terms is not the same as a dataset with permissive terms.
The repository
Usually the most authoritative source, because it is maintained by the publisher and carries the operative text. The Open X-Embodiment repository is explicit and structured: all software is licensed under Apache 2.0, and all other materials under the Creative Commons Attribution 4.0 International License. It also distributes everything on an as-is basis without warranties of any kind.
The mirror
Copies on hosting platforms carry license metadata entered by whoever uploaded them. Checked on the same date, a HuggingFace copy of DROID declared cc-by-4.0 in its dataset card, and a HuggingFace copy of Open X-Embodiment declared the same. Those declarations tell you what the uploader believed or intended. They are not statements by the original publisher, and they are the single most common source of false confidence in this whole area.
What the widely used collections say
The Open X-Embodiment case deserves attention because it is the most used and the most instructive. Its dual structure separates code from data cleanly, which is helpful. The part that matters more is what the repository says about its constituents: it points users to per-dataset citations and terms rather than treating the collection as uniformly licensed.
That is the crux. A collection assembled from many contributing datasets can only grant what its contributors granted. A permissive collection-level license sits on top of constituent datasets that may carry their own conditions, and the subset you sampled for training is governed by whatever applies to those constituents. The meaningful question is therefore never "what license is this collection under," it is "what terms apply to the specific constituents I used." Two teams training on the same collection can be in different positions entirely, depending on which subsets they drew from.
The mirror problem
Mirrors solve a real problem, since original hosting is often slow, partial, or gone. They also break the chain between the terms and the publisher who set them.
The failure is rarely anyone acting in bad faith. An uploader fills in a required license field, picks the option that matches their understanding, and the platform then displays that choice with the same visual weight it would give a publisher's own declaration. Nothing in the interface distinguishes a carefully verified value from a reasonable guess, and downstream users have no way to tell which they are looking at.
When a copy's stated license differs from the publisher's, or when the publisher stated nothing and the copy states something, the copy has not created permission. Deciding what to do in that situation is a legal question rather than a technical one. What you can do as an engineer is notice the discrepancy, record both sources with dates, and escalate before training rather than after. The pattern to watch for is a dataset whose own site is silent while its mirrors are confident.
A check you can run before training
Four steps, none of which requires a lawyer, all of which make the eventual legal review tractable.
Identify the exact artifact
Record the precise URL, the version or commit, the file set, and the date you downloaded it. "We used DROID" is not specific enough to answer a licensing question later.
Find the publisher's own terms
Go to the publisher's repository or official distribution rather than a mirror. If no terms are stated, record that as a finding rather than an absence of a problem.
Check the constituents separately
For any collection, enumerate the subsets you sampled and check each one's terms. This is the step that most often changes the answer in practice, and it is the step most often skipped.
Record what you found and when
Keep a short table of artifact, source URL, stated terms, and date checked. Terms change, pages move, and a screenshot with a date answers questions a later review cannot otherwise resolve.
This record is cheap at the time and unreconstructable later, which is the same property that makes collection metadata worth capturing during a session. A year after training, the question will not be what the license said; it will be what the license said when you downloaded it, and only a dated record answers that.
What to require when commissioning instead
Commissioning removes the ambiguity, which is a substantial part of its value and rarely counted. Teams comparing build against buy tend to weigh collection cost against download cost and treat the legal position as equivalent, when it is the dimension on which the two differ most. When you contract for collection you specify the rights in the agreement: commercial training, derivative model distribution, exclusivity, and retention. Weigh that against the cost of collecting, which is substantial once yield is accounted for, and against the work of specifying a capture campaign of your own. The same tradeoff appears in other modalities, where buying an existing set against commissioning one turns on rights as much as on fit, and it should inform how you read where teams are sourcing this data alongside what you would be training the policy on.
Commission a dataset whose terms you set
When the licensing answer is unclear and the model is headed for production, commissioned collection converts an open legal question into a contract term. HumanSignal Services runs collection with the rights specified up front, so commercial use, derivative distribution, and exclusivity are settled before the first session. Book a scoping conversation to work through what your deployment requires.
Does open source mean I can train a commercial model on it?
Not by itself, and the phrase usually describes access rather than permitted use. Open licenses vary widely in what they allow, some datasets are released for research use specifically, and collections can mix terms across their parts. Read the operative license text for the exact artifact you downloaded, and treat a summary in a paper or on a download page as a pointer rather than an answer. This is general information and not legal advice.
Where do I find the real license for a robotics dataset?
Start at the publisher's own repository or official distribution, which usually carries the operative text, then check the project page and paper for statements that qualify it. Mirrors should be checked last and treated as a claim by the uploader. If the publisher states nothing, record that finding explicitly rather than reading silence as permission.
Is a HuggingFace license tag authoritative?
It is a field completed by whoever uploaded that copy, so it reflects their understanding rather than the publisher's terms. It can be correct and it can also differ from what the original publisher stated, or state something where the publisher stated nothing. Use it as a lead, then verify against the publisher's own distribution.
What if a collection mixes licenses across its parts?
Then the terms that matter are the ones on the subsets you sampled, not the collection's headline license. Open X-Embodiment's repository directs users to per-dataset terms for exactly this reason. Enumerate the constituents you drew from and check them individually, because two teams using the same collection can end up in different positions.
What does CC-BY 4.0 require me to do?
Broadly it requires attribution to the original creator alongside other conditions set out in the license text, and the precise obligations for your use case should be read from that text rather than from a summary. How attribution applies to a trained model, as opposed to a redistributed dataset, is a question worth putting to counsel rather than settling from a blog post. Nothing here should be read as advice on whether a particular use complies.
What should I require from a vendor if I commission collection instead?
Written rights to use the data for commercial model training, to distribute models derived from it, and to retain the data for a stated period, plus clarity on exclusivity and on whether the vendor may resell the same data. Get those in the agreement rather than inferring them from a delivery note. The advantage of commissioning is precisely that these become negotiated terms rather than research questions.