Short answer
A public dataset is not automatically free for commercial use. Check the license text, any access agreement, the terms of each underlying source, and whether trained weights count as a derivative. Published examples include EPIC-KITCHENS under CC BY-NC 4.0 and Ego4D behind a license agreement, and an audit of text datasets found license labels often more permissive than the author's. This is not legal advice.
Why this matters for licensed demand
Public robot and egocentric datasets are the obvious first stop for a humanoid program, and many were released for research. When a commercial team wants to ship a model trained on one, the license is a gating item. A restriction found late can force a change of data source, which is a reason to read the terms early. Terms differ per dataset and per version; the examples below show kinds of checks, not rulings. This is not legal advice; have counsel review the text that applies to you.
What the sources show
- Access agreement. Ego4D requires users to review and accept a license agreement before they can access data or annotations, executed as an individual or on behalf of an institution; the visualization tool requires an approved license too (Ego4D consortium, n.d.). An access agreement is a contract, so who signs, and whether it covers a company, matters.
- NonCommercial license. EPIC-KITCHENS video files are published under CC BY-NC 4.0, meaning attribution, a link to the license, indication of changes, and no commercial use; the VISOR extension is also CC BY-NC 4.0, with commercial licenses negotiated through the University of Bristol (EPIC-KITCHENS team, n.d.).
- What NonCommercial means. Creative Commons describes NonCommercial as not primarily intended for or directed towards commercial advantage or monetary compensation, judged by the use rather than the user. NoDerivatives restricts public sharing of adaptations, and licenses with NC or ND terms are considered non-free (Creative Commons, n.d.).
- Labels can be wrong. An audit of more than 1,800 text datasets found that, after re-annotation, 66% of analyzed Hugging Face licenses sat in a different use category from the author's license, often labeled more permissive (Longpre et al., 2023). That audit covers text, not robot data, but it shows why a repository tag is not enough.
Illustrative license-audit checklist
| Check | Pass condition (buyer sets) | Evidence to keep |
|---|---|---|
| Original license text | Read from the dataset's own site, not a mirror | Dated copy of the text |
| Commercial use | Explicitly permitted for your use, or a commercial license obtained | License or written grant |
| Access agreement | Signed by an entity that can bind your company | Executed agreement |
| Derivatives | Treatment of trained weights, outputs and annotations understood | Counsel's note |
| Attribution and share-alike | Obligations can be met in your release process | Compliance plan |
| Underlying sources | Video, images or scenes in the set carry no extra restrictions | Source list from the authors |
| People in the data | Consent basis covers your use | Consent description |
| Version | Terms for the exact version you downloaded | Version and checksum |
| Mirrors and derivatives | Copies on other hubs carry the original terms | Provenance record |
| Change clause | What happens if terms change or the host removes data | Contract term |
For web-scale video collections, see public instructional video vs licensed footage. For the general dataset review, use the dataset evaluation checklist; for what happens to trained weights, see model weights, derivatives and redistribution.
Red flags
- The only license information is a repository tag.
- "Research use" appears without a statement about commercial use.
- The dataset is a mix of sources, and only the compilation has a license.
- Downloaded through a mirror with no upstream terms.
- Annotations are licensed differently from the media.
- Nobody can say who at your company accepted the access terms.
What a permitted public dataset still cannot give you
Even with clean terms, a public set is shared with everyone, was built for its authors' embodiment and tasks, and says nothing about your sites. It will not hold your robot's action space or your business's tasks. Terms may allow use for evaluation but not training, or the reverse; check separately. See exclusivity and field of use.
Questions to ask before you rely on one
- Which exact license applies to the media, the annotations and the metadata?
- Does it permit training a commercial model, and distributing the weights?
- Who accepted the access agreement, and for which entity?
- Are there upstream restrictions from the source recordings?
- Can the authors or host grant a commercial license, and on what process?
- Have we archived the terms as they stood when we downloaded?
For the contract side of licensed alternatives, see licensing terms for robotics training data, provenance and chain of custody, and data license terms explained.
Combining public data with licensed data
Where a public dataset's terms fit, it can serve as a baseline or for evaluation-only comparison, while licensed data covers the commercial training set. Keep a register that records, per dataset, the version, license, the entity that accepted terms, and which models consumed it. That register makes later audits, customer questionnaires and retention decisions manageable; see retention, deletion and audit rights.
Also record what you did to the data. Filtering, relabeling or converting formats can create adapted material, which some licenses treat differently. The CC description of NoDerivatives, for example, concerns sharing adaptations publicly, so whether your internal use is affected depends on the license text and your plans. Ask counsel rather than inferring.
Explore your options
If a public dataset's terms rule it out, describe what you need in a request to SourceX. SourceX reviews the requirement and follows up about potential sources, checks suppliers' licensing permissions and agrees allowed uses in a license; nothing is contracted until a supplier agrees, and a request does not guarantee a match. See the procurement hub and the humanoid data guide.
Explore sourcing and licensing options with SourceX.
Sources
- Ego4D Start Here — license agreement and access process.
- EPIC-KITCHENS — CC BY-NC 4.0 terms for video and VISOR.
- The Data Provenance Initiative — audit of license labels.
- About CC Licenses — NonCommercial and NoDerivatives definitions.