How to License Pretraining Data for Commercial Robot Models

It depends entirely on the specific dataset’s license text, not on the dataset’s scale or quality. A large, well-labeled egocentric video corpus can still be legally unusable for a commercial model if its license is research-only or non-commercial. The published pretraining corpora referenced elsewhere on this site — EgoScale, GEN-0, EgoVerse, Egocentric-10K — were each released under whatever terms their publishing lab chose, and those terms vary by dataset. Egocentric-10K, for example, is released under an Apache 2.0 license according to its official Hugging Face dataset page, which is a permissive commercial-compatible license; other corpora may carry different terms. The only reliable way to know is to read the specific dataset’s current license text before planning around it.

This page is general information about how dataset licensing typically works for pretraining corpora, not legal advice. Confirm the current license text of any specific dataset directly, and consult counsel before relying on it for a commercial roadmap.

A useful habit is to treat "is this dataset good?" and "can I actually license this dataset for my model?" as two separate questions that must each be answered independently before a pretraining plan is finalized. A corpus can score well on both counts and still fail the licensing question, or the reverse: a smaller, clearly commercially licensed corpus can be the safer foundation-layer choice even if a larger research-only corpus looks more attractive on paper. Teams that plan their pretraining mixture around scale and diversity alone — without checking licensing at the same stage — tend to discover the restriction only once a model is close to shipping, which is the most expensive point to discover it.

Four terms decide whether a pretraining dataset is usable in a commercial model: whether the grant is commercial or research-only, whether it is perpetual or revocable, whether redistribution of derivatives (including a trained model) is permitted, and who ultimately owns the resulting model. Pretraining specifically raises the derivative-work question early, because pretraining is usually the first and largest training stage — if the base layer of a data pyramid is built on a research-only corpus, that restriction can propagate through mid-training and post-training even though those later stages use different, properly licensed data.

A research-only license permits academic study and benchmarking but does not grant the right to train a model that ships in a commercial product. A commercial license explicitly extends that permission. Whether a trained model itself counts as a "derivative work" of its training data is a separate and unsettled legal question that varies by jurisdiction — which is exactly why pretraining on a research-only or non-commercial corpus carries risk that tends to surface only after a model is already deployed, not before. See /guides/research-only-vs-commercial-dataset-licenses for the fuller breakdown of how these license families are typically structured.

This distinction is unrelated to a dataset’s technical merit. A published pretraining corpus can be exactly the right scale and diversity for a team’s target embodiment and still be a poor fit for a commercial pretraining mixture if the license does not extend to commercial deployment. Because pretraining sits at the base of a data pyramid and typically consumes the largest share of hours in a training plan, a licensing mistake made at that layer is harder to unwind later than the same mistake made in a smaller mid-training or post-training slice.

A commercially licensed custom collection is built specifically to avoid the research-only problem at its source, rather than requiring a buyer to audit someone else’s public dataset after the fact. On Firsthand, that means a documented consent chain — a signed release per participant recorded before capture, with a consent artefact ID per episode listed in the data card — paired with a perpetual, buyer-owned commercial license rather than a revocable subscription. The buyer owns the resulting rights outright: the license cannot later expire or be renegotiated by the provider, and it explicitly covers pretraining-scale use from the start.

Consent and license answer two different questions, but a weak consent chain undermines a license even when the license text itself looks commercial. Consent covers what the people in the footage agreed to; the license covers what a buyer can legally do with the footage. A commercial license backed by an undocumented or assumed consent basis is a liability that surfaces after deployment, when a participant or regulator challenges whether the underlying capture was ever properly authorized. Irreversible pixel-level anonymization of bystanders who did not consent is part of the same chain: it is what keeps a dataset commercially usable without depending on every incidental bystander’s individual consent.

• Does the license explicitly permit commercial use, or is it research-only?
• Is the grant perpetual, or can it be revoked or expire?
• Is consent documented per participant, with a traceable artefact per episode?
• Who owns the resulting trained model — the buyer, or is it licensed back from the provider?
• Is bystander footage irreversibly de-identified at the pixel level?
• Does a data card record the jurisdictions of capture and the applicable license terms?

• Confirm the license text explicitly grants commercial rights, not just research or benchmarking use
• Confirm the grant is perpetual and cannot be revoked after training has already used the data
• Confirm consent is documented per participant with a traceable artefact, not just asserted
• Confirm the resulting model’s ownership is clear — owned by the buyer, or licensed back under specific terms
• Confirm bystanders are de-identified irreversibly at the pixel level
• If any of the above cannot be confirmed for a public corpus, treat it as unusable for the commercial layer of the pretraining mix

This checklist reflects general licensing considerations, not legal advice. Confirm the current terms of any specific dataset and consult counsel before finalizing a commercial pretraining plan.

It is worth running this checklist against every layer of a planned pretraining mixture, not just the largest public corpus in the plan. A mixture typically blends web data, egocentric human video, synthetic data, and real-robot data, and each of those sources can carry different licensing terms. A mixture that is diverse in scenes, tasks, and embodiments but licensed inconsistently across its layers creates the same downstream risk as a single research-only corpus — the weakest link in the licensing chain determines what the finished model can legally do, regardless of how strong the other layers are.

Teams that need a pretraining-scale corpus cleared for commercial use from the start, rather than auditing public datasets after the fact, can request a custom pretraining collection built with a documented consent chain and a perpetual, buyer-owned commercial license from day one. See /custom-collection/pretraining-data-collection for how the collection is scoped, and /guides/robot-pretraining-data for how pretraining data fits into the broader training pipeline.

A custom collection also lets the licensing terms be negotiated and settled before capture begins, rather than accepted as fixed the way a public dataset’s terms are. That means the consent chain, the commercial grant, and the data card format can all be specified up front as part of the collection brief, so the resulting corpus is licensable for the intended commercial use by construction rather than by later audit.