Free is not yours: what the robot base-model licences actually say

On 11 September 2026, at 21:55 Singapore time, we deleted a sentence from our own thesis page. Commit 91f448e. It read:

"Chinese robots do have brains, and some of them are open. Unitree's UnifoLM, AgiBot's GO-1 and GO-2, Alibaba's RynnBrain and Qwen-Robot, Ant's LingBot-VLA 2.0, X Square's WALL-OSS and BAAI's RoboBrain are released with open weights."

Eight releases across six vendor families: Unitree, AgiBot, Alibaba, Ant/Robbyant, X Square, BAAI. RynnBrain and Qwen-Robot are both Alibaba Group, DAMO Academy and Tongyi Lab. Ant was spun out of Alibaba and is a separate company today, so Robbyant counts on its own; collapse the two and the number is five [inference].

Earlier that day we had read the licence of each out of the publisher's own weight repository, as bytes, not off a badge or a press release. Four of the eight did not support the sentence. Twenty minutes after the commit, between 14:15 and 14:35 UTC, we re-ran all thirteen checks against the four registry APIs named under the table, and every cell below is from that re-run.

The finding that survives all thirteen is one line long and anyone can check it in a minute. Of the twenty Hugging Face weight repositories this article names, two ship a LICENSE file, ten carry a license tag, and not one carries both.

That is the argument. The base layer is a commodity in price and a dependency in law: it costs nothing, and what you are permitted to do with it is mostly not written down beside the thing you downloaded. This piece ends in a table rather than a verdict, because two of the thirteen rows are questions we are not qualified to answer.

We are not a base-model lab and we have measured nothing on a machine. What follows is a reading of published documents by a company that has to read them. It is not legal advice, and no sentence here predicts what any licensor will do.

The Chinese column, one repository at a time

Start with the two everyone cites, because they are the two that cannot go into a commercial product.

unitreerobotics/UnifoLM-VLA-Base ships a real LICENSE: 20,930 bytes of Creative Commons Attribution-NonCommercial-ShareAlike 4.0. §2(a)(1) grants "a worldwide, royalty-free, non-sublicensable, non-exclusive, irrevocable license … for NonCommercial purposes only", and §1(k) defines NonCommercial as "not primarily intended for or directed towards commercial advantage or monetary compensation." The card points at the file. What the repository does not carry is a license tag — cardData is {"tags": ["robotics"]} — so a Hugging Face filter for non-commercially licensed models does not return it. Five of the seven models in the org ship that identical file, and not one of the five carries a tag.

The other two are this article's own clock. UnifoLM-ER-1 and UnifoLM-ER-Flow were published on 11 September carrying a LICENSE our check read as that same CC BY-NC-SA 4.0 file. At 14:40 and 14:41 UTC, five minutes after our audit window closed, both repositories' next commits read Delete LICENSE. Today each carries license: apache-2.0 in its card front matter and no file behind it. We report the two commits and nothing about why. Unitree is not careless about licensing. It is careless about metadata, and the metadata moves.

agibot-world/GO-1 is the inverse. The entire grant is one bullet of prose, card line 27: "- License: CC BY-NC-SA 4.0". No LICENSE file, no license field in cardData, no tag. To a licence filter, an SBOM scanner or a procurement check, this is a repository with no licence at all. It is also, on our reading, the most coherent position of the six families [inference]: OpenDriveLab/AgiBot-World states that "All the data and code within this repo are under [CC BY-NC-SA 4.0]" and carries the badge. Code, data and weights are non-commercial, and AgiBot says so in three places, just never in a file. agibot-world/GO-1-Air is a second non-commercial release on the same terms.

GO-2, announced 9 April 2026 [secondary], has no weight release we could find. The agibot-world org listing is six models and holds no GO-2; ModelScope returns nothing of AgiBot's under that name; GitHub search returns no official repository; OpenDriveLab/AgiBot-World's model list ends at GO-1 and GO-1 Air although it was pushed after the announcement. What would falsify it: a weight repository under any AgiBot-controlled namespace on any hub, or a GO-2 entry added to that README. What we did not search: an AgiBot download portal, a partner programme, a Chinese object store. And the discipline that makes such a negative worth anything: HTTP 401 is not evidence of absence. Hugging Face returns 401 for repositories that do not exist, verified against a nonsensical control, so every negative here rests on a complete org listing, full-text search across two registries, or a vendor's own words.

Then the line that made this article worth writing. Alibaba-DAMO-Academy/RynnBrain-8B, front matter line 6, unchanged since April:

license: apache-2.0 # TODO: 如果不是 Apache-2.0 请改

"TODO: if it is not Apache-2.0, please change." That comment is the only licence statement in the card. No licence section in the body, no LICENSE file in either the 8B or the 2B repository. Nor is it one stray note: the same front matter carries a second unreviewed comment, custom_code # 如果需要 trust_remote_code=True. A template was filled in and never cleaned. It is not an artefact of Hugging Face mirroring either — the identical front matter is published on ModelScope, on Alibaba's own hub.

Here the article has to correct itself. Our audit searched Hugging Face. On ModelScope, DAMO Academy publishes eleven RynnBrain models, including a whole 1.1 generation absent from Hugging Face, whose cards carry a clean license: apache-2.0 with no TODO. DAMO fixed the template. The TODO survives on the older 8B and 2B cards, which are the ones our thesis page named and the only ones on Hugging Face. A Hugging Face-only licence audit of Chinese robot models undercounts them, and ours did. The registry we reached for shaped the finding we got.

The second correction is owed to Alibaba outright. Our page listed Qwen-Robot among the open-weight releases; Alibaba never claimed it was one. QwenLM/Qwen-RobotManip and QwenLM/Qwen-RobotNav say so at line 23 of both READMEs:

"Note: There is currently no plan to release the model weights for Qwen-RobotManip or Qwen-RobotNav. We will continue adding report resources that can be publicly released to this repository."

Both hold a README and an assets folder: no code, no weights, no LICENSE. There is no Qwen-RobotWorld either. Our registry checks agree, in one clause: 465 models in the Qwen org, 41 in Alibaba-DAMO-Academy, none matching "robot". But the vendor's sentence is the finding, and it replaces our search. A bounded negative reads as suspicion of a company; the company's own sentence is a fact. The likeliest origin of our error is the coverage, which reported Qwen-RobotManip as being "trained on more than 38,100 hours of fully open-source data" [secondary] — a claim about training data, not weights [inference — our reading of our own error, and no source says it].

Ant's robbyant/lingbot-vla-v2-6b publishes six safetensors shards and a licence sentence: "This project is licensed under the [Apache-2.0 License](LICENSE)." The file behind that link returns HTTP 404. cardData is null, and it is not one repository: all 30 models in the robbyant org carry no license field, and robbyant/lingbot-vla-4b repeats the same prose and the same 404. Robbyant/lingbot-vla-v2 on GitHub is Apache-2.0 with a badge, so the code is properly licensed.

X Square's WALL-OSS is the total case. All four x-square-robot wall-oss weight repositories — wall-oss-flow, wall-oss-0.5, wall-oss-fast and wall-oss-flow-0.1 — open straight at # WALL-OSS, with no YAML front matter. cardData carries no licence, there is no LICENSE file, and the string "licen" does not occur anywhere in the card. ModelScope's licence field is empty too. X-Square-Robot/wall-x on GitHub carries a real Apache-2.0 LICENSE, and its README says nothing about the weights. One more fact, and a fact is all it is: the same Hugging Face org ships LICENSE files in four other repositories. The wall-oss repositories do not have one. We make no claim about why, and none about what X Square intends.

BAAI read as the simplest row at our first check on 11 September and did not at the second. All four RoboBrain repositories carry license: apache-2.0 in cardData, and all four are gated: auto: the prompt reads "You agree to not use the model to conduct experiments that cause harm to human subjects", the gate fields collect Company/Organization and Country, and the card returns 401 unauthenticated. The one unambiguously permissive row in this column is also access-controlled. You give your employer and your country and accept a use restriction Apache-2.0 does not contain before you can download. gated: auto grants access automatically on acceptance — an identification and acceptance gate, not a discretionary approval. Gated on the date we checked, and the phrasing is deliberate. The three RoboBrain2.0 repositories' metadata changed within thirty-six seconds of each other on 10 September, consistent with a batch change, but the commit history is behind the gate and RoboBrain-X0-Preview has been gated since 2025. The GitHub repository is Apache-2.0 and the ModelScope copy carries the same tag.

Two rows we will not answer

nvidia/GR00T-N1.7-3B is ungated, has 82,420 downloads, and ships a LICENSE file of 4,063 bytes. Open it. It is headed "NVIDIA License", and its §3.3 says:

"The Work and any derivative works thereof only may be used or intended for use non-commercially. Notwithstanding the foregoing, NVIDIA Corporation and its affiliates may use the Work and any derivative works commercially. As used herein, "non-commercially" means for research or evaluation purposes only."

Its grant, §2.1, is "a perpetual, worldwide, non-exclusive, royalty-free, copyright license", and the word "revocable" does not appear in that file at all.

The model card in the same repository says: "This model is ready for commercial/non-commercial use", and "This model is released under the [NVIDIA Open Model License Agreement]."

NVIDIA/Isaac-GR00T's README says, at line 70, "GR00T N1.7 is fully commercially licensable under Apache 2.0", and at lines 601–602, "- Code: Apache 2.0" and "- Model weights: [NVIDIA Open Model License]." It contradicts itself inside one document.

Three documents, three licences. The reading that is fair to NVIDIA, stated as a reading: GR00T N1.6's card links its licence directly to NVIDIA-OneWay-Noncommercial-License-22Mar2022.pdf, the file N1.7 and the whole lineage still ship, and N1.5's card says plainly that the model is ready for non-commercial use. N1.7 is the generation at which NVIDIA moved to commercial licensing; the card was rewritten in April 2026 and the LICENSE file was not replaced. On the evidence this looks like a stale file, and Isaac-GR00T's §License is NVIDIA's clear statement that the weights are under the Open Model License. That is a reading. We publish no verdict on which document governs the weights, and anyone putting a training budget behind this model should read all three rather than the badge.

The Open Model License itself, at its canonical URL, Last Modified 24 October 2025, grants "a perpetual, worldwide, non-exclusive, no-charge, royalty-free, revocable (as stated in Section 2.1) license". The parenthetical is part of the string; we have dropped it before. Three further clauses, quoted and not extrapolated from. Circumvent a guardrail without substituting a similar one and "your rights under this Agreement will automatically terminate." Use must be "consistent with NVIDIA's Trustworthy AI terms", incorporated by reference from a second page. And: "NVIDIA may update this Agreement to comply with legal and regulatory requirements at any time and You agree to either comply with any updated license or cease Your copying, use, and distribution of the Model and any Derivative Model." Those are words in a licence, and that is all we say about them.

On the same document, in NVIDIA's favour: the preamble says the models "are intended to be used permissively and enable the further development of AI technologies" and that "Models are commercially usable", and the definition of "Model" reaches "checkpoints, learnt weights, algorithms, parameters" by name, which is less common than it should be. One structural fact belongs in the taxonomy rather than in any verdict: GR00T's vision-language backbone nvidia/Cosmos-Reason2-2B is gated behind acceptance of the Open Model License, and every GR00T checkpoint loads it on first use, so the ungated N1.7 download cannot be executed without accepting that licence on a different repository.

GR00T N2's licence could not be established. There is no weight release — the 934-model nvidia org listing contains no N2, search returns one unrelated third-party repository, and Isaac-GR00T's README does not mention it — and the 16 March 2026 press release says only that it is "slated to be available by the end of the year" [secondary].

The other unanswered row is openpi. Physical-Intelligence/openpi ships two licence files at its root: LICENSE, 11,356 bytes, and LICENSE_GEMMA.txt, 8,431 bytes. A recursive walk of all 177 entries in the tree finds those two and no third. GitHub's metadata reports the repository as Apache-2.0, because its detector reads the file named LICENSE and does not see the one beside it. Three facts bound the question and none resolves it. The README says nothing: "licen", "gemma" and "copyright" each occur zero times in it. The weights are not in the repository that holds the licence files — π0, π0-FAST and π0.5 download from a Google Cloud Storage bucket. And that bucket holds no licence document: the GCS JSON API lists the checkpoint prefix as two folders and no objects, and /openpi-assets/LICENSE returns 404.

What LICENSE_GEMMA.txt says, if it governs, is not small. §3.2: Google "reserves the right to restrict (remotely or otherwise) usage of any of the Gemma Services that Google reasonably believes are in violation of this Agreement." §3.1 requires you to carry the §3.2 restrictions down "as an enforceable provision in any agreement … governing the use and/or distribution of Gemma or Model Derivatives". For a robot vendor, that is a clause in a customer contract. §4.5: on termination "you must delete and cease use and Distribution of all copies of Gemma and Model Derivatives in your possession or control."

We do not say the Gemma terms govern openpi's weights. We do not say they do not. No probability, no "most likely", no recommendation: this needs a lawyer, not a researcher, and the useful public act is to show that the question exists and the repository does not answer it. π0.7, published 16 April 2026, has no weight release at all — the physical-intelligence org contains one model, the FAST tokenizer, and openpi's checkpoint table lists π0, π0-FAST and π0.5 only. Our Brain comparison already records π0.7 as a partner programme rather than something you self-host.

Then the contrast, from the vendor of the most contested row: nvidia/Cosmos3-Edge. OpenMDW 1.1, ungated, card says ready for commercial and non-commercial use, licence under 3,000 characters of operative text. "Model Materials" expressly includes "one or more machine learning models (including architecture and parameters)"; permission is granted "to deal in the Model Materials without restriction"; the words "revocable", "irrevocable" and "perpetual" appear nowhere in it; the only termination trigger is patent or copyright litigation retaliation. No use restrictions, no ethics terms by reference, no unilateral amendment clause. NVIDIA publishes both the least ambiguous grant in this table and the most contested one, and the least ambiguous one has twenty times the downloads — 1,632,790 against 82,420, relative magnitudes on a 30-day rolling measure and not adoption claims.

The table

Every licence string above was read as bytes from the publisher's own repository between 14:15 and 14:35 UTC on 11 September 2026, through the Hugging Face model API and raw file fetch, the GitHub REST API, the ModelScope model API and the Google Cloud Storage JSON API. The audit covered three hubs: Hugging Face, ModelScope and GitHub. Read the cells as written: not stated is a different mark from no, and a row can be silent about permanence without being revocable. Some rows cover more than one repository. The commercial-use column reports what the published licence permits, not whether a vendor would grant other terms on request — we wrote to none of them. Counted as Hugging Face repositories rather than as rows, this article names twenty that publish weights. Two of the twenty ship a LICENSE file: unitreerobotics/UnifoLM-VLA-Base and nvidia/GR00T-N1.7-3B. Ten carry a license tag: RynnBrain-8B and -2B, all four RoboBrain repositories, nvidia/Cosmos3-Edge, the gated backbone nvidia/Cosmos-Reason2-2B, and the two UnifoLM-ER repositories described above. None carries both. Every cell was re-read on 2026-09-12 and not one had changed; the only change that re-read found is the Unitree one described above, in two repositories that are not in this table. The twenty-repository count was run again on 2026-09-14 and holds. Licence estates move — two Unitree repositories deleted their licence file five minutes after our window closed — so treat every cell as of its date and read the file again before you rely on it.

"Open" names four different things

Our taxonomy, marked as ours [inference]. The word carries at least four separate claims, and the table is what they look like when they come apart: the weights are downloadable; the code is permissively licensed; the weights are permissively licensed, said about the weights specifically; and the grant is irrevocable. Most of these releases satisfy one or two. Our retired sentence named eight as though they satisfied all four, which is how this article started.

The fourth is where we were wrong ourselves, and it does not take two values. It takes three. Explicitly irrevocable: CC BY-NC-SA 4.0 §2(a)(1) uses the word, so UnifoLM and GO-1 carry it. Explicitly revocable: the NVIDIA Open Model License uses the word, in the grant. And silent: Apache-2.0, OpenMDW 1.1, and every repository that states no licence at all, which is most of the table. Silent is its own mark. Drawing it as "no" would be the same error as drawing "not stated" as "no", one column to the left.

Which produces the inversion. The two releases that may not be used commercially are the two carrying the word irrevocable. The one with a commercial-use clause on its card is the one carrying the word revocable. Permission and permanence are traded against each other in the documents, whatever anyone intended by it.

On a strict reading, no release affirms all four, and nvidia/Cosmos3-Edge is the only one that fails none: weights downloadable and ungated, OpenMDW 1.1 reaching "architecture and parameters" by name and granting use "without restriction", commercial use stated on the card, and the licence silent on irrevocability rather than affirming it. That is the honest form of the answer, and the one we publish.

Checked, no weight release found

AgiBot GO-2no weight repository on any of the three hubs we searched

Qwen-RobotManip, Qwen-RobotNavthe vendor states there is no plan to release them

NVIDIA GR00T N2no weight release found, and the licence could not be established

Physical Intelligence π0.7no weight release found; our own Brain comparison records it as a partner programme

Outside every region, because the first claim is never made: four of the thirteen checks found no weight release at all, one of them on the vendor's own statement and three on the bounds of our search.

1 · Downloadable — the weights are published and fetchable

UnifoLM-VLA-Basea real LICENSE, non-commercial; the grant carries the word irrevocable

GO-1, GO-1-Airnon-commercial in card prose; the code repository is non-commercial too, which keeps it out of the next region

openpi — π0, π0-FAST, π0.5two licence files at the root, a silent README, weights in a bucket holding no licence document; no verdict

2 · Permissive licence on the code

wall-oss repositoriesthe weights state no grant of any kind — the string licen does not occur in the card

RynnBrain-8B, -2Bthe only statement on the weights is a tag carrying an unresolved TODO

LingBot-VLA v2-6b, 4bcard prose names Apache-2.0 and links a LICENSE that returns 404

GR00T N1.7three documents naming three licences; we publish no verdict on which reaches the weights

3 · Permissive licence stated about the weights specifically

RynnBrain 1.1 generationa clean apache-2.0 tag with no TODO, on the hub our own audit did not search

RoboBrain, four repositoriesapache-2.0 on the tag, behind an identification and acceptance gate on the date we checked

Cosmos 3 EdgeOpenMDW 1.1 reaches architecture and parameters by name — and is silent on permanence, not affirming it

4 · Irrevocable — the grant cannot be withdrawn

Empty. No release in the table affirms all four, and this is the inversion the documents make: the word irrevocable appears in the two licences that bar commercial use, and the word revocable in the one whose card offers it.

Three rows state it about the weights themselves. None says whether the grant can be withdrawn.

The code repository is Apache-2.0 in every row here. None gets past the next boundary, because none makes a settled permissive statement about the weights.

Reaching this boundary and no further: the weights are there, and the grant beside them is non-commercial or unresolved.

• One sense of "open". Each boundary is a stronger claim than the one outside it.
• A release we checked, at the innermost boundary its own documents put it past.
• Checked, but no weights published, so no region applies.

What it does to a strategy built on it

The pitch an investor has heard five times is that the robot brain has commoditised, so the defensible layer must be somewhere else. Half of that is right, and it is the cheap half. The base layer is free. What it is not is yours.

None of the thirteen rows is an accusation. In two of them the prose is clearer than the metadata: Ant's card states Apache-2.0 in words while no model in the 30-model robbyant org carries a license field, and X Square's code repository ships a real Apache-2.0 LICENSE while the wall-oss weight repositories say nothing about a licence at all. Our own reading of the pattern, and it is ours: most of what the table records is metadata debt rather than intent [inference]. But someone asking what a robotics company owns should notice that the answer is not the base model, and that it is not made more yours by nobody having charged you for it.

Two rows stay open, which is the discipline the rest of the piece rests on. We also wrote to no vendor. We looked for a public statement from Unitree, AgiBot and X Square about commercial licensing and found none, and absence of a public statement is not absence of a programme: a commercial licence may well be available on request from any of them, and a buyer who needs one should ask rather than read our table as a closed door.

The check cost an afternoon. It corrected our thesis page, then it corrected our method, and the second correction was the more useful one: we had searched one registry and mistaken the shape of our instrument for the shape of the field. If you are about to build on one of these models, the part worth copying is not our conclusion. It is the order of operations. Open the weight repository, list its files, read the one named LICENSE and the one beside it, check whether the card and the file agree, and write down the date you did it.