A repository only a few days old, a code license stated plainly, 238 stars, and yet the model weights you came for are nowhere inside it. That is not a case of a project failing to open up; it is a case where the line marking how far it opened has to be drawn word by word through the README and the LICENSE. This piece reads only what can be checked inside the repository: the fields returned by the API, the exact wording of the README, and the full text of the LICENSE. Any figure that is not written down verbatim in the repository is not repeated here as fact. As of November 12, 2026, the conversation around LingBot-Touch already carries a fair number of very specific-sounding results, and separating what is in the repo from what is not is more useful than memorizing any of those retellings.
Pin Down the Repo: What Sits Behind 238 Stars
Start by getting the object right. Its exact path on GitHub is Robbyant/lingbot-touch, with the organization name capitalized as Robbyant, which is not the same as the all-lowercase identifier used on the model hosting platform; the two are separate registrations. The repository API returned 200 on both checks, and its fields are readable item by item: the primary language is Python, the default branch is main, it is neither a fork nor archived, it was created on November 9, 2026, its most recent push was on November 11, 2026, and its star count is 238. These are structured fields returned by the API, not paraphrases and not estimates.
The repository description reads "Learning Heterogeneous Tactile Representation for General Dexterous Manipulation", while the first line of the README is the title "LingBot-Touch: A Foundation Model for Heterogeneous Tactile Representation Learning". Both phrases land on heterogeneous tactile representation: the subject is tactile signals and the goal is a representation, not a model that generates text or images. The license field gives Apache-2.0, and the badge alt text reads "License: Apache-2.0" as well.
Where Heterogeneous Lands: Three Sensors, One Representation
The center of gravity here is neither the brand nor the star count but the word "heterogeneous" in the title. Sensing principles alone split into three kinds, listed verbatim in the README as magnetic, piezoresistive, and visuotactile: one reads changes in a magnetic field, one changes in the resistance of a piezoresistive material, one converts tactile information into a visual form. Their output shape, density, and units do not naturally line up, so the hard part is not chasing a high score on one sensor but landing unrelated signals in a single reusable representation.
The README is unusually specific about which point layouts pretraining supports: for magnetic sensing, 96, 76, 64, 48, 32, 24 points per finger; for piezoresistive sensing, 61, 16, 9, with "124 points for the thumb in the 61-point layout" called out on its own; and for visuotactile sensing, 200, 96, 32. The three groups span a wide range, from single digits to two hundred, and even one kind of sensor comes in several densities, which is exactly the point: a single representation has to accept inputs of very different sparsity. These are supported sensor layouts on the input side, a compatibility list, not accuracy, not a score, and not a benchmark result.
Along that representation the repository offers two model sizes, named LingBot-Touch-Tiny and LingBot-Touch-Small. All the names themselves establish is that a smaller and a larger option exist; figures such as parameter count, speed, and memory footprint are not written down verbatim in the README, so they should not be filled in on its behalf. The related paper is at arxiv.org/abs/2611.12806, and anyone who wants the method details should read that paper rather than reverse a conclusion out of a list of point layouts.
What Code Under Apache License 2.0 Covers, and What It Does Not
The license is the part of this piece that most needs to be read word by word. The LICENSE file in the repository returns 200 and runs 11,358 characters, the standard full text of the Apache License 2.0, neither a rewrite nor a summary. The matching sentence in the README reads "LingBot-Touch is licensed under the Apache License, Version 2.0. Applicable third-party components retain their original licenses", and it points to a separate third-party notices file. Read together, the boundary is clear: what can be confirmed is that the code is licensed under Apache 2.0, and at the same time the applicable third-party components keep their own original licenses, which are not flattened to match just because the main repository is Apache 2.0.
So the phrase "code under Apache 2.0" has to be taken literally. It covers the code in this source repository. It does not mean the entire project, model files included, is open under one and the same license, and it gives no basis for expanding that scope into a blanket claim about commercial use. On licensing, the only thing to write is the scope that has been verified verbatim; whatever has not been verified is left honestly blank. Existing single-project open-source pieces on the site treat the same boundary for image or video generation models: separate the source repository from the model files, then check each license.
What Is Not in the Repo: Where the Weights Went
Once "what is in the repo" is listed, the more important question is "what is not in the repo". The README contains one decisive sentence: "Model files are hosted separately from this source repository. Access requires an account with permission to the model repository." The first half says the model files live somewhere other than this source repository; the second half says using them requires an account with permission on that model repository. The weights sit elsewhere, behind a permission gate, and what this repository hands you is the code, the README, the LICENSE, and the third-party notices.
That boundary has to be held firmly, because it decides directly which claims can be written and which cannot. Since the weight files are not inside the source repository, their license status cannot be inferred from this repository's LICENSE. Since access requires permission, there is no ground for any blanket commercial-use claim on the weights side. The license on the weights side has not been verified, and the careful phrasing is to defer to whatever license text that weight repository eventually carries; this piece does not pre-decide it.
The corresponding page on the model hosting platform failed its handshake repeatedly during checking and never returned content, so fields such as download counts, likes, tags, and creation date have no verifiable reading and are neither written nor estimated here. Anyone who wants to check directly can open the plain address https://github.com/Robbyant/lingbot-touch, read those two sentences in the README, and compare them against the LICENSE full text.
A Deliberate Angle: Not the World Model, Not the Wuzhen Pieces
The site already carries several pieces about LingBot and a few more about the Lingbot brand, but the specific name LingBot-Touch has not appeared as a main subject before, and even the tactile angle is new. The distinction to draw first is this: this repository is a tactile foundation model doing a unified representation of heterogeneous tactile signals, whereas the world model piece from September 14, the related open-source piece from October 16, and the Wuzhen summit piece from November 4 discuss LingBot-World 2.0, LingBot-VLA, and summit news. They share a brand with this repository but are different products, different tasks, and different source material. The fact that every name contains LingBot is no reason to carry over progress on world models or the cadence of a summit launch and retell it as an achievement of the tactile model.
So this piece takes a deliberate angle rather than filling in a brand overview. The duplicate-check conclusion was likewise "needs an angle" rather than "discard": the main object had no dedicated piece before, and the focus on a unified tactile representation is a concrete technical thread that stands on its own, provided the boundary with those three earlier pieces is made clear in the body. Readers arriving for the world model or the summit have complete coverage in those pieces; this one handles only the tactile repository that was created a few days ago, and only the part of it that can be checked inside the repo.
Why the Numbers in Secondhand Roundups Cannot Be Copied
The retellings around this repository circulate a batch of hard-sounding numbers: training hours, frame counts in the hundreds of millions, success rates rising from one figure to another, resolution down to fractions of a millimeter, parameter counts in the millions, totals of real-robot tasks, gains in percentage points, and even a claim about commercial use. Check them item by item against the repository README and LICENSE and none of them has a corresponding statement in the primary material; their sources are only a handful of secondhand sites, and the repository itself gives neither these readings nor any original table from which they could be derived. Since this piece accepts only what can be checked inside the repo, those numbers cannot be quoted as primary facts, still less rewritten as "officially announced" or "measured in testing".
This mirrors the point layouts above: supported layouts are an input-side compatibility list, while training duration, success rates, and task counts are a different kind of quantity with no basis in the repo. The method is one line: write what the repository writes, and if the repository does not write it, leave it out of the body no matter how concrete the secondhand version sounds.
A Checklist You Can Verify Yourself
The above can be folded into a short checklist to verify item by item. The exact repository path is Robbyant/lingbot-touch, and the repo homepage opens from here. In the API fields: primary language Python, default branch main, star count 238, created on November 9, 2026, last pushed on November 11, 2026, license field Apache-2.0. The README title identifies a foundation model for heterogeneous tactile representation, and the three sensing principles listed verbatim are magnetic, piezoresistive, and visuotactile. The pretraining layouts come in three groups: magnetic at 96, 76, 64, 48, 32, 24; piezoresistive at 61, 16, 9 with 124 points for the thumb in the 61-point layout; and visuotactile at 200, 96, 32. The two model sizes are LingBot-Touch-Tiny and LingBot-Touch-Small, and the paper is at arxiv.org/abs/2611.12806.
On the license side, the LICENSE file is the standard full text of the Apache License 2.0, the code is licensed under Apache 2.0, and third-party components keep their original licenses. On the weights side, the README states that model files are hosted outside the source repository and that access requires an account with permission, with the weights license unverified. Every item above can be found verbatim in the repository; none of it comes from outside summaries or calculation.
FAQ
Q1: The code is under Apache 2.0, so is it free for any commercial use? A1: Only the code side can be confirmed. The LICENSE is the standard full text of the Apache License 2.0, and the code is licensed under Apache 2.0, while the README notes that applicable third-party components retain their own original licenses. The weight files are not inside the source repository and access requires an account with permission, so the weights license has not been verified.
Q2: Where are the model weights, and how do I get them? A2: The README states verbatim that "Model files are hosted separately from this source repository. Access requires an account with permission to the model repository." Following this repository yields the code, the README, the LICENSE, and the third-party notices, not the weights themselves. The corresponding page on the model hosting platform failed its handshake repeatedly during checking, so download counts, likes, and tags have no verifiable source and are neither written nor estimated here.
Q3: The three sensor kinds differ so much, how do they fit one representation? A3: That is exactly where the "heterogeneous" in the README title lands. The three sensing principles are magnetic, piezoresistive, and visuotactile, and their output signals do not line up in shape, density, or units; the pretraining layouts range from single digits to two hundred, which shows a single representation has to accept inputs of very different sparsity. Note that the layouts are input-side compatibility specifications, not accuracy, scores, or benchmark results.
Q4: How do I choose between LingBot-Touch-Tiny and LingBot-Touch-Small? A4: The repository only establishes that a smaller and a larger option exist. Figures such as parameter count, speed, and memory footprint are not written down verbatim in the README, so they should not be filled in on its behalf. For method details, follow the paper at arxiv.org/abs/2611.12806.
Q5: Can I quote the training hours, success rates, and parameter counts from secondhand roundups? A5: Not as primary facts. Checked item by item against the repository README and LICENSE, those numbers have no corresponding statement in the primary material and their sources are only a handful of secondhand sites. This piece accepts only what can be checked inside the repo: write what the repository writes, leave out what it does not, and never rewrite the gaps as officially announced or measured in testing.
Related reading
- How a single open-source project draws its license boundary: a different kind of generative model and how it separates the code license from the weights status.
- Another single-project verification method: organizing a teardown around what can be checked inside the repo.