Executive Summary
Recordings from three kinds of touch sensor, each built on a different physical principle, now read as the same language inside one model. The model is LingBot-Touch, a tactile foundation model released on November 10, 2026 by Robbyant, the company inside Ant Group that builds the AI which moves robots. Its raw material is 4,983 hours gathered from piezoresistive sensors, whose resistance changes under pressure, from magnetic sensors, which read how far a magnet has been displaced, and from vision-based sensors, which film a gel from behind with a camera. The method was not to grow the model. It threw away the taxel index each sensor assigns on its own, wrote down instead where that taxel sits on the fingertip as a coordinate, and put all of them in one space. This article follows what that method achieves and what it gives up, through the full paper and the public repository.
The results split in two. Attached with no parameter changed to a fourth sensor that had never entered pretraining, the model located the point of contact within 0.16 mm on average. That is a figure an order of magnitude tighter than a comparison model trained with an adapter built for that very sensor. Laid on top of four existing policies, its tokens raised the average score across twenty-six real-robot tasks for all four, while the setting that fed the same signal in unprocessed delivered a third to a half of that, depending on the policy.
The paper also wrote out the bill. When the hours you can spend are capped, a model trained on a single sensor's data wins seven of eight pairs on its own sensor. Force and dynamics carry across devices, but spatial precision comes only from seeing your own layout at your own resolution. Pooling is therefore not a capability but an exchange. The attempt to hit the same wall from the opposite side is also still in progress. A proposal to add a tactile sensor module to LeRobot, the de facto container for robot learning data, opened in August and has stayed open for three months, and what it snagged on is the same thing: a 9-taxel pad and a 239-taxel array cannot be expressed in one feature. The question left at an operating site is not whether tactile data was collected. It is whether your records are in a format that travels when the device changes, and whether pooling gains more than it gives up.
0.16 mm
Contact-center error on a sensor absent from pretraining
FTP-1, trained with an adapter built for that same sensor, was at 5.89 mm. Not one parameter was changed
+23.2pp
Change in the policy T-Rex, averaged over 26 tasks
From 45.4 to 68.6. Feeding the same signal in unprocessed stops at +11.2pp
7 / 8
Sensor-task pairs won by the single-sensor model at the same 1,000-hour budget
The place where the paper measures directly how much spatial precision pooling gives up
22 / 104
Task-policy pairs that got worse once the raw signal was added
With the same signal converted into tokens, 99 of the 104 pairs were at or above the original
The figures in this article come from the body, tables and appendix of arXiv:2611.12806 v1, registered on November 9, 2026, and are current as of November 17, 2026. We reopened the abs page that day and confirmed that v1 is the only version.
One Contact, a Different Set of Numbers on Every Sensor
Press a fingertip once against the edge of a desk. Depending on which device that single contact lands on, it survives as an entirely different bundle of numbers. On a piezoresistive grid the resistance drops in each pressed cell. On a magnetic sensor the displacement of a magnet is written along three axes. On a vision-based sensor with a camera behind a gel, the distorted gel is simply a photograph. The channel count differs, the layout differs, the resolution differs, and so does how many times per second it is read. The paper lists this as the first of three gaps.
"There is no common input format: each sensor type defines its own channels, layout, resolution, and sampling rate, so a reading from one sensor means nothing to a model trained on another."
§1, the first gap. Everything downstream follows from this one sentence.
Once the format splits, the corpus cannot pile up either. One sensor's recording will not add to the next one's, so hours scatter by device, and buying one more hour means a hand wearing that sensor has to go out and touch the world. The paper names the root of all three gaps in a single sentence, then immediately writes the way out alongside it.
"The three gaps share a root: tactile data are tied to the physics of the sensor that produced them. Contact itself is not. A fingertip pressing on an edge is the same event whether a magnetometer, a resistive grid, or a camera behind a gel records it."
§1, where three separate problems are folded into one and the exit appears in the same breath.
How far apart "different formats" actually pulls is visible in a single composition table for the pretraining corpus. Inside one corpus, a piezoresistive glove puts nine taxels on one finger and reads them 100 times a second. A robot fingertip carrying a vision-based sensor photographs that same single finger as a 240×240 image 30 times a second. The two recordings share neither order of magnitude, nor unit, nor physical quantity.
| Principle | Collected on | Hours | Episodes | Taxels per finger | Rate |
|---|---|---|---|---|---|
| Piezoresistive | Human-worn tactile glove | 2,197 | 86,906 | 9 / 16 / 124 thumb, 61 others | 100 / 60 / 60 Hz |
| Magnetic | Human-worn tactile glove | 1,136 | 294,871 | 116 thumb / 96 others | 41.7–83.3 Hz |
| Vision-based | Bimanual robot fingertips, teleoperated | 1,650 | 82,654 | 240×240 image resolution | 30 Hz |
| Total | 4,983 | 464,431 | about 1.33 billion frames |
Composition of the pretraining corpus by sensor (Tab. 4). The three numbers in the piezoresistive row mean that three glove types each use a different taxel count and rate. The 240×240 under the vision-based sensor is an image resolution, not a taxel count. That the two columns cannot be compared in one unit is the point of this section.
1.1The Same Wall Stands Outside the Paper
The problem written down as "there is no common input format" is one LeRobot has still not closed. A proposal to add a tactile sensor module shaped like the existing camera module (PR #4469) opened on August 18, 2026, and as of November 17 it remains open. The proposal defines four data types: displacement and force as a 3D point set, optical sensors as 6D, and force-torque as a wrench. What the discussion snagged on was not an algorithm.
"a dataset that records a 9-taxel pad and a 239-taxel array cannot express both in one feature."
From the review thread on LeRobot PR #4469. That thread also notes that the two tactile sensors documented so far already span 9 to 239 taxels.
The overlap in those numbers is not a coincidence. The piezoresistive glove with nine taxels per finger in the table above is the same class of pad the LeRobot thread brings up. One side has been stuck for three months trying to make the bytes agree; the other left the bytes untouched and had them meet inside the model instead. Two parties, hitting one wall from opposite directions.
1.2Where Our Earlier Articles Stopped
Pebblous has covered tactile data four times. The Tactile Data Humanoid Capital Left Behind asked why it fails to accumulate, looking at the hole left when money flowed only toward video. This article's question sits one step further on. It accumulated, so why will it not pool?
In Why Is Robot Learning Data Still Scarce When Video Is Everywhere? we wrote that the bottleneck sits not in the policy but in the translation machinery, and the four pieces named there were auto-labeling, retargeting, physics-grounded 3D inference and success judgement. All four stand on the video and action axes. Touch was not on that list. LingBot-Touch amounts to a fifth translation machine bolted onto it.
The 4,983 Hours Came from Gloves, Not Robot Hands
Two thirds of the 4,983 hours came from tactile gloves worn on human hands, and only the remaining third is teleoperated recording on the robot hands that actually get deployed. From the point of view of collecting tactile data, that is the wrong way round. Instead of collecting on the best-performing sensor, the project bought hours on gear a person can wear all day while washing dishes and opening drawers. The paper defends the choice in one sentence.
"the sensor to scale with is not the most capable one but the one at the sweet spot between capability and ease of capture, which is why two thirds of our corpus come from gloves a person can wear all day rather than from the fingertips we later deploy on."
§3.1, the design argument for collecting on gloves rather than on the hand the model is later deployed on.
What a glove buys also includes breadth. No robot has to be set up, so one person can record for several hours a day in ordinary homes and offices, and the first-person recordings that pile up carry objects and motions that would take a long time to script and reproduce on a robot. Breadth is what was bought by taking the capability down one notch.
A glove and a robot hand differ in shape and in degrees of freedom. What was aligned in order to put both into one corpus was behavior, not hardware. Every episode is classified into one of 24 task categories such as opening and closing, assembly, wiping, pouring and tool use, and most categories are covered by more than one source. The paper puts it this way: "The two sources are matched on behavior, not on hardware." Instead of matching the gear, they matched what the gear was doing.
Labels are almost absent. Each episode carries only which task category it belongs to. Nothing records how many newtons of force there were, where the contact happened, or when an object slipped. This is where the requirement in §3 of this article, that pretraining has to run without labels, comes from.
It is also worth knowing in advance that the word "episode" points at a different length for each sensor inside this one corpus. Hours divided by episode count put the piezoresistive glove at an average of 91 seconds per episode and the vision-based fingertip at 72, while the magnetic glove sits at 14. Hence the magnetic glove's odd position: the fewest hours, yet the most episodes at 294,871. Any average that mixes the three sources carries this difference in length along with it.
The sampling ratio does not follow the collection ratio either. When clips are drawn during pretraining, the piezoresistive, magnetic and vision-based sources are given equal probability. The hours inside the corpus tilt 2,197 to 1,136 to 1,650, while the amount training sees is flat. In a pooled corpus, how much was collected and how much reaches the model are settled separately.
2.1What Is Released and What Is Not
As of November 17, 2026 the repository is open and the license is Apache 2.0. What you can take is the inference code, the tooling that extracts tokens from a frozen encoder, checkpoints in the two sizes Tiny and Small, the per-sensor data interface code, and 120 sample recording files. The samples are 30 bimanual pairs each from two platforms.
The 4,983 hours used for pretraining cannot themselves be downloaded. What has been released is not the corpus but the encoder trained on it, together with the method for attaching that encoder to your own sensor. This is a release for application, not for reproduction. If you come across a summary saying a 5,000-hour dataset has been released, that summary is wrong.
A Coordinate on the Fingertip Instead of a Taxel Index
To unify its input, LingBot-Touch threw away the taxel index. Taxel 3 on one glove has nothing to do with taxel 3 on another. The index is a name tag the manufacturer assigned in wiring order. In its place the model writes where that taxel sits on the fingertip surface as a continuous 3D coordinate, together with which finger it belongs to. Start from the geometry asset each sensor ships with, map it onto a unit-radius sphere, and the taxels of all three device families become points in one space.
"we describe where a measurement was taken by a continuous 3D coordinate and a finger identity, and place the coordinates of every sensor in one canonical fingertip space."
§3.2, the sentence that swaps a manufacturer's index for geometry.
Matching the places alone does not make the input one thing. Under identical force, the number a resistive grid reports and the distance a magnet is displaced still start from different units. So magnitude gets a second treatment. Each sensor's value is divided by its own response range, which moves the normal-direction value into zero to one and the signed tangential value into minus one to one. That unitless value and the coordinate before it together form the common input specification.
After that it is the model's job. One small two-layer network per sensor type turns its own signal into tokens, and from there all three run through a single shared encoder. Sensor-specific parameters survive only in the shallow layers at the edge of the network.
A redrawing of the design in §3.2 and §3.3. The only parts that differ by sensor are the tokenizer at the inlet and the predictor interface at the outlet; the body in the middle is single. The taxel counts in the drawing are one example device per family from the corpus.
One token is the six numbers a single taxel produced over two consecutive frames. Each token carries a five-axis address made of time, finger and the 3D coordinate, attached through rotary positional encoding. What matters about this scheme is that the token count no longer has to be fixed in advance. Nine taxels or 116, they enter the same backbone unchanged. The paper names exactly that as the basis for transfer.
"This positional encoding does not require a fixed token count or a sensor-specific positional table, allowing a shared backbone to process different sensor layouts and supporting transfer to unseen sensors."
§3.2, on why an unfamiliar layout needs no new lookup table.
3.1Why the Prediction Target Is a Latent
With no labels, the training objective has to be manufactured. The common approach is to mask part of the input and have the original signal restored, and that approach does not work well on tactile recordings. Nearly every cell in a clip sits close to zero. A model that is good at predicting zero is good at learning the background, not at learning contact. So LingBot-Touch predicts the latent representation of the masked region rather than its signal. It is the joint-embedding predictive architecture carried over to touch, and the same family of architecture and its design logic were covered in Did Yann LeCun's Small Model Use Less Data, or Pair It Better?, so they are not unpacked again here.
The gap against reconstruction was measured directly under matched conditions. With model size and data held equal and 1,000 hours of training, contact-center error was 0.90 mm against 3.03 mm for a VAE and 3.07 mm for an MAE, and recall for finding the matching video from touch alone was 32.1 percent against 4.4 and 5.6. The more striking part is what happened when the data grew. With more training hours, the VAE got worse on four of five metrics, contact-center error widening from 3.03 mm to 4.08 mm. Seeing more background does not teach you more about contact.
3.2Full Sharing Was Not the Answer
Pushed to its end, the word "unified" ties the inlet and the outlet into one as well. That setting was actually run, and the result went the other way. A model with fully shared tokenizers and predictor interfaces lost on all five metrics. Grouping them by modality and routing accordingly brought contact-center error down from 2.32 mm to 0.70 mm.
"Grouped tokenizers with group-routed predictor interfaces beat fully shared interfaces on all five metrics, cutting contact-center error from 2.32 to 0.70 mm; the sensor-specific boundary layers of Sec. 3.2 earn their place." (§4.5) The answer to pooling was not to make everything the same but to keep only the boundary per sensor. The layers where the physics differs stay different all the way out; the middle, where meaning is made, is shared.
One line about the training setup is worth recording too. The masking ratio is 80 percent, but masking at 80 percent from the first step makes gradients explode. So it climbs from 40 to 80 percent over 300 steps. The encoders actually used are small. Tiny has 5.46 million parameters and Small 21.76 million, and both are twelve layers deep.
Contact Found on a Hand That Pretraining Never Saw
Attached as a frozen encoder to a hand that appears nowhere in the pretraining data, the model located the point of contact within 0.16 mm on average. That hand is XHand1, which carries magnetic sensing, and none of the three bundles — piezoresistive, magnetic, vision-based — holds a recording from that device. Not one parameter was changed. Once the model is given the sensor's geometry, each taxel's place on the fingertip is settled, and those coordinates go straight into the same encoder.
What it can do was measured with five probing tasks: reconstructing the force at the fingertip, pinpointing where contact occurred, estimating the area touched, predicting the force at the next instant, and retrieving the matching video from the tactile signal alone. The force sensor's measuring range is 0 to 30 N. The one comparison is FTP-1, which handles a new sensor by attaching an adapter built for it and training that.
There is a reason the comparison narrows to FTP-1 alone. Several earlier efforts tried to tie heterogeneous sensors together, and the paper gathers into one section where each of them stopped. Some stayed inside the single optical family that puts a camera behind a gel (Sparsh, AnyTouch). Some tied several optical sensors together while keeping a dedicated encoder per sensor (T3). Some need paired recordings, the same contact captured by two sensors at once, before alignment is possible (HTT). Crossing all three principles at once while also waiving the paired recordings leaves FTP-1 as the remaining opponent.
| Probing task | Sharpa Wave in pretraining |
〃 FTP-1 | XHand1 not in pretraining |
〃 FTP-1 |
|---|---|---|---|---|
| Fingertip force reconstruction error | 0.81 N | 2.31 N | 1.32 N | 3.72 N |
| Contact-center distance error | 0.12 mm | 2.19 mm | 0.16 mm | 5.89 mm |
| Contact area error | 0.66pp | 2.90pp | 0.43pp | 4.39pp |
| Next-force prediction error | 1.04 N | 1.67 N | 1.45 N | 2.00 N |
| Touch-to-video retrieval, Recall@5 | 72.6% | 16.5% | 47.1% | 6.0% |
The five probing tasks of §4.2. On the four error rows lower is better; on the last row higher is better. Orange marks LingBot-Touch. The left two columns are a vision-based robot fingertip that was in pretraining (Sharpa Wave), the right two a magnetic-sensing hand that was not (XHand1).
On the contact-center row alone, the gap opens by an order of magnitude. The 0.16 mm LingBot-Touch posts on the sensor it never saw is smaller than the 2.19 mm FTP-1 posts on the sensor it did see. The paper writes that comparison out as a sentence: "contact-center error on the unseen sensor is more than an order of magnitude below FTP-1's error on the seen one."
The same numbers rendered as lengths. FTP-1 was at 5.89 mm on XHand1 even after being trained with an adapter built for it, while LingBot-Touch was at 0.16 mm with nothing attached.
The design difference surfaces here. FTP-1 adapts by bolting on a dedicated part for each new piece of hardware; LingBot-Touch has no part to bolt on. The sensor's geometry is enough on its own: its taxels turn into coordinates and enter the same encoder. The model with a dedicated adapter lost; the model with nothing attached won.
The unseen sensor did not beat the seen one, though. The paper records the other direction in the same section. "Errors on XHand1 are higher than on Sharpa Wave on four of the five tasks, as expected for a sensor with no pretraining exposure." (§4.2) On four of the five tasks the XHand1 error is larger, and the explanation is that this is what you would expect from a sensor never exposed in pretraining. Saying transfer happens and saying nothing is lost are two different statements.
More interesting is how the XHand1 scores improve. Not one hour of data from that sensor was added, yet raising the hours on the magnetic side, or adding hours from an entirely different sensor family, lifts XHand1 along with them. The reading offered is this: "What improves with more and more varied data is a representation of contact, not of any particular transducer." What grows with the data is knowledge of contact as an event, not knowledge of any one transducer.
Same Signal, Different Format, Different Robot Scores
Same robot, same tasks, same number of demonstrations, and only the format in which the touch signal arrives was swapped — and the scores split visibly. The experiment ran on 26 real-robot tasks: 20 on a Sharpa Wave hand carrying vision-based sensing and 6 on the XHand1 hand that was absent from pretraining. Each task trained a policy on 200 demonstrations (500 for bimanual tasks) and was run over 20 trials, with three settings. The original policy using no touch, a policy with the raw signal added, and a policy with the tokens LingBot-Touch produces added. The signal entering the model comes from the same place in all three.
5.1Before the Numbers: This "Success Rate" Is Not a Pass Rate
The success rate reported here is not how many times out of twenty the robot pulled it off. Every task has its own 4-point or 5-point rubric, and what goes in the cell is the sum of the points earned across twenty trials divided by the maximum. The Appendix B header gives the scoring rule and the results-table caption gives the trial count.
"Numbers in parentheses denote points; allocations are additive unless an ordinal scale is stated, in which case the highest level reached determines the score." / "Success rates are reported in percent over 20 trials per task."
Appendix B header and the Tab. 2 caption, quoted together. Parenthesised numbers are points; unless an ordinal scale is stated, points add up, and on an ordinal scale the highest level reached is the score for that trial.
The arithmetic of the ticks proves this reading. Cap twisting is out of 4 points, so twenty trials make 80 the maximum and one point equals 1.25 percent. The twelve cells in that row read 56.25, 70.00, 76.25, 78.75, 85.00, 91.25, 75.00, 75.00, 87.50, 56.25, 62.50 and 90.00, every one of them a multiple of 1.25. Zipper closing and SD card ejection are out of 5 points, where one point equals 1.00 percent, and every cell in those rows is an integer. Had this been a count of passes and failures, there would be no reason for the tick size to differ from task to task.
The rubric for cap twisting runs like this: approach, 1; grasp and rotate effectively, 2; fully unscrew, 3; lift clear, 4. So 90 percent means an average of 3.6 stages over twenty trials, and 56.25 percent means an average of 2.25 stages. Render the numbers below as "it opened the bottle nine times in ten" and you have them wrong.
5.2Four Policies, Averaged over 26 Tasks
Each of four policies was given all three settings and averaged over the 26 tasks. The order is identical in every model. Raw signal beats no touch, and tokens beat raw signal. The size of the two steps is not the same. What the raw signal brings is a third to a half of the total, depending on the policy.
| Policy | No touch | + raw signal | + LingBot-Touch | Change |
|---|---|---|---|---|
| Diffusion Policy | 21.6 | 28.7 | 35.6 | +13.9pp |
| π0.5 | 48.1 | 54.7 | 69.9 | +21.8pp |
| LingBot-VLA | 48.3 | 56.5 | 73.1 | +24.8pp |
| T-Rex | 45.4 | 56.6 | 68.6 | +23.2pp |
The 26-task averages of Tab. 2. The unit is the percentage of the stage-based rubric score explained in 5.1 above. The raw-signal share for the four models was +7.0, +6.6, +8.2 and +11.2pp respectively. The encoder attached to the real robots is the Tiny trained on 4,000 hours; the one trained on the full 4,983 hours is the Small used in the probing experiments of the previous section.
LingBot-VLA, the largest mover in the table, is a policy built by Robbyant itself, the company that published the paper. No affiliation is printed in the paper, so it does not show on the surface, but the company's product line dates it to January 2026. This is not a thing to hold against the work. The other three are all outside models and they moved in the same direction by much the same margin, which if anything supports the result. Still, the figure this article carries as its headline is T-Rex's +23.2pp, from a third party.
5.3Cells Where the Raw Signal Hurts
Averaged out, the raw signal looks like it helps. Go task by task and model by model and the picture changes. Across 26 tasks × 4 policies, all 104 pairs, the policy with the raw signal added came out worse than the policy using no touch at all in 22 of them. LingBot-VLA dropped from 28.8 percent to 5.0 on mahjong sorting and from 65.0 to 28.8 on egg transfer, and π0.5 dropped from 39.0 to 15.0 on zipper closing. The explanation offered is simple.
"Given 200 demonstrations, a policy cannot learn to read a sparse, high-dimensional sensor from scratch, and the attempt can distract it from the signals it already reads. The frozen tokens arrive already meaning something."
§5.2, on why an unprocessed signal can cost a policy something it already had.
On the hand that pretraining never saw, the difference is sharper still. Across XHand1's six tasks, what the raw signal added to the four policies was +4.6, −1.7, 0.0 and +2.0pp, which is close to nothing. Converted into LingBot-Touch tokens, the same signal turns those slots into +16.3, +24.5, +17.8 and +13.8pp. Same sensor, same tasks, same demonstration count. Only the format changed.
Some cells reverse the order. In 9 of the 104 pairs the raw signal beat the tokens, and five of those nine are Diffusion Policy. The clearest case is SD card ejection. π0.5 scored 69.00 with the raw signal against 64.00 with tokens, and LingBot-VLA 72.00 against 64.00. So this is not a result that wins every cell, and that fact is printed in the paper's own table.
| Task (hand) | Policy | No touch | + raw signal | + LingBot-Touch |
|---|---|---|---|---|
| Cap twisting Sharpa Wave · 4-point ordinal | Diffusion Policy | 56.25 | 70.00 | 76.25 |
| π0.5 | 78.75 | 85.00 | 91.25 | |
| LingBot-VLA | 75.00 | 75.00 | 87.50 | |
| T-Rex | 56.25 | 62.50 | 90.00 | |
| Zipper closing Sharpa Wave · 5 points | Diffusion Policy | 14.00 | 9.00 | 22.00 |
| π0.5 | 39.00 | 15.00 | 59.00 | |
| LingBot-VLA | 11.00 | 16.00 | 55.00 | |
| T-Rex | 15.00 | 17.00 | 42.00 | |
| SD card ejection Sharpa Wave · 5 points | Diffusion Policy | 6.00 | 10.00 | 16.00 |
| π0.5 | 47.00 | 69.00 | 64.00 | |
| LingBot-VLA | 34.00 | 72.00 | 64.00 | |
| T-Rex | 32.00 | 84.00 | 92.00 | |
| Stacked cup separation XHand1 · 4 points | Diffusion Policy | 27.50 | 32.50 | 41.25 |
| π0.5 | 32.50 | 32.50 | 72.50 | |
| LingBot-VLA | 58.75 | 85.00 | 86.25 | |
| T-Rex | 41.25 | 46.25 | 55.00 |
Four of the 26 tasks (Tab. 2). Orange marks the winner in each row. The "56.25 to 90" that several outlets carried right after release is the T-Rex row of cap twisting; on the same task π0.5 moved from 78.75 to 91.25 and LingBot-VLA from 75.00 to 87.50. Stacked cup separation appears under both hands, so the hand is named alongside it.
Where the gain comes from is visible in the table too. The widest gaps sit at the one moment vision cannot adjudicate: whether the zipper slider is engaged, whether the trigger has been pulled far enough, whether the card has come all the way out. In the paper's words, "The largest per-task gains sit at the step of a task that vision cannot resolve."
Another experiment swaps a different encoder into the same slot. Built on LingBot-VLA — the authors' own policy — and run over five Sharpa Wave tasks (block flipping, SD card ejection, spray trigger actuation, stacked cup separation, plate wiping), the policy alone scored 55.9 percent and the policy carrying LingBot-Touch's Tiny tokens scored 82.2. FTP-1's tactile encoder in the same slot lands at 67.5, less than half the gain. FTP-1 helps on three of the five but drops stacked cup separation from 95 percent to 85 and plate wiping from 74 to 67. LingBot-Touch was at or above the original policy on all five.
The same five tasks also show what pretraining hours are worth. Trained on 1,000 hours and then on up to 5,000 in thousand-hour steps, and slotted in each time, the identical Tiny encoder lifts the policy to 69.8, 74.3, 77.4, 81.4 and 82.2 percent. None of the added hours came from these five tasks or from demonstrations on this robot. Not one extra hour of task data was supplied; only the contact the encoder had seen grew, and the robot's scores rose with it. The last thousand hours added 0.8pp, the smallest of the five steps.
Twelve simulated reinforcement-learning tasks put things in the same order. LingBot-Touch 69.0 percent, a CNN trained separately per task 45.6, raw signal 38.5, a binary signal carrying only whether contact occurred 30.3, and no touch 21.4. What deserves a second look is not the average but the wobble. Rerun with different seeds, LingBot-Touch held a standard deviation that peaked at 9.2 and averaged 3.9, while the alternatives widened to 55 on pen spinning. Whether a rerun of the same setting lands in the same place matters as much as the average once you are operating a site.
Pooling Comes with a Bill
In the very section that shows what pooling is worth, the paper writes out what it gives up as well. This is the least-quoted part of the work and, to anyone running data, the most usable. The question was split in two and each half measured. First, what happens when you keep all of your own data and add someone else's. Second, when the total hours available are fixed, is it better to give them all to one sensor or to share them across three.
| Question | What the paper measured |
|---|---|
| Keep your own data and add someone else's hours single-source is 331 vision hours or 228 magnetic hours; the multi-source model sees both of those plus 441 piezoresistive hours, 1,000 in all |
Seven of eight sensor-task pairs improve. The magnetic sensor improves on every task, with force error falling from 4.34 N to 2.52 N and next-force prediction from 3.77 N to 1.84 N. Contact-center error on the vision-based sensor, though, worsens from 0.41 mm to 0.68 mm |
| Cap the total budget at 1,000 hours and let them compete one sensor with 1,000 hours against the same 1,000 hours shared by three sensors |
The single-sensor model wins seven of eight pairs on its own sensor. The gap is largest on the spatial tasks (contact center 0.34 against 0.68 mm on vision, 0.33 against 1.12 mm on magnetic). On tasks predicting change over time, multi-source is the better one on the vision side (1.49 against 1.67 N) |
§4.3 and Fig. 8. The two "seven of eight" figures point in opposite directions. The upper row counts pairs won by the model that added someone else's data; the lower row counts pairs won by the single-sensor model at an equal budget.
What separates the two rows is whether the time spent on your own sensor shrinks. In the upper case the vision sensor's 331 hours and the magnetic sensor's 228 both stay where they are, and 441 hours of piezoresistive recording are laid on top to reach 1,000. In the lower case the same 1,000 hours are divided three ways, so the time the vision sensor spent looking at its own layout at its own resolution falls from 1,000 hours to 330. The gap therefore opens first on the spatial tasks.
The two numbers are not a contradiction. The questions differ. When extra hours arrive free, pooling wins; when the hours available have to be divided, pooling loses. Which of the two your site is in decides the call.
What crosses over and what does not also splits. The paper's summary draws that boundary exactly.
"Force and dynamics transfer across transducers, so hours from another sensor substitute for a sensor's own when those are scarce. Spatial precision depends on seeing a sensor's own layout at its own resolution, so at a fixed budget pooling costs some of it, in exchange for one encoder that serves three sensor families."
§4.3, the line that states both what pooling buys and what it bills.
6.1More Data Needs a Bigger Model
The effect of adding hours hangs on model size. Small keeps improving from 1,000 to 5,000 hours on every task on both hands, with force error on Sharpa Wave moving from 1.68 N to 0.81 N and retrieval recall from 42.5 percent to 72.6. Tiny follows the same curve to somewhere around 4,000 hours, then flattens, and at 5,000 its retrieval recall actually falls on both hands. The diagnosis reads "the data are not the limit; the model is, so capacity has to grow with the corpus." Collecting more data and growing the model that receives it have to move together.
6.2The Sensor Type That Failed the Quality Bar
Three limitations are stated. Hand types need to be more varied, the sensor range needs to be wider, and the coordinate frame and layout lean toward the fingertip, so contact on the palm and the inner surfaces of the fingers is under-represented. The note attached to the middle item is the heaviest line in this article.
"Capacitive sensors are currently absent from pretraining because the capacitive recordings available to us did not meet our quality requirements."
§6, limitations. A fourth sensor family is missing, and the reason given is the state of the recordings rather than the design.
However well the coordinate frame is designed, a recording that fails to clear the bar keeps a fourth sensor type out of the merge altogether. That is a confession that the precondition for pooling is in the end a quality gate, and it is the sentence that runs straight into the closing section of this article.
6.3Two Roads: Matching Bytes and Matching Representations
The same problem is being worked from two directions right now. One side matches the bytes. That road fixes the container that holds the data, as LeRobot and ROS do, then tries to agree on what data type touch goes in as. The container already exists; what is left is the semantic layer. What to name the keys, what units to use, how to write down the sensor geometry, what to treat as the baseline when nothing is touching. That is why the proposal seen in §1 has stayed open for three months. The other side matches the representation. It leaves the bytes as they are and has them meet in the embedding, and LingBot-Touch is the first large-scale report from that road.
It is too early to say which one wins. The two roads do ask for different things, though. Matching bytes has to wait for agreement; matching representations demands up front enough hours and enough quality to learn that representation at all.
TouchScale: One Tactile Glove, and Robot Success More Than Doubles, covered here five weeks ago, stands on the opposite axis. That project fixed the sensor to one type and grew only the hours, isolating what quantity alone carries. This article looks at what a project that made sensors mixable carries and what it gives up. Read the two together and you can see how much of robot tactile data is quantity and how much is format.
Why Pebblous Is Watching
From here on, this is not a fact recorded in the paper but a reading Pebblous has attached, and it should be taken separately.
The Method Was to Move the Input
What LingBot-Touch did was move the input into one place rather than build a bigger model. Rewrite the records as coordinates on a fingertip instead of sensor indices and data that could not be read across devices becomes a single training asset. It is a worked example of what Pebblous calls AI-Ready Data, and at the same time evidence that the work has to finish on the data side, outside the model. For a customer handling robot or equipment data, what this case shows is which question belongs after "let us collect more data."
A Quality Gate Shows Up Three Times
The quality gate turns up three times in this article, and two of the three sit inside the paper. One is the capacitive sensor dropped out of pretraining for failing the bar, the other is the 22 of 104 pairs where feeding the raw signal straight in cost the policy performance. Both say the same thing. Having data and being able to add it are different states. The third piece of evidence sits outside the paper. What blocked the discussion on the LeRobot proposal was not an algorithm either, but metadata: key names, units, sensor geometry, baselines. A sensor recording that does not carry its measurement conditions and its coordinate frame gets thrown away the moment the device changes. That is why DataClinic asks about collection conditions first when a dataset arrives.
A Conditional Answer to the Question Sites Actually Ask
"What happens to the records we have built up if we change the equipment?" is a real question at any site running robots or measurement gear. This paper gives an answer with conditions attached. Force and change over time carry to other equipment; spatial precision comes only from seeing your own layout at your own resolution. So what to weigh when deciding on pooling is not whether it can be done but which axis the precision you need sits on. And only an organization that left coordinates, specifications and baselines behind at the collection stage is in a position to make that call.
Where Pebblous Stands
Pebblous does not build tactile models. Its work is to leave records from different devices in a shape that can be pooled later. LingBot-Touch is a case of that work being done inside a model, and most sites do not build models. What remains for them is planting collection specifications, coordinates and lineage on the data side, and that work has to be finished before you can try your own records in an encoder someone else builds one day.
The figures and quoted lines in sections 1 through 6 were checked directly against the body, tables, appendix and limitations of the full text of arXiv:2611.12806 v1. The release scope and license were checked against the repository and the project page, and the LeRobot proposal against that pull request's body and review thread, all opened on November 17, 2026. Everything inside quotation marks is the English original. The paper's introduction puts the gain at "14 to 26 percentage points", while the 26-task averages in the results table run from 13.9 to 24.8pp, and this article followed the results table. Specifications and prices from the suppliers of the sensors and robot hands could not be confirmed, so no company names are listed in the body, and the size of the tactile sensor market was left out because the spread between research firms is too wide to use. Pebblous has no interest in the authors, in Robbyant, or in any institution named in the text.
References
Primary — the paper this article read
- 1.Bing Han et al. (19 authors; project lead Wei Ning). "Learning Heterogeneous Tactile Representation for General Dexterous Manipulation." arXiv: 2611.12806 — v1, November 9, 2026, cs.RO, CC BY 4.0. Every figure and quotation in the body was checked against the body, tables and appendix of this full text. As of November 17, 2026, v1 is the only version.
- 2.FTP-1 — arXiv: 2606.13102. The one direct comparison in this article, and the approach that adapts to new hardware by attaching a sensor-specific encoder.
- 3.The prior work item 1 draws its boundary against in §2.2 — T3 (per-sensor optical encoders plus a shared backbone), AnyTouch and AnyTouch 2 (alignment across optical sensors), HTT (optical and array sensors, requiring paired sensor recordings), Sparsh, Sparsh-skin and Sparsh-X (optical, magnetic, and multimodal within one device, respectively). This article went only as far as the scope description written in item 1 and did not read each original separately.
- 4.I-JEPA and V-JEPA — the lineage item 1 names for Tactile-JEPA. The architecture itself is not explained in this article.
- 5.Roland S. Johansson, J. Randall Flanagan. "Coding and use of tactile signals from the fingertips in object manipulation tasks." Nature Reviews Neuroscience, 2009 — the human-side evidence cited in §1 of item 1.
Primary — release status and the company
- 6.github.com/robbyant/lingbot-touch — Apache 2.0. Inference code, token extraction tooling, Tiny and Small checkpoints, per-sensor data interface code, and 120 sample recording files. Accessed November 17, 2026. The pretraining corpus is not included.
- 7.Project page
technology.robbyant.com/lingbot-touch· Hugging Facerobbyant/lingbot-touch· ModelScope — accessed November 17, 2026. - 8.Robbyant press release (November 10, 2026) and company profile — legal entity Shanghai Ant Robbyant Technology Co., Ltd., describing itself as "an embodied AI company under Ant Group, building one brain for all robots". The make-up of the LingBot model family, including that LingBot-VLA was released in January 2026, was confirmed in the same material.
Primary — the surrounding ecosystem
- 9."Add tactile sensor module with unified point-set interface (roadmap #3832)", huggingface/lerobot PR #4469 — opened August 18, 2026 and still open as of November 17, 2026. The body and the review thread were opened and checked directly.
Secondary — press coverage
- 10.AI Times (Korea), "Ant Group unveils an AI that gives robots a sense of touch" (November 10, 2026) · Special Times (November 11, 2026) — two Korean reports. The two render the company name differently in Korean.
- 11.roboticsandautomationnews.com (November 15, 2026) · interestingengineering.com — two English secondary reports. Both carried the introduction's "14 to 26 percentage points" as written.
Earlier Pebblous articles
- 12.TouchScale: One Tactile Glove, and Robot Success More Than Doubles (October 9, 2026) — the counterpart in 6.3. It covered a project that fixed the sensor to one type and grew only the hours.
- 13.The Tactile Data Humanoid Capital Left Behind (July 12, 2026) — the article that asked why it fails to accumulate, picked up in 1.2.
- 14.Why Is Robot Learning Data Still Scarce When Video Is Everywhere? (August 11, 2026) — the source of the "fifth translation machine" in 1.2.
- 15.Did Yann LeCun's Small Model Use Less Data, or Pair It Better? (November 8, 2026) — the article 3.1 hands the joint-embedding predictive architecture off to.
- 16.AgiBot Starts Recording the Texture of Contact in Touch (June 27, 2026) · Deform360 Touch Data for Deformable Objects (July 29, 2026) · What Do You Feed a Robot? (August 1, 2026) · Physical AI Datasets Hub — earlier articles in the same cluster.