Ant Lingbo Open-Sources Tactile Foundation Model LingBot-Touch, 5.32M Parameters Integrates with Mainstream Action Models

Fable 5.2 assembles a Rubik's Cube with its eyes open, while Chinese AI plays mahjong with its eyes closed! Human "sense of touch" tokenized for the first time Xin Zhiyuan November 10, 2026 11:38:36 From Beijing Updated November 10, 2026 12:02:24

Without any warning, Anthropic dropped its next-generation flagship Fable 5.2 late at night, plunging directly into the physical world.

Its all-around performance crushes Opus 5.5, and it even demonstrates astonishing generalization ability—even unfamiliar new tasks are handled with ease.

Assembling a Rubik's Cube bare-handed, twisting off bottle caps, drawing robots, installing gears and testing gear meshing—Fable 5.2 achieved a 64.9% success rate, easily crushing a host of opponents in the RoboDojo test, more than double that of GPT-6 Astra.

Just two months ago, GPT-6 Astra likewise showed off a heaven-defying feat—after being hooked up to a pair of dexterous hands by geeks, it played a Chopin nocturne on the spot.

Looking at this, wouldn't you think large models are about to take over the physical world?

In that very same evaluation, the RoboDojo test team put this "former leaderboard champion" Astra onto real hardware to take on plugging in a charger and inserting a test tube.

As a result, because it had only eyes and no sense of touch, even though the plug was already jammed in crooked and stuck fast, it could not feel the resistance at all.

So the unsuspecting robotic arm, following visual instructions, kept ramming straight ahead and smashed the real hardware on the spot!

Researchers, terrified, rushed to pull the emergency stop, and in the end, across 33 real-machine tests, these two tasks were never completed even once—the success rate was a flat zero (0%).

Why such a huge contrast?

Because tasks like assembling a Rubik's Cube and drawing only require "eyes" to see, but when it comes to fine handiwork with the hands, cameras alone are far too slow and simply cannot see.

On one hand, the moment a cup starts slipping at the fingertips, by the time the camera captures the frame it has long since shattered.

On the other hand, whether a bottle cap is actually being turned or the fingers are just spinning in place looks exactly the same in the camera, and the contact point is often blocked by the fingers themselves.

For this very reason, for tasks like inserting wiring harnesses in car factories, precisely because the split-second feel at the fingertips is missing, Audi admits that the industry-wide automation rate is still below 10% to this day.

In the final analysis, what robots lack today is no longer better eyes, but a sensitive "sense of touch."

Just moments ago, Ant Lingbo Technology officially released and open-sourced the general-purpose tactile foundation model LingBot-Touch, filling in this most critical shortcoming in one stroke!

For the first time, it uniformly converts three different tactile signals—vision-based tactile, magnetic tactile, and piezoresistive—into "universal tactile Tokens" that action models can use directly.

This way, robots can finally develop a set of "sense of touch" that keeps working even when swapped onto a different dexterous hand.

Therefore, in order not to compete for compute with vision and action models, the team directly deployed the lightest Tiny version, with only 5.32 million parameters.

As a result, after connecting to the three mainstream action models, the parameter count increases by only 0.13% to 0.16%, adding almost no burden, yet instantly giving dexterous hands a "sense of proportion."

Take "grabbing a spray bottle and spraying water": the plastic bottle body is very slippery, and relying purely on vision, the moment force is applied, the index finger slips and loses its grip.

But after connecting LingBot-Touch, the instant there is the slightest slip at the fingertip, it can immediately fine-tune its grip posture and steadily pull the trigger, with the success rate soaring directly from 42.50% to 87.50%.

Next, look at "squeezing toothpaste," which tests two-arm coordination even more: squeeze the tube too lightly and nothing comes out, squeeze too hard and it splatters everywhere.

With the help of tactile Tokens, the two arms can not only steadily align with the narrow brush head, but the thumb can also apply delicate pressure following the resistance of the tube body, squeezing the toothpaste precisely onto the bristles:

Even the "hole alignment" task that at the start caused the vision large model to smash real hardware is easily handled when switched to the more delicate "using a key to open a lock."

The fingertips pinch the small key and, following the faint collision resistance of the keyhole, probe and fine-tune while aligning, then crisply twist the padlock open in one go:

Even more interesting is "blindly feeling mahjong," with rules so strict that if the tile face is exposed to the camera even once, it is immediately scored zero.

The dexterous hand relies entirely on no eyes throughout; a gentle press of the fingertip can feel out the embossed patterns and sort them into different plates, with the success rate directly doubling to 70%:

For another example, separating two stacked soft paper cups—the fingers block the camera completely, squeeze a bit too hard and the cup crumples, squeeze too lightly and it cannot be pulled out.

But with the help of tactile Tokens, the two arms can follow the faint friction changes at the fingertips and, with a gentle rub, pull out the inner paper cup.

Beyond these everyday fine tasks, the team used two dexterous hands of different configurations paired with four mainstream action models to test 26 real-machine tasks and 12 simulated operations in one go, and the combinations connected to LingBot-Touch won across the board:

The three major action models all improved their scores comprehensively—T-Rex, π0.5, and LingBot-VLA saw average success rates rise by 23.18, 21.81, and 24.81 percentage points respectively;

Fine manipulation was steady, accurate, and forceful—turning a single page of a book surged to 91.25%, pressing to eject an SD card soared to 92%, and carrying soft paper cups reached a perfect 100% at best;

12 simulated operations passed with high scores—spinning a plate and spinning a pen both exceeded 97% success, and "balancing a rolling ball on a tray," which scored zero with pure vision, also reached 67.7%.

5,000 hours of real tactile data

Dexterous hands can use it right away

Since touch is so useful, why is a general large model so hard to train? The main bottlenecks are "sensors each speaking their own language" and "an extreme shortage of real data."

First, the hardware: the three mainstream types of tactile sensors currently available have completely different underlying principles:

Visual-tactile sensing hides a tiny camera in the fingertip to capture the deformation of soft gel, allowing it to discern extremely subtle contact and making it suitable for twisting off bottle caps, but its reaction speed is limited by frame rate and is not fast enough;

Magnetic tactile sensing relies on soft magnetic materials changing the magnetic field under force, and is good at sensing multi-directional forces and slipping, making it suitable for pulling zippers and plugging in plugs, but it is susceptible to interference from surrounding magnetic fields;

Piezoresistive sensing, meanwhile, works by the material's resistance changing under pressure, and can cover fingers and palms at low cost, at the price of limited resolution.

As a result, even when pressing against the wall of a cup in the same way, the three types of sensors record three completely incompatible "dialects," and switching to another hand means retraining from scratch.

Even more troublesome is the lack of real data. In the past, everything relied on human teleoperation of robotic arms to collect it, which was slow and expensive and simply could not feed large models.

To solve the data problem, Ant Lingbo came up with a clever approach—having real people wear lightweight tactile gloves to directly collect "human touch feel" in everyday life!

That way, whether twisting off a bottle cap, wiping a table, or plugging in a plug, however the human hand applies force, the glove records it.

Ultimately, by combining humans wearing gloves with real-machine teleoperation, the team amassed the world's largest real tactile dataset of nearly 5,000 hours (4,983 hours) and 1.33 billion frames, fully two-thirds of which came from human gloves.

But human hands and dexterous hands are not shaped the same. Is the collected data actually useful for robots?

On this point, real-world tests gave a pleasantly surprising answer—the "touch feel" accumulated from human hands could actually transfer zero-shot directly to a brand-new dexterous hand!

Keep in mind that 100% of the magnetic tactile data in pretraining came from human gloves, and the model had never seen the new magnetic tactile dexterous hand used in testing.

Yet after the team completely froze the tactile model parameters and directly installed them on this new dexterous hand, the average success rate of π0.5 immediately soared from 48% to 74.17%, and for carrying soft paper cups it jumped from 58% to 90%!

This means the ability to understand physical contact has crossed hardware forms and been passed unchanged from human hands to dexterous hands.

Unifying three types of tactile signals

Developing a real sense of touch by predicting applied force

So at the underlying level, how do you fit data from three types of sensors, human hands, and dexterous hands into the same model? The Lingbo team seized on the most core commonality:

No matter whether a sensor measures images, magnetic fields, or resistance, in the end it answers the same four questions—where the finger touched, how much force was used, which way the force went, and whether it is slipping.

Since the questions are the same, the team pulled out "understanding contact" separately and built it into a general foundation model in three steps.

Step one: "receive" each type separately according to its principle.

Specifically, the team set up an extremely simple entry point for each of the three sensor types, converted them into a unified format, and then handed them to the same tactile backbone network for learning, so new hardware based on the same principle can be plugged in directly.

Step two: put all fingertips on the same "map" (5D-RoPE).

Unifying the format alone is not enough; the contact point layouts of different dexterous hands also vary enormously.

Therefore, the team mapped all contact points uniformly into 3D fingertip spherical coordinates and randomly perturbed their density and positions to prevent the model from rote-memorizing the layout of a single piece of hardware.

Step three: solve a "contact completion" problem (Tactile-JEPA).

Why design this problem specifically? Because when grasping an object, more than 90% of contact points actually touch nothing at all, and their readings are all zeros and background noise.

If you simply copy vision models and mask a patch to reconstruct the raw readings, the model can get 90 points for free just by filling in all zeros.

As a result, with 4,000 hours of data, the traditional approach achieved only 3.06% and 8.69% recall at top-5 in cross-modal retrieval!

So LingBot-Touch drew on the JEPA architecture championed by Turing Award winner LeCun and directly masked up to 90% of the area.

On that basis, instead of making the model reconstruct zero-value background noise, it makes the model infer how force is applied in the masked areas, lifting retrieval recall all the way to 49.32%!

To this end, in five physical contact tests, the team allowed the mainstream open-source model FTP-1 to attach a hardware-specific adapter, while LingBot-Touch kept its weights fully locked and did not change a single parameter.

Even after giving its opponent this handicap, LingBot-Touch still reduced fingertip blind-touch localization error from the opponent's 2.19 mm to 0.12 mm, and also reduced it to 0.16 mm on the new magnetic tactile dexterous hand.

At the same time, fingertip 3D force sensing error also dropped by about 65%, and when installed in the same action brain to test five real-machine tasks, it won by a wide margin with an average success rate of 82.2% versus 67.5%.

More importantly, at the scale of 5,000 hours, the team observed that scaling also exists in touch, revealing two key trends:

On one hand, when data from three different types of sensors is trained together, all metrics comprehensively beat training them separately, and data from each hardware route finally converges in one place, producing a "1+1+1>3" effect.

On the other hand, as data accumulated to nearly 5,000 hours, the lightweight Tiny version gradually saturated after 4,000 hours, while the larger-capacity Small version kept rising.

This shows that the larger the model and the more data, the stronger the robot's "sense of touch"!

Hardware need not be unified; intelligence can be shared

The value of LingBot-Touch actually goes beyond giving robots one more sense.

Keep in mind that touch is a typical "physically native modality"—text and images are everywhere online, yet large models in the digital world have never truly "touched" anything, which is also why Astra at the beginning did not even notice when the plug got stuck.

In addition, because sensor principles differ, touch has long remained scattered and fragmented, and has never formed a unified representation.

Now, LingBot-Touch has proven that even touch, the most hardware-dependent sense of all, can be distilled into a unified representation.

This has been validated on touch, and looking further, it can fully be extended to embodied intelligence as a whole.

The reason physical manipulation is hard to generalize largely comes down to hardware: bodies, dexterous hands, and sensor approaches all differ, and the experience painstakingly accumulated ends up locked inside each set of hardware.

Ant Lingbo's approach is very clear on this: hardware need not be unified; intelligence can be shared.

What it aims to build is the layer of intelligence between hardware and scenarios—turning the physical experience accumulated across different bodies and different sensors into a unified representation that can be shared, then offering it out in an open-source, decomposable form.

This year, Lingbo has already open-sourced, one after another, the spatial perception model responsible for seeing more clearly, the world model responsible for thinking more clearly, and the cross-body action model responsible for acting more deftly.

LingBot-Touch, open-sourced this time, not only fills in the more sensitive-to-touch piece of the "embodied-native" brain full stack 2.0, but also applies the same approach to touch, the most hardware-dependent sense of all—the fact that touch can work shows that this approach can likewise be applied to other new sensing modalities.

When it comes to industry, this layer of shared intelligence directly saves upstream and downstream players the effort of reinventing the wheel.

For touch sensor manufacturers, in the past every time a new sensor was launched, they had to collect data themselves and train a touch model from scratch before the action model could use it, which meant a high barrier and a long cycle.

Now, new hardware can plug directly into LingBot-Touch to produce general-purpose touch Tokens, so manufacturers can focus their efforts on precision, durability, and cost, and their products are also easier for various robots to adopt.

For dexterous hand makers and complete-machine manufacturers, the biggest change is that selection is finally free.

On a single hand, different parts inherently need different things: the fingertips handle fine actions such as pinching, twisting, and rubbing, and need high resolution to perceive subtle contact; the palm is mainly responsible for supporting and gripping, and needs large-area, low-cost sensing of force.

The ideal approach is naturally to let each part play to its strengths, but in the past, using two kinds of sensors on one hand meant configuring two sets of touch models to train separately, adding an extra share of development and maintenance costs.

Now, the same LingBot-Touch can understand touch signals based on different principles, and a cost-effective combination like "fingertip visuotactile + palm piezoresistive array" can finally be confidently installed on a hand and truly put to use.

At present, Ant Lingbo has joined forces with a number of robot body manufacturers and industry partners to bring this embodied brain capability into retail pharmacies, warehousing and logistics, and factory production lines for real work.

Looking back, the reason large models could explode in the digital world is that massive amounts of text and images can be learned in a unified way and reused repeatedly.

Now that even the most hardware-dependent "sense of touch" can be pretrained and reused, physical intelligence finally has the possibility of accumulating ever more thickly, just like language intelligence.

Hardware can flourish in a hundred ways, yet intelligence can keep growing on the same layer.

The inflection point at which robots truly enter every industry may well start here.