← AI glossary

AI glossary · Basic · Robotics & physical AI · Multimodal

LingBot-Touch

LingBot-Touch is an open-source tactile foundation model released by Robbyant, a unit of Ant Group. It converts signals from tactile sensors that work on different principles into a single common 'tactile token,' so that robot control models can use fingertip sensing.

About this term
Level
Basic · Technical terms you often meet in AI news
Added
2026-11-17
In plain words It is similar to an interpreter who translates reports arriving in many languages into one language. It translates each sensor's different numbers into a common language of 'where contact happened, how hard, and in which direction the force is leaning,' and passes that to the robot's decision-making model. Unlike a human interpreter, though, it translates not into words people can read but into bundles of numbers only the model understands.

What it is

LingBot-Touch is a foundation model (a large model pretrained in advance for use across many tasks) specialized in understanding tactile information from robot hands. Robbyant released it as open source in November 2026, and a related paper appeared on arXiv around the same time. It unifies signals from three kinds of tactile sensors, piezoresistive, magnetic and vision-based (which use a camera to capture deformation of a gel at the fingertip), into one Representation (a shared internal representation in the model). According to reports, the code and pretrained weights were released on GitHub and Hugging Face under the Apache 2.0 license, and the model comes in two sizes, including a lightweight version, 'LingBot-Touch Tiny.'

How it works

Each sensor type has its own lightweight input stage that turns raw readings into tokens, and after that all sensors share a single tactile transformer backbone. Because sensors differ widely in the number and layout of taxels (tactile points), it was described as mapping all taxels into a standardized fingertip space, attaching an 'address' made of time, finger and 3D position, and feeding this to the model with RoPE. Pretraining follows the JEPA approach: most taxels (90%, according to reports) are masked, and the model looks at the remainder to predict representations of the masked parts in feature space rather than their raw values. This is because most tactile data is zero, meaning 'no contact,' so asking the model to reconstruct raw values tends to push it toward the shortcut of learning only the background well. The resulting tactile tokens can be plugged into diffusion policies or VLA models such as π0.5 and LingBot-VLA.

Architecture of LingBot-TouchPiezoresistive sensor → Per-sensor input stage; Magnetic sensor → Per-sensor input stage; Vision-based sensor → Per-sensor input stage; Per-sensor input stage → Shared tactile transformer; Per-sensor input stage → Shared tactile transformer; Per-sensor input stage → Shared tactile transformer; Shared tactile transformer → Tactile tokens; Tactile tokens → Robot control modelPiezoresistivesensorMagnetic sensorVision-basedsensorPer-sensor inputstageRaw values → tokensPer-sensor inputstageRaw values → tokensPer-sensor inputstageRaw values → tokensShared tactiletransformerJEPA-stylepretrainingTactile tokensContact position ·force · directionRobot controlmodelDiffusion policy · VLA
Architecture of LingBot-Touch Three kinds of sensors pass through their own input stages into a shared transformer backbone, and the tactile tokens it produces are used by multiple robot control models.

Why it matters

Changing the type of tactile sensor or robot hand has often meant collecting data again and retraining the model, and this is an attempt to reduce that burden. According to reports, it was trained on about 5,000 hours of real tactile data collected from tactile gloves worn by people and from robot hands. One notable finding is an experiment in which feeding raw tactile signals directly into a control model sometimes lowered the success rate. According to Chinese media reports, on closing a zipper bag, the success rate was 39% with cameras only, dropped to 15% when raw tactile input was added, and rose to 59% when passed through LingBot-Touch. In other words, it shows that 'organizing signals so the model can understand them' matters more than 'adding more sensors.'

Zipper bag closing success rate

Cameras only
Raw tactile added
LingBot-Touch
Success rates with the same robot hand and π0.5, changing only how tactile input is added (per Chinese media reports). Feeding raw signals directly actually lowered it.

Things to know

Robbyant said that across 26 tasks evaluated on two kinds of robot hands, the average scores of four existing control models rose by 13.9 to 24.8 percentage points. This evaluation gives partial credit for intermediate steps, so it differs from the rate at which tasks were fully completed. The performance figures come from the company's own evaluation, and further validation is needed to see whether they hold on other hardware or in real industrial settings. Some figures, such as the parameter count of the lightweight model, were reported slightly differently across outlets, so it is safer to check the model card in the official repository.

Example

Picture a robot hand opening a bottle cap. The moment a finger slips slightly on the cap is hard to see with a camera. Robbyant said the Sharpa Wave robot hand's evaluation score on bottle-cap opening was 56.25% without touch and rose to 90% with LingBot-Touch attached. Developers can keep the VLA or diffusion policy models they already use and simply connect tactile tokens on top, letting them try tasks where force control matters, such as turning a key or squeezing toothpaste.

Where it came up on this site

Related terms

References