Robotics

Ant Group's Robbyant Open-Sources LingBot-Touch, a Model That Lets Robots Share a Sense of Touch Across Different Sensors

Trained on nearly 5,000 hours of real tactile data from three kinds of sensors, LingBot-Touch turns their different signals into one common format that any robot policy can use. Robbyant says it lifted task success by roughly 14 to 26 points across four action models, and took one bottle-cap test from 56 per cent to 90 per cent.

DailyMech Staff

Ant Group's Robbyant Open-Sources LingBot-Touch, a Model That Lets Robots Share a Sense of Touch Across Different Sensors
Tumisu / Pixabay — Pixabay Content License

A white robotic hand on a black mechanical arm reaches in from the right to touch the fingertip of a smiling young woman against a pale turquoise background

Illustrative photo of a robotic hand. It does not show Robbyant's robots or sensors.

Robots can see well and, increasingly, reason well. What most of them still cannot do is feel. Robbyant, the embodied-AI company inside Ant Group, is trying to close that gap with LingBot-Touch, a "tactile foundation model" it has released as open source.

The model was pre-trained on nearly 5,000 hours of real-world touch data, according to Robbyant's announcement, published by Robotics & Automation News on Sunday. Its job is to sit between a robot's touch sensors and the software that decides how the robot moves, and to translate whatever those sensors report into a single, shared language.

The code is on GitHub under the Apache 2.0 license, and two pre-trained checkpoints, LingBot-Touch-Tiny and LingBot-Touch-Small, are on Hugging Face and ModelScope. A technical paper, "Learning Heterogeneous Tactile Representation for General Dexterous Manipulation," was posted to arXiv on Nov. 9.

At a glance

Item Detail Source
Developer Robbyant, part of Ant Group Robotics & Automation News
Training data Nearly 5,000 hours of real tactile data Robotics & Automation News
Sensor types Magnetic, visuotactile, piezoresistive Announcement, README
Real-robot tests 26, on Sharpa Wave and XHand dexterous hands Robotics & Automation News
Action models tested LingBot-VLA, π0.5, Diffusion Policy, T-Rex Robotics & Automation News, arXiv
Average gain About 14 to 26 percentage points in task success Robotics & Automation News
Smallest model LingBot-Touch-Tiny, about 5.3 million parameters Robotics & Automation News
License Apache 2.0 GitHub, Hugging Face

The problem: every sensor speaks a different language

Touch sensors for robot hands are still young, and there is no standard. Some use magnets, some use small cameras looking at a soft skin, and some measure changes in electrical resistance. Each approach has its own strengths and costs, and Robbyant expects all of them to stay in use for a long time.

That creates a data problem. The same press of a fingertip produces completely different raw signals depending on the sensor and where its contact points sit. A model trained on one sensor usually cannot use another's data, so when a team changes its sensor or its robot hand, it often has to collect new data, retrain and re-integrate. Robbyant argues that this repeated work raises costs for hardware makers, integrators and model developers alike.

How LingBot-Touch works

The model gives each kind of sensor its own input pathway, then maps contact points from different layouts onto a shared fingertip coordinate system. From there, one shared model learns where contact is happening, how force is spread and how both change over time. The output is a set of "tactile tokens" that a robot's control policy can read, whatever hardware produced them.

Technically, the GitHub page describes it as a JEPA-style model: a shared Transformer that learns by predicting hidden parts of the touch signal in a compressed internal space, rather than reconstructing raw readings. The public release includes the frozen encoder and tools for pulling out token-level features, which is what developers need to plug it into their own systems.

Because the format is shared, data from different sources can be trained together. Robbyant says it combined recordings from people wearing tactile gloves with data from remotely operated robot hands.

The results Robbyant reports

Robbyant ran 26 real-world tests on two dexterous hands, Sharpa Wave and XHand. It added LingBot-Touch to four different action models: its own LingBot-VLA, the π0.5 vision-language-action model, Diffusion Policy and T-Rex. Average task success rose by about 14 to 26 percentage points compared with running without touch input, the company said.

One example shows why the shared format matters more than simply adding sensors. In a test where T-Rex controlled a Sharpa Wave hand to unscrew a bottle cap, success was 56.25 per cent with no touch input. Feeding the raw sensor signals straight into the policy raised it only to 62.5 per cent. Using LingBot-Touch's processed representation took it to 90 per cent.

The XHand test was about transfer. That hand's magnetic sensor was not in the pre-training data at all; the model's only magnetic experience came from human glove recordings. Robbyant kept LingBot-Touch frozen and trained only the action model on XHand demonstrations, to check whether the touch knowledge carried over to new hardware. The project's GitHub page says the model transfers to new tactile sensors and real robot hands without additional training.

Robbyant also says the tasks it targets include turning pages, unscrewing caps and telling Mahjong tiles apart by feel.

Small enough to run on the robot

All 26 tests used the Tiny version, at about 5.3 million parameters. Added to LingBot-VLA, π0.5 and T-Rex, it increases total model size by only about 0.13 to 0.16 per cent, Robbyant said. That matters because touch needs to be processed fast, on the robot, while it is gripping something.

A larger version did better in perception tests, and its performance kept improving as training data approached 5,000 hours. Robbyant reads that as a sign that tactile pre-training has a lot of room to grow with more data.

Why it matters

Most of the recent progress in robot learning has come from vision and language. But many of the jobs people want robots to do, from assembling cables to handling food or folding laundry, depend on feel: knowing when a grip is slipping or a cap is tight. A shared touch layer that works across sensors could let the robotics industry pool its scarce tactile data, much as large vision models let developers stop training image recognition from scratch.

Robbyant is pitching LingBot-Touch as a general-purpose perception layer rather than a module for its own robots. It says LingBot-Touch is not tied to its own models, and its tests on third-party policies support that.

The caveats are real. Every number here comes from Robbyant's own tests, on hardware it chose, and none has been reproduced by outside researchers yet. Twenty-six real-world tests is a modest sample.

What to watch

Watch whether other labs and sensor makers adopt the shared format, whether independent groups reproduce the bottle-cap and transfer results, and whether Robbyant releases the larger model and training code, not just inference tools.

How we reported this

DailyMech read Robbyant's announcement as published by Robotics & Automation News, the project's GitHub README and the arXiv abstract. We did not read the full paper. All performance figures are the company's and were not independently tested. DailyMech contacted no one for this article.

Published November 15, 2026. Facts checked against the cited sources by DailyMech's AI writer.

Sources

DailyMech used AI tools to research and draft this article. Facts were checked against the cited sources by DailyMech's AI writer, not by a human editor.