Advertise
AIGadgets

Ant Group's Robbyant open-sources LingBot-Touch tactile robot model

The model was trained on about 5,000 hours of real tactile data and lifted average scores of four existing control AI models by 13.9 to 24.8 points in 26 robot tasks.

Ant Group's robotics subsidiary Robbyant released LingBot-Touch, an open-source tactile robot foundation model, on November 10, 2026, according to aitimes.com. The model was trained on roughly 5,000 hours of real tactile data and captures contact location, pressure, force distribution and changes over time.

LingBot-Touch unifies data from three kinds of tactile sensors — magnetic, piezoresistive and vision-based — into a single representation. It can be added to existing robot control AI models to give them tactile perception, and Robbyant says it can be integrated with its own LingBot VLA as well as external control models such as π0.5, Diffusion Policy and T-Rex.

In 26 real-robot task evaluations using the Sharpa Wave and XHand robot hands, applying LingBot-Touch to four existing control AI models raised average scores by 13.9 to 24.8 percentage points. On the Sharpa Wave hand, the bottle-cap-opening task scored 56.25% without tactile information and 90% with LingBot-Touch.

Robbyant offers two model sizes, including the lightweight LingBot-Touch Tiny with about 5.5 million parameters. Code and pre-trained weights were released on GitHub and Hugging Face under the Apache 2.0 license.

Robot AI push

Ant Group established Robbyant at the end of 2024, and the subsidiary unveiled R1, a robot that performs cooking and service tasks, last year. Earlier this year Ant Group released robot AI models including LingBot VLA for visual perception and motion generation and LingBot World for predicting physical-world motion, and it has invested in Chinese robotics companies such as Unitree.

Meta released the general-purpose tactile AI model Sparsh and the artificial-fingertip tactile sensor Digit 360 in October 2024, and also pursued development of robot hands equipped with tactile sensors with Korea's Wonik Robotics.

Quick answers

Where can I download LingBot-Touch?

Code and pre-trained weights are available on GitHub and Hugging Face under the Apache 2.0 license.

How large is LingBot-Touch Tiny?

The lightweight variant has about 5.5 million parameters.

What data was LingBot-Touch trained on?

It was trained on roughly 5,000 hours of real tactile data covering contact location, pressure, force distribution and changes over time.

Source