Ant's Lingbo open-sources LingBot-Touch, a general sense of touch for robots

Ant Group's embodied AI unit Lingbo has open-sourced LingBot-Touch, a tactile foundation model built to carry touch experience from human hands onto robot hands it has never encountered, without retuning a single parameter for new hardware.

The clearest evidence comes from experiments on XHand. Its magnetic tactile sensor model never appeared in pretraining data, and all magnetic tactile data in pretraining came from gloves worn by humans. With LingBot-Touch attached unchanged, π0.5's average success rate across six tasks rose from 48.00% with raw signals to 74.17%.

On the task of separating stacked cups it went from 32.5% to 72.5%. The job requires taking only the top cup: one hand steadies the stack while the other pinches the rim and pulls. If resistance suddenly drops while the fingertip stays on the rim, the cup has come loose and force must be released; if the fingertip itself is slipping, the grasp has to be redone. The two sensations are entirely different.

A comparison with the earlier tactile foundation model FTP-1 shows the generalization gap. Judging contact centers on Sharpa Wave, FTP-1's error was 2.19 mm against 0.12 mm for LingBot-Touch; on XHand the error fell from 5.89 mm to 0.16 mm. FTP-1 used dedicated adapters for each hand, while LingBot-Touch used one frozen encoder for both.

Scaling behavior is already visible. The Tiny version saturates around 4,000 hours, while the larger Small version was still improving at 5,000 hours, and real-robot success rates rose steadily with pretraining data.

LingBot-Touch slots into a wider Lingbo stack: LingBot-Vision and LingBot-Depth 2.0 for spatial perception, LingBot-Video and LingBot-World 2.0 for environment interaction, LingBot-VLA 2.0 and LingBot-VA 2.0 for dexterous manipulation. Most of those solve what happens before contact — where the object is, where the hand should go, what to do next. Once fingers touch, vision degrades fast: a fingertip against a bottle cap is occluded, and a key inside a lock is invisible.

The division of labor is simple: vision handles the approach, touch handles what follows. Two-handed unlocking shows it. Gripping the padlock and aiming the key rely on the eyes; pushing the key in can tilt the lock or jam halfway, requiring a partial retreat and a new angle, and once the key is inside the camera is largely useless — only the resistance transmitted through the key to the fingertips remains.

Lingbo's argument for making touch a standalone foundation model is that every hand is controlled differently but all of them must understand contact. Isolating that layer means experience from one piece of hardware becomes the starting point for the next. The design choices follow: gloves collect data because they do not depend on a specific robot body, which lets the dataset scale, and inputs are organized by sensing principle rather than product model to leave room for future sensors.

LingBot-Touch is already fused with LingBot-VLA and LingBot-Depth to improve hand-eye coordination and fine manipulation in Lingbo's general-purpose brain. Commercially, the company is running tests with hardware partners including Leju and Tihu, and deployment partners such as Guoda Pharmacy and Longsheng, in retail sorting, logistics sorting and industrial settings.