Ant Group's Robbyant open-sources LingBot-Touch, a touch model that lifts four robot policies by 14 to 26 points

Robbyant, the embodied-AI company inside Ant Group, has open-sourced LingBot-Touch, a tactile foundation model trained on 4,983 hours of fingertip recordings from three kinds of sensors: piezoresistive, magnetic and visuotactile. Added to four existing robot policies as a frozen 5.46 million-parameter Tiny encoder, pretrained on 4,000 of those hours, it raised their average success on 26 real-world two-arm tasks by 14 to 26 percentage points, the authors report in a technical report posted to arXiv on Nov. 9. The policies were Diffusion Policy, Physical Intelligence's π0.5, Robbyant's own LingBot-VLA and T-Rex.

The control run is the real finding. The team also gave the same policies the raw tactile signal. That helped at best about half as much, 7.0 to 11.2 points, and made results worse in 22 of the 104 task-policy pairs. With raw touch, LingBot-VLA fell from 28.8% to 5.0% on mahjong sorting. A robot hand that can feel is not the hard part anymore. Turning what it feels into something a policy can use is.

A hand the model never felt still gained

The magnetic fingertip sensor on XHand1 was absent from pretraining. With no parameter updated for it, LingBot-Touch tokens added 13.8 to 24.5 points to the four policies' mean success across XHand1's six tasks, while the raw signal moved them by between −1.7 and +4.6 points. On the dexterous hand Sharpa Wave, whose visuotactile sensors were in the training data, π0.5's mean success across 20 tasks rose from 47.6% to 68.6%.

The authors say no tactile sensor today has 5,000 hours of data of its own, and their experiments suggest it does not need them: hours from one sensing principle improved the model on another. The corpus mixes human tactile gloves with teleoperated robot hands, 464,431 episodes over 24 task categories. The Tiny encoder adds 0.13% to 0.16% to the parameter counts of π0.5, LingBot-VLA and T-Rex. Robbyant says the code and model are on GitHub and Hugging Face in two sizes.

Where touch beats sight

The gains were not universal. LingBot-Touch fell below the original policy in 5 of the 104 pairs, and each task was judged over 20 trials after 200 demonstrations, or 500 for tasks that need both hands. The biggest jumps came at steps a camera cannot settle: whether a bottle cap has broken loose (T-Rex, 56% to 90%), whether a spray trigger has been pulled far enough (LingBot-VLA, 42.5% to 87.5%), and whether a zipper slider has caught, where LingBot-VLA went from 11% to 55%.