Ant LingBot open-sources the LingBot-Touch tactile foundation model on November 10, raising success rates by 14 to 26 percentage points across 26 robot tests
LingBot-Touch is pretrained on nearly 5000 hours of real tactile data across magnetic, visuotactile and piezoresistive sensing, and was evaluated in 26 real robot tests on the Sharpa Wave and XHand dexterous hands
On November 10, Ant LingBot open-sourced LingBot-Touch, a tactile foundation model that gives robot brains a missing tactile perception capability. Pretrained on nearly 5000 hours of real tactile data, the model covers three mainstream sensing principles, magnetic, visuotactile and piezoresistive, and converts their signals into one shared tactile representation so that accumulated feel can be reused across sensors, dexterous hands, models and tasks.
The team ran 26 real robot tests on the Sharpa Wave and XHand dexterous hands, covering fine operations such as turning book pages, unscrewing bottle caps and identifying mahjong tiles. After adding the tactile representation, four action models, LingBot-VLA, pi 0.5, Diffusion Policy and T-Rex, improved their average task success rate by roughly 14 to 26 percentage points over the no-touch baseline. In the T-Rex task of unscrewing a bottle cap on Sharpa Wave, success was 56.25 percent without touch, 62.5 percent when raw sensor signals were fed in directly, and 90 percent once the LingBot-Touch representation was used.
Architecturally, LingBot-Touch sets up separate encoding entrances for each sensing principle, processes differing raw signals first and then maps contacts with different layouts into a unified fingertip coordinate space, emitting unified tactile tokens for different action models to consume. That design lets data collected from people wearing tactile gloves and teleoperating a dexterous hand be trained on together, and leaves room to plug in new kinds of tactile sensors later.