Robbyant, an embodied AI unit of Ant Group, open-sourced LingBot-Touch on November 10, 2026, a tactile foundation model designed to translate signals from different types of touch sensors into a single format that robot control systems can use. The model was pre-trained on nearly 5,000 hours of real-world tactile data and covers three common sensing types: magnetic tactile, visual tactile and piezoresistive sensing. In testing across 26 real-world tasks performed with two dexterous robotic hands, Sharpa Wave and XHand, adding LingBot-Touch lifted success rates by 14 to 26 percentage points depending on the task and underlying action model. On one example task, unscrewing a bottle cap, success climbed from 56.25% with no tactile input, to 62.5% with raw sensor signals, to 90% once LingBot-Touch processed the same signals.
Touch Has Lagged Vision in Robot AI for a Structural Reason
Robot perception research has focused overwhelmingly on vision because cameras produce a standardized signal that's easy to pool across manufacturers and build large training sets from. Tactile sensors do not share that uniformity: a magnetic sensor, a camera-based visuotactile sensor, and a piezoresistive sensor each output fundamentally different signal types, which has historically forced developers to build sensor-specific software for every hand they support. LingBot-Touch addresses this by giving each sensing principle its own encoding pathway while mapping all of them into a shared fingertip coordinate space, so a model trained on one sensor type can still interpret signals from another.
Robbyant kept the model lightweight on purpose. The smaller variant, LingBot-Touch-Tiny, has roughly 5.3 million parameters and adds only 0.13% to 0.16% to the size of the action models it plugs into, a design choice aimed at devices with limited onboard compute rather than requiring a data-center connection to process touch in real time.
A Narrower But More Exploited Gap Than Vision Models
Tactile foundation models remain far less developed than their vision counterparts; most dexterous-hand research to date has relied on raw, sensor-specific signals or has skipped touch entirely in favor of vision and force estimation. LingBot-Touch was tested against four different action models, including LingBot-VLA, π0.5, Diffusion Policy and T-Rex, rather than being locked to Robbyant's own proprietary stack, which widens its potential user base beyond Ant Group's own robots.
The open release puts touch sensing in the same category as vision-language-action models: a shared, swappable component rather than something every hardware maker rebuilds from scratch. For companies building dexterous hands, that lowers one of the costlier integration barriers, though the model's real test will be whether gains measured in a research setting with two specific hands transfer cleanly to the dozens of other tactile sensor designs already on the market.
Source: Robotics Tomorrow