Ant Group Open-Sources LingBot-Touch Foundational Tactile Model for Robotics
Ant Group's LingBot-Touch has been open-sourced as a foundational tactile model trained on nearly 5,000 hours of real-world data, enabling cross-sensor and cross-task tactile perception. In 26 real-world manipulation tests, integrating this model improved average task success rates by approximately 14 to 26 percentage points. This open-source release addresses a major bottleneck in embodied AI by standardizing heterogeneous tactile data into a unified representation. It allows robots to move beyond vision-only limitations, greatly enhancing performance in dexterous manipulation tasks like turning bottle caps or flipping book pages. The model supports magnetic, vision-based, and piezoresistive tactile sensors, mapping different layouts into a shared fingertip coordinate space. The lightweight LingBot-Touch-Tiny version contains approximately 5.3 million parameters and adds only 0.13% to 0.16% parameter overhead when integrated into action models.
## BACKGROUND
Embodied AI focuses on intelligent agents like robots that interact physically with the real world using sensors and actuators. Tactile sensing allows robots to detect physical forces, contact locations, and slippage, which are crucial for fine manipulation tasks that vision alone cannot resolve.