Learning Heterogeneous Tactile Representation for General Dexterous Manipulation

This paper introduces LingBot-Touch, a tactile foundation model designed to unify data from diverse sensor types (piezoresistive, magnetic, visuotactile) into a common representation. The model produces plug-and-play tokens that can be integrated into various policy frameworks, including diffusion policies and vision-language-action (VLA) models like π0.5 and LingBot-VLA. Experiments across multiple hardware platforms in both simulation and real-world settings demonstrate that the unified tactile representation enhances dexterity for complex manipulation tasks while maintaining computational efficiency. The work highlights the potential of cross-sensor data to scale tactile learning and informs future hardware and model design.