LingBot-Touch: Learning Heterogeneous Tactile Representation for General Dexterous Manipulation

Tactile sensing is essential for contact-rich dexterous manipulation, but existing commercial tactile sensors vary wildly in physics and design, making cross-sensor data scaling difficult. To address this, the researchers developed LingBot-Touch, a tactile foundation model designed to learn a unified and extensible representation across heterogeneous tactile sensors including piezoresistive, magnetic, and visuotactile devices. The model outputs plug-and-play tactile tokens that integrate seamlessly into existing policy architectures, such as diffusion policies and vision-language-action (VLA) models like pi_0.5 and LingBot-VLA. Extensive experiments spanning simulation and real-robot deployments across multiple hardware platforms show that the proposed representation consistently improves manipulation dexterity while maintaining high computational efficiency. As a preprint, further independent replication and code or weight releases will help confirm the generalizability of these cross-sensor tactile representations.