Learning Heterogeneous Tactile Representation for General Dexterous Manipulation
LingBot-Touch gives robot policies a shared representation of touch across three sensor types.
The model combines data from piezoresistive, magnetic and visuotactile sensors, with a design intended to accommodate additional types. Its tactile tokens plug into diffusion policies and vision-language-action models, including π₀.₅ and LingBot-VLA. The authors report improved dexterity in simulation and real-robot experiments across multiple hardware platforms, while maintaining computational efficiency. The abstract provides no numerical results to quantify those gains.
HF Daily Papers' note
The model combines data from piezoresistive, magnetic and visuotactile sensors, with a design intended to accommodate additional types. Its tactile tokens plug into diffusion policies and vision-language-action models, including π₀.₅ and LingBot-VLA. The authors report improved dexterity in simulation and real-robot experiments across multiple hardware platforms, while maintaining computational efficiency. The abstract provides no numerical results to quantify those gains.
HF Daily Papers' note
score 5