AI News
All stories
Official

Robbyant's LingBot-Touch gives different tactile sensors one shared representation

Robbyant introduces LingBot-Touch, a tactile foundation model that maps piezoresistive, magnetic and visuotactile sensors to one virtual fingertip.

Research · · 1 min read

Robbyant introduces LingBot-Touch, a tactile foundation model that gives three kinds of touch sensors one shared representation. The model was pretrained on 4,983 hours of real touch data and maps piezoresistive, magnetic and visuotactile sensors onto the same virtual fingertip. The authors say it works on the XHand1 robot hand, whose sensor the model never saw during training.

The arXiv paper, whose first author is Bing Han, says the resulting tactile tokens are plug-and-play. They can be attached to policies ranging from diffusion policy to vision-language-action models such as π0.5 and LingBot-VLA. Because every sensor maps to the same virtual fingertip, downstream policies can work from one shared representation.

On the XHand1, a frozen LingBot-Touch with a small trained add-on reached 0.16 mm average contact-center error. FTP-1, with an XHand1-specific adapter, reached 5.89 mm.

Across six real-world XHand1 tasks, LingBot-VLA averaged 57.21% success whether or not it received raw tactile input. With LingBot-Touch features, it averaged 75.04%. The post does not name the six tasks or say how many trials each was run for, so the spread around those averages is not visible. The abstract also says the tactile representation benefits manipulation across simulation and real-robot deployments on multiple hardware platforms, and that it is easily extensible to new sensor types.

What changed

LingBot-Touch maps piezoresistive, magnetic and visuotactile sensors onto one shared virtual fingertip, so its tactile tokens can be attached to policies such as diffusion policy, π0.5 and LingBot-VLA.

Type: New capability · Applies to: LingBot-Touch, XHand1, π0.5, LingBot-VLA

Why it matters

Robotics teams could reuse one tactile model across sensor hardware instead of training an adapter for each, based on the authors' reported results.

Original source

Where we saw it

1 source

DAIR.AI (Elvis Saravia)

Robbyant

AI wrote this summary from the sources listed, and nobody reviewed it individually before publication. Check the original source before you rely on it. How it works