Robbyant, the embodied AI unit of China's Ant Group, has released LingBot-Touch, an open-source model that lets a robot use data from different kinds of touch sensors, Humanoids Daily reported on 10 November.
The model converts signals from magnetic, vision-based and piezoresistive sensors into shared tokens, so a robot hand does not need a newly trained model when its sensors change. It was pretrained on 4,983 hours of touch data, from people handling objects in sensor gloves and from teleoperated robots.
In tests on 26 tasks with two kinds of multi-fingered hands, adding LingBot-Touch to four existing robot control models raised average success rates by 13.9 to 24.8 percentage points. It also improved results with the touch sensors of a hand left out of training, without retraining the encoder.
Robbyant published versions with 5.46 million and 21.76 million parameters, with the code and weights on GitHub and Hugging Face under the Apache 2.0 licence. The tests used only two kinds of hand, and capacitive sensors were not in the training data.
Background
LingBot-Touch
Robot touch AI model
An open-source AI model that processes data from different kinds of tactile sensors for robot hands.
