Ant Lingbo open-sources tactile AI model LingBot-Touch that boosts robot success rates by up to 26 points

Ant Lingbo Technology has open-sourced a tactile AI model called LingBot-Touch, extending robots' sensing from vision to their fingertips and giving embodied AI a real sense of touch.

The model was pre-trained on nearly 5,000 hours of real touch data. It supports three main types of touch sensors and turns different signals into one shared touch format, so the learned "feel" can be reused across sensors, robot hands, models, and tasks.

In 26 real-world tests on two robot hands, four action models saw their average success rates rise by 14 to 26 percentage points with this model. For example, the success rate for twisting bottle caps went from 56.25% to 90%.

The team offers two model sizes. All 26 tests used the light version with about 5.3 million parameters, adding only 0.13% to 0.16% to the action models, balancing sensing power with on-device computing.