Ant Group’s robot unit releases a free AI model that helps robot hands make sense of touch
Robbyant says robots using it opened bottle caps successfully 90% of the time, up from 56% without touch data. The results are the company’s own.
Robbyant, a robotics company within Ant Group, the Chinese financial technology company behind the Alipay payment app, has released a free AI model meant to give robots a better sense of touch, the company announced. The model, called LingBot-Touch, and its code are posted publicly online.
The problem it tries to solve
Most robots that pick things up rely mainly on cameras. But many jobs, such as turning a page or twisting off a bottle cap, depend on feel: how hard to squeeze and whether something is slipping. Robot hands use several kinds of touch sensors, including ones that sense magnetic changes, tiny cameras that watch a soft fingertip bend and pads whose electrical resistance changes under pressure. Each produces a different kind of signal, so a robot AI trained on one type of sensor usually cannot use another.
What Robbyant built
LingBot-Touch was trained on nearly 5,000 hours of real-world touch data from all three sensor types and converts their signals into one shared format that different robot hands and robot control AIs can use, Robbyant says. The company describes it as a foundation model, a large AI model trained on broad data that other systems can build on. Smaller versions of the model are posted for download on Hugging Face, a site for sharing AI models.
Results so far
In 26 real-world tests on two robot hands, adding LingBot-Touch raised the average success rates of four different robot control AIs by about 14 to 26 percentage points, the company says. In one test, a robot hand twisting off a bottle cap succeeded 56% of the time without touch input, 63% when fed raw sensor readings and 90% with LingBot-Touch.
These results come from Robbyant’s own tests and have not been independently checked. The company did not say whether the model is being used in any commercial robots.
Sources: Robbyant announcement via Business Wire; LingBot-Touch code on GitHub; LingBot-Touch model on Hugging Face.