SimToolReal teaches a robot hand to use many different tools by practicing in simulation and then working in the real world without extra training.
Robots learn faster and more flexibly when they can use human touch data, but humans and robots feel touch with very different sensors.
Computers usually click like a woodpecker, but they struggle to drag smoothly like a human hand; this paper fixes that.
GR-Dexter is a full package—new robot hands, a smart AI brain, and lots of carefully mixed data—that lets a two-handed robot follow language instructions to do long, tricky tasks.