UC Berkeley robot hand learns to twist and turn objects without dropping them, preprint says
Trained only in simulation, the hand reached 95% of its targets in real-world tests.
A robot hand trained entirely in computer simulation can rotate delicate objects in the air and rarely drops them, according to UC Berkeley and UC Davis researchers. Their study, called DexSafe, is a preprint that has not been peer reviewed.
The team used a 16-joint robot hand called a LEAP hand and trained it with no human demonstrations. The simulated fingertips started out extra grippy and became more realistic during training, so the hand learned not to rely on unrealistic grip.
On the real robot, the researchers tested 4,800 unseen target positions over 13.3 hours with a lightbulb, a 3D-printed knife, a wineglass and a rubber duck. The hand reached 95.2% of them. A safety add-on called ResSafe cut drops from 16 to four while still reaching 91.5% of goals. Using only a wrist camera, it reached 72%. It also screwed in a lightbulb until it lit.
Each object needs its own trained controller. The authors, led by Ian Chuang and including Berkeley professor Claire Tomlin, plan to release their code.
Source: arXiv preprint 2610.37652 and DexSafe project page