Berkeley researchers’ robotic hand drops fewer objects with a software safeguard
Ian Chuang and colleagues at the University of California, Berkeley and Davis trained a 16-joint robotic hand in simulation to rotate and reposition objects between its fingers. In tests, a separate safety system cut drops from 16 to four across 1,200 movement sequences, while reducing the hand’s target success rate from 95.2% to 91.5%, according to Electronics For You.
Bottom line — The safeguard trades some accuracy for fewer drops; the researchers say its settings can be adjusted independently of the hand’s manipulation skill.
Go deeper 6
-
Electronics For You reports that training used computer simulations, without human demonstrations, and incorporated measured motor behaviour from a real robotic hand.
-
The researchers tested four 3D-printed objects—a knife, wine glass, duck and light bulb—with a separate controller for each, Electronics For You reports.
-
The full evaluation involved 4,800 requests across four objects and six wrist orientations, according to Electronics For You.
-
Yahoo Tech says the safety system does not know the target position or receive rewards for reaching it; it adjusts movement to preserve grip while disturbing the main task as little as possible.
-
Yahoo Tech reports a demonstration in which the hand rotated a light bulb, moved it to a socket and screwed it in, though the safeguard was not equally effective in every demonstration.
-
The findings are posted on the arXiv preprint server, according to Yahoo Tech and Lifeboat News.