Holding an object with one hand is a balancing act: the fingers must release and re-grasp the object without letting it slip, even when the palm is turned downward. Researchers at the University of California, Berkeley and Davis (USA) taught a robotic hand to perform such manipulations and added a separate software safeguard to reduce the number of drops. The work was published on the arXiv preprint server.
In tests with external motion-tracking cameras, the safety algorithm cut the number of drops from 16 to four per 1,200 requested motion sequences. The price was that the hand missed its targets slightly more often.
The LEAP robotic hand has 16 joints. It was trained exclusively in computer simulation, without human demonstrations. The task was to move objects between the fingers, changing their position and orientation, rather than simply rotating them endlessly.
Training began with hundreds of thousands of computer-generated grasps. At first, friction at the fingertips in the simulation was unrealistically high, which made holding easier. As the model improved, this support was gradually reduced, bringing conditions closer to reality. In addition, the researchers measured the motor behavior of a real hand and incorporated it into the simulation.
Tests were conducted with a real light bulb, as well as with plastic 3D-printed copies of a knife, a wine glass and a duck. Each object required its own controller, a control program. The safeguard was trained separately: it adjusted the motions proposed by the controller so as to prevent drops while changing those motions as little as possible. The safety filter did not know the target position and received no reward for reaching it. "This separates task learning from safety learning," the authors write.
The evaluation included 4,800 requested movements of four objects at six wrist orientations. The trials ran in series of 50 targets; if an object fell or got stuck, the hand was reset. With external cameras tracking the object, the robotic hand reached 95.2% of targets without the safety algorithm and 91.5% with it. A task was considered successful if the object ended up within 2 centimeters and about 23 degrees of the target for 10 seconds.
Although the safeguard reduced the overall share of targets reached, it allowed the hand to perform more movements before the first drop. This difference matters when a dropped object interrupts the task. The researchers also demonstrated screwing a light bulb into a socket until it lit up, combining the learned finger movements with programmed arm movements. In these small demonstrations, the safeguard did not improve success, showing that its benefit does not automatically extend to any task.
For now, the system offers a tunable trade-off between reaching the manipulation goal and maintaining the grasp. The next tasks are to make the safeguard robust and to train a single controller to work with multiple objects.