A 5-year-old can tie their shoelaces, but robot hands aren’t nearly so nimble. A new system, however, has greatly improved their dexterity.
Hard-coding a robot to coordinate multiple joints is daunting. So computer scientists have turned to machine learning, a field of artificial intelligence (AI) in which computers build skills on their own. Such learning takes time and repetition, however, and robot hardware is slow and breakable. Some researchers instead train algorithms with virtual robots, but reality is always slightly different from simulation.
The new work overcame this “reality gap” by slightly randomizing elements of the simulation during training, such as friction and object size. (Most of the work, in both simulation and reality, was done with a child’s building block with letters on its sides.) They also gave the program short-term memory, so after a few seconds of handling the cube, it got a sense of the block’s exact size and other factors and adjusted for them.
The researchers used the commercial Shadow Dexterous Hand, which resembles a human hand, attached to a wall, along with a digital simulation of the hand for training. In both virtual training and a physical test to see how well the training transferred to the real hand, the hand was instructed to manipulate a cube in a series of new orientations so that, for example, the side with the A on it was facing up and side with the P on it was facing out. No robot hand had ever done something nearly as complicated.
In the real world, the system “saw” the cube using three cameras placed above the hand. The virtual hand, after the equivalent of 100 years of trial-and-error practice (sped up in simulation), performed an average of 30 consecutive reorientations without getting stuck or dropping the cube. The physical hand completed an average of 15 consecutive reorientations without getting stuck or dropping the cube, the researchers report today. The system, called Dactyl, also discovered common human tricks such as spinning the cube between two fingertips or taking advantage of gravity to shift the block.
The advance might improve the assembly of delicate electronics or the ability of health care or domestic robots to help around the house. Omelet, anyone?