Researchers from OpenAI, a nonprofit artificial intelligence research group founded in 2015, today announced that they are teaching robotic hands to turn a block of colorful letters into the desired page of the block.
The task is simple. But the breakthrough was in the way the hand gained the ability: all learning was done in a software simulation and then relatively easily transferred to the physical world.
This solves a challenge for robot hands that resemble the fist of a robot in the 1980s science fiction movie Terminator. Hands have been commercially available for years, but are hard for engineers to program. Engineers can write specific computer codes for each new activity whenever a new expensive program is needed. Or the robots can be equipped with software that allows them to "learn" through physical training.
Physical training takes months or years and has specific problems, for example, when a robot hand releases a piece, a human must take it and replace it. It's very expensive: Researchers have tried to share those years of physical training and distribute them on multiple computers for a software simulation that can be done in hours or days without human assistance.
Ken Goldberg, a professor of robotics at the University of California at Berkeley, who did not attend OpenAI research but reviewed it, defined the work of OpenAI, which was released on Monday as "a major achievement" in achieving this goal.
"It's the beauty of having many computers," said Goldberg. "You do not need robots, you just have a lot of simulation."
