Zero to Expert in Eight Hours: These Robots Can Learn For Themselves
Zero to Expert in Eight Hours: These Robots Can Learn For
Themselves
By Pavel Alpeyev
December 3, 2015 — 4:00 AM PST
A yellow robotic arm pauses over a pile of metal
cylinders, snaps a photo, then proceeds to confidently pick pieces out of the
jumble. What’s impressive is that just eight hours ago, its bin-picking skills
were about zero.
Fanuc Corp. is showing off the first results of its
partnership with artificial intelligence startup Preferred Networks Inc. at
Tokyo’s International Robot Exhibition this week. The world’s largest maker of
automation equipment is using so-called “deep learning” to enable machines that
can acquire skills independently.
Left overnight, a robot empowered by the algorithms can
use trial and error to figure out how to pick up randomly positioned objects
with 90 percent accuracy. Eight machines working simultaneously and sharing
their lessons can do that in an hour. A veteran Fanuc engineer would need
several days to write a teaching program that even approaches that performance.
“Deep learning allows machines to take vast amounts of
data and distill useful rules and lessons all by themselves,” Preferred
Networks Chief Executive Officer Toru Nishikawa said at the expo. “For a robot,
that means understanding not only why one movement was successful and another
one not, but also how to improve its performance.”
Fanuc plans to make the self-learning functionality
commercially available next year, Nishikawa said. The two companies are also
working on applications that predict machine failures to prevent costly factory
stoppages.
The deep learning algorithms, inspired by the way living
things process information, already help minimize human involvement when
Facebook Inc. tags user photos or Google Inc. serves up ads. Now hardware
companies from Fanuc to Toyota Motor Corp. and Samsung Electronics Co. are
stepping up investments to add problem-solving capabilities to their products.
Fanuc earlier this year paid 900 million yen ($7.5
million) for a 6 percent stake in Preferred Network, after rival ABB Ltd.
invested several million dollars into AI startup Vicarious. Facebook’s Mark
Zuckerberg, Amazon.com Inc.’s Jeff Bezos, actor Ashton Kutcher and Samsung are
also among Vicarious’s shareholders.
Nishikawa is a world finalist of the prestigious ACM
Programming Contest, whose winners include a former chief technology officer of
Facebook and the first employee at Google. He founded Preferred Networks with
his University of Tokyo classmate Daisuke Okanahara in March 2014. Little more
than a year later, the company counted Fanuc, Toyota and Panasonic Corp. among
its partners.
“This technology doesn’t have to be limited to Fanuc
robots,” Nishikawa said. “There are all kinds of machines that can benefit from
it.”
Comments
Post a Comment