A Warehouse Robot Learns to Sort Out the Tricky Stuff - Major Advance in AI...
A
Warehouse Robot Learns to Sort Out the Tricky Stuff
LUDWIGSFELDE,
Germany — Inside a warehouse on the outskirts of Berlin, a long line of blue
crates moved down a conveyor belt, carrying light switches, sockets and other
electrical parts. As they came to a stop, five workers picked through the small
items, placing each one in a cardboard box.
At Obeta, an electrical parts company
that opened in 1901, it is the kind of monotonous task workers have performed
for years.
But several months ago, a new worker
joined the team. Stationed behind protective glass, a robot using three suction
cups at the end of its long arm does the same job, sifting through parts with
surprising speed and accuracy.
While it may not seem like much, this
component-sorting robot is a major advance in artificial intelligence and the
ability of machines to perform human labor.
As millions of products move through warehouses run by Amazon,
Walmart and other retailers, low-wage workers must comb through bin
after bin of random stuff — from clothes and shoes to electronic equipment — so
that each item can be packaged and sent on its way. Machines had not really been
up to the task, until now.
“I’ve worked in the logistics industry
for more than 16 years and I’ve never seen anything like this,” said Peter
Puchwein, vice president of Knapp, an Austrian company that provides automation
technology for warehouses.
Standing nearby at the Obeta warehouse,
the California engineers who made the robot snapped pictures with their
smartphones. They spent more than two years designing the system at a start-up
called Covariant.AI, building on their research at the University of
California, Berkeley.
Their technology is an indication that,
in the coming years, few warehouse tasks will be too small or complex for a
robot. And as the machines master tasks traditionally handled by humans, their
development raises new concerns about warehouse workers losing their jobs to
automation.
Because the online retail business is
growing so quickly — and most companies will be slow to adopt the latest
robotic technologies — economists believe the advances will not cut into the
overall number of logistics jobs anytime soon. But the engineers building these
technologies admit that at some point most warehouse tasks will be done by
machines. Human workers will need to find other things to do.
The engineers at Covariant specialize in
a branch of artificial intelligence called reinforcement learning. The
machines are wired to learn new tasks on their own through extreme trial and
error. And the best place to learn is in the real world.
“If you want to advance artificial
intelligence, you don’t just do it in a lab,” said Peter Chen, Covariant’s
chief executive and co-founder. “There is a huge gap in bringing it to the real
world.”
Warehouses are already highly automated. At
the facility outside Berlin, inside a fenced-off room larger than a football
field, other robots have long been used to fetch large boxes from shelves several
stories high.
But that is a relatively easy task for a
machine. Engineers can program a robot to perform the same motion over and over
again. The boxes are uniform. A robot can pick them up with the same motion
every time.
Picking through a bin of random items is
different. Shapes vary, as do surfaces. One light switch might be upside down,
the other right-side up. The next electrical gadget might be in a plastic bag
that reflects light in ways a robot has never seen. A human touch has been
needed.
Programming a robotic arm to deal with
every situation, one rule at a time, is impossible. At Knapp, Mr. Puchwein and
his partners had tried and failed for years to create a robot with the
dexterity and flexibility needed for the job.
Covariant, which is working with Knapp,
built software that could learn through trial and error. First, the system
learned from a digital simulation of the task — a virtual recreation of a bin
filled with random items. Then, when Mr. Chen and his colleagues transferred
this software to a robot, it could pick up items in the real world.
The robot could continue to learn as it
sorted through items it had never seen before. Inside the German warehouse, the
robot can pick and sort more than 10,000 different items, and it does this with
more than 99 percent accuracy, according to Covariant.
Late last year, the international robot
maker ABB ran a contest. It invited 20 companies to design software for its
robot arms that could sort through bins of random items, from cubes to plastic
bags filled with other objects.
Ten of the companies were based in
Europe, and the other half were in the United States. Most came nowhere close
to passing the test. A few could handle most tasks but failed on the trickier
cases. Covariant was the only company that could handle every task as swiftly
and efficiently as a human.
“We were trying to find weaknesses,” said
Marc Segura, managing director of service robotics at ABB. “It is easy to reach
a certain level on these tests, but it is super difficult not to show any
weaknesses.”
Knapp, which helped deploy the system
outside Berlin, and ABB believe this technology can be used in similar
warehouses.
Covariant engineers believe their robots
will improve with practice. As a robot in one warehouse learns better ways for
picking up certain items, the information feeds back to what is essentially a
central brain run by Covariant that will help operate machines.
Dirk Jandura, the managing director of
Obeta, said companies like his were under extreme pressure to be more
efficient. Automation is a key way to keep costs low.
Like many warehouse operators,
Obeta has trouble finding workers willing to do the monotonous work. Each
picker handles about 170 orders an hour, or about three per minute, over an
eight-hour day. In the summer, temperatures in the warehouse reach more than
100 degrees. It is hard to keep employees for longer than six months.
For Obeta, the new robot is an ideal
solution. A job that requires three humans is done by one tireless robot.
“It doesn’t smoke, is always in good
health, isn’t chatting with its neighbors, no toilet breaks,” Mr. Jandura said.
“It’s more efficient.”
Knapp is also considering the design of
warehouses staffed by robots rather than humans that would allow for packages
to be more densely packed into spaces and retrieved by robots trained to
perform the task.
“The new warehouses will be built around
A.I. robots and not humans,” Mr. Puchwein said.
Knapp plans to make it hard for companies
to say no to replacing human workers with robots. Mr. Puchwein said they would
charge a fee that was always lower than what a company would pay a human. If a
company paid $40,000 per year to a worker, Knapp would charge about $30,000, he
said.
“We just go lower,” he said. “That’s
basically the business model. For the customer, it’s not very hard to decide.”
Beth Gutelius, associate director of the
Center for Urban Economic Development at the University of Illinois at Chicago,
who has studied the impact of automation on work, said this kind of technology
was unlikely to shift the job market any time soon.
The greater problem, she said, is that as
humans work alongside robots, they will be judged in new ways. “As we start to
compare the speed and efficiency of humans to robots, there is a whole new set
of health and safety issues that emerge,” she said.
Pieter Abbeel, a Berkeley professor who
is a co-founder of Covariant as well as its president and chief scientist, said
humans would continue to work alongside machines in these kinds of warehouses.
But he acknowledged that the job market would significantly shift as machine
learning improved.
“If this happens 50 years from now, there
is plenty of time for the educational system to catch up to the job market,” he
said.
At the German warehouse, a woman in a
baggy T-shirt diligently sorted through the boxes, occasionally looking up at
the English-speaking visitors who were taking pictures of the robot and were
marveling at its effectiveness.
A Covariant engineer walked over to the
group to share that the robot had filled more than 200 orders in the past hour,
enough to receive a bonus if it were a human.
Adam Satariano reported from
Ludwigsfelde, and Cade Metz from San Francisco.
Comments
Post a Comment