Robots will use the latest computer-vision and machine-learning
algorithms to try to perform the work done by humans in vast fulfillment
centers.
Packets
of Oreos, boxes of crayons, and squeaky dog toys will test the limits
of robot vision and manipulation in a competition this May. Amazon is
organizing the event to spur the development of more nimble-fingered
product-packing machines.
Participating
robots will earn points by locating products sitting somewhere on a
stack of shelves, retrieving them safely, and then packing them into
cardboard shipping boxes. Robots that accidentally crush a cookie or
drop a toy will have points deducted. The people whose robots earn the
most points will win $25,000.
Amazon
has already automated some of the work done in its vast fulfillment
centers. Robots in a few locations send shelves laden with products over
to human workers who then grab and package them. These mobile robots,
made by Kiva Systems, a company that Amazon bought in 2012 for $678 million,
reduce the distance human workers have to walk in order to find
products. However, no robot can yet pick and pack products with the
speed and reliability of a human. Industrial robots that are already
widespread in several industries are limited to extremely precise,
repetitive work in highly controlled environments.
Pete Wurman,
chief technology officer of Kiva Systems, says that about 30 teams from
academic departments around the world will take part in the challenge,
which will be held at the International Conference on Robotics and
Automation in Seattle (ICRA 2015).
In each round, robots will be told to pick and pack one of 25 different
items from a stack of shelves resembling those found in Amazon’s
warehouses. Some teams are developing their own robots, while others are
adapting commercially available systems with their own grippers and
software.
The 25 items that participating robots will need to retrieve from shelves.
The
challenge facing the robots in Amazon’s contest will be considerable.
Humans have a remarkable ability to identify objects, figure out how to
manipulate them, and then grasp them with just the right amount of
force. This is especially hard for machines to do if an object is
unfamiliar, awkwardly shaped, or sitting on a dark shelf with a bunch of
other items. In the Amazon contest, the robots will have to work
without any remote guidance from their creators.
“We
tried to pick out a variety of different products that were
representative of our catalogue and that pose different kinds of
grasping challenges,” Wurman said. “Like plastic wrap; difficult-to-grab
little dog toys; things you don’t want to crush, like the Oreos.”
The video below shows the approach taken by a team at the University of
Colorado. The team is using off-the-shelf software and building a robot
arm specialized for the task, says Dave Coleman, a PhD student
involved.
The
contest could offer a way to judge the progress that has been made in
the past few years, when some cheaper, safer, and more adaptable robots
have emerged (see “How Technology Is Destroying Jobs”)
thanks to advances in the technologies underlying machine dexterity.
New types of robot manipulators are making machines less ham-handed at
picking up fiddly or awkward objects, for example. Several startups are
developing robot hands that seek to copy the flexibility and sense of
touch found in human digits. Progress in machine learning could help
robots perform far more sophisticated object manipulation in coming
years.
A
key breakthrough in this area came in 2006, when a group of researchers
led by Andrew Ng, then at Stanford and now at Baidu, devised a way for
robots to work out how to manipulate unfamiliar objects. Instead of
writing rules for how to grasp a specific object or shape, the
researchers enabled their robot to study thousands of 3-D images and
learn to recognize which types of grip would work for different shapes.
This allowed it to figure out suitable grips for new objects.
In
recent years, robotics researchers have increasingly used a powerful
machine-learning approach known as deep learning to improve these
capabilities (see “10 Breakthrough Technologies 2013: Deep Learning”). Ashutosh Saxena,
a member of Ng’s team at Stanford and now an assistant professor at
Cornell University, is using deep learning to train a robot that will
take part in the Amazon challenge. He is working with one of his
students, Ian Lenz.
While
the Amazon challenge might seem simple, Saxena believes it could
quickly make an impact in the real world. “If robots are able to handle
even the light types of grasping tasks the contest proposes,” he says,
“we could actually start to see a lot of robots helping people with
different tasks.”