Robots That Learn Through Repetition, Not Programming
Robots That Learn Through Repetition, Not Programming
A startup says getting a robot to do things should be
less about writing code and more like animal training.
By Tom Simonite on September 22, 2014
Making it easier to give robots intelligent behavior
could make them cheaper and more widely used.
In an onstage demonstration this week, Todd Hylton of
Brain Corporation used gestures to train a wheeled robot to come when he
beckoned to it.
Eugene Izhikevich thinks you shouldn’t have to write code
in order to teach robots new tricks. “It should be more like training a dog,”
he says. “Instead of programming, you show it consistent examples of desired
behavior.”
Izhikevich’s startup, Brain Corporation, based in San
Diego, has developed an operating system for robots called BrainOS to make that
possible. To teach a robot running the software to pick up trash, for example,
you would use a remote control to repeatedly guide its gripper to perform that
task. After just minutes of repetition, the robot would take the initiative and
start doing the task for itself. “Once you train it, it’s fully autonomous,”
says Izhikevich, who is cofounder and CEO of the company.
Izhikevich says the approach will make it easier to
produce low-cost service robots capable of simple tasks. Programming robots to
behave intelligently normally requires significant expertise, he says, pointing
out that the most successful home robot today is the Roomba, released in 2002.
The Roomba is preprogrammed to perform one main task: driving around at random
to cover as much of an area of floor as possible.
Brain Corporation hopes to make money by providing its
software to entrepreneurs and companies that want to bring intelligent,
low-cost robots to market. Later this year, Brain Corporation will start
offering a ready-made circuit board with a smartphone processor and BrainOS
installed to certain partners. Building a trainable robot would involve
connecting that “brain” to a physical robot body.
The chip on that board is made by mobile processor
company Qualcomm, which is an investor in Brain Corporation. At the Mobile
Developers Conference in San Francisco last week, a wheeled robot with twin
cameras powered by one of Brain Corporation’s circuit boards was trained live
on stage.
In one demo, the robot, called EyeRover, was steered
along a specific route around a chair, sofa, and other obstacles a few times.
It then repeated the route by itself. In a second demo, the robot was taught to
come when a person beckoned to it. One person held one hand close to the
robot’s twin cameras, so that EyeRover could lock onto it. A second person then
maneuvered the robot forward and back in synchronization with the trainer’s
hand. After being led through a rehearsal of the movements just twice, the
robot correctly came when summoned.
Those quick examples are hardly sophisticated. But
Izhikevich says more extensive training conducted over days or weeks could
teach a robot to perform a more complicated task such as pulling weeds out of
the ground. A company would need to train only one robot, and could then copy
its software to new robots with the same design before they headed to store
shelves.
Brain Corporation’s software is based on a combination of
several different artificial intelligence techniques. Much of the power comes
from using artificial neural networks, which are inspired by the way brain
cells communicate, says Izhikevich. Brain Corporation was previously
collaborating with Qualcomm on new forms of chip that write artificial neural
networks into silicon (see “Qualcomm to Build Neuro-Inspired Chips”). Those
“neuromorphic” chips, as they are known, are purely research projects for the
moment. But they might eventually offer a more powerful and efficient way to
run software like BrainOS.
Brain Corporation previously experimented with
reinforcement learning, where a robot starts out randomly trying different
behaviors, and a trainer rewards it with a virtual treat when it does the right
thing. The approach worked, but had its downsides. “Robots tend to harm
themselves when they do that,” says Izhikevich.
Training robots through demonstration is a common
technique in research labs, says Manuela Veloso, a robotics professor at
Carnegie Mellon University. But the technique has been slower to catch on in
the world of commercial robotics, she says. The only example on the market is
the two-armed Baxter robot, aimed at light manufacturing. It can be trained in
a new production line task by someone manually moving its arms to direct it
through the motions it needs to perform (see “This Robot Could Transform
Manufacturing”).
Sonia Chernova, an assistant professor in robotics at
Worcester Polytechnic Institute, says that most other industrial robot
companies are now working to add that type of learning to their own robots. But
she adds that training could be tricky for mobile robots, which typically have
to deal with more complex environments.
Izhikevich acknowledges that training a robot via
demonstration, while faster than programming it, produces less predictable
behavior. You wouldn’t want to use the technique to ensure that an autonomous
car could detect jaywalkers, for example, he says. But for many simple tasks,
it could be acceptable. “Missing 2 percent of the weeds or strawberries you
were supposed to pick is okay,” he says. “You can get them tomorrow.”
Comments
Post a Comment