National Public Radio said it well: How can robots learn new
tasks? Practice, practice practice.
Stefanie
Tellex, Brown University assistant professor in computer science department,
talked to NPR recently about getting robots to do better at real-world
tasks—specifically, grasping, why grasping is hard for a robot, and what her
team would like to do about it.
Tellex works with a "Baxter," the industrial robot made by Rethink Robotics with its
box-like torso and arms.
Joe Palca, NPR science correspondent, watched the Baxter try to
pick up a battery; he said watching it was a little like watching paint dry. He
told her this looked pathetic. She laughed, knowingly.
Stuff that is really hard for a robot to do is almost effortless
for a person to do. A human spends little conscious awareness of exercising
perception, planning and control in picking up an object. But the robot from
scratch doesn't know from kazoos, batteries and pens.
It gets information from its cameras. And that information, he
said, is just a bunch of numbers.
Palca said Tellex thinks the way robots will get better and faster
in picking up unfamiliar objects is to give them programs which let them learn
from experience just as a child would.
Unsurprisingly, Tellex has had her Baxter working around the
clock. It picks up objects. It puts them down. Repeat. Repeat. And so on. But
Tellex also has an idea for speeding up the learning curve.
She
hopes to recruit some more Baxter robots elsewhere which are left idle
off-hours during robotics research projects to do the same tasks as her Baxter
to speed up the process. Pace said it evokes the saying, many hands do light
work.
Simply put, said Will Knight of MIT
Technology Review, "hundreds of robots could accelerate the process by
sharing knowledge."
Just how fast would this learning curve speed up learning?
According to Knight, Tellex counted around 300 Baxter robots in
research labs around the world and if each of those robots were to use both
arms to examine new objects, it would be possible for them to learn to grasp a million
objects in 11 days.
The
concept of a robot teaching itself how to do tasks is of interest among
researchers beyond Brown. Last month, MIT
Technology Review looked at
the work of Lerrel Pinto and Abhinav Gupta at Carnegie Mellon University. They
worked with a Baxter, giving it deep learning capabilities, placing it in front
of a table with
different sized objects and left it to learn how to grasp them.
The robot was left for up to 10 hours a day. If the robot dropped an object on the floor, there
were others it could continue with.
Lab
work dedicated to robots gaining better grasping skills may result in robots
that will be part of daily life.
Tellex said, "In twenty years, every home will have a personal robot which can perform tasks such as clearing the dinner table, doing
laundry, and preparing dinner. As these machines become more powerful and more
autonomous, it is critical to develop methods for enabling people to tell them
what to do."
