Investors like Yahoo founder Jerry Yang and PayPal
founder Peter Thiel have backed Osaro so far.
“What most people think of as ‘artificial intelligence’
– even deep learning technologies – fail to go beyond
basic pattern recognition to replicate intelligent, decision-
making skills,” says Osaro CEO Itamar Arel. “Our deep
reinforcement learning approach allows human mentors to
teach computers and robots how to take intelligent actions
In December, President and COO Derik Pridmore talked
about what deep reinforcement learning is and how it works.
Ideally, deep reinforcement learning will make robots
easier to deploy, he says. Set-up time will be reduced,
because the robots will be able to understand a goal
instead of simply performing pre-set motions.
“The basic idea is to make robot control really fast and
easy to deploy,” Pridmore says. “That means you want them
to automatically interpret what they’re seeing (in this case,
seeing could be input from touch sensors and things, but in
general, processing perceptual data) and then automatically
acting as well.”
Factory robots typically learn to make movements through
imitation. They mimic the movements of a human operator
using a recording, or are programmed to move to certain
fixed points in space. Even if the robot is presented with an
object of a different shape than the ones it has worked with
before, it will still try to hit the same points.
That’s where deep reinforcement learning comes in. The
software generalizes, Pridmore says, understanding the
goal of a task instead of parroting movements.
LEARNING WITH VIDEO GAMES
In order to teach machines how to navigate the world in
this way, Osaro lets the robots play video games.
“We can compare the problem of learning complex control
to teaching a child how to ride a bike,” says Arel. “Existing
DRL architectures attempt to learn from a blank slate, which
is like providing a bike to a child and walking away – hoping
the child will eventually develop bike-riding proficiency through
pure trial-and-error. In the real world, we would guide the
Artificial intelligence and deep reinforcement learning startup Osaro launched on Dec. 2, 2015
with $3.3 million in funding and the goal
of making factory robots smarter by
teaching them to learn from humans.
By Megan Crouse, Real-Time Digital Reporter
TEACHING ROBOTS TO
LEARN FROM HUMANS