through hands-on demonstration and practical assistance until
they can reliably practice and improve their skill on their own.”
Once a human demonstrates how to play the game, the
machine can figure it out much faster than if it was exposed to a
‘trial-and-error’ approach used in other types of machine learning.
It doesn’t have to play millions of games in order to understand
the rules. Instead, the human shows it a goal and the tools it can
use to get there.
Osaro is currently teaching robots to play video games using
a simulation environment called Gazebo, in which commercial
industrial robot manufacturers can expose models of physical
machines to virtual environments. A camera in the virtual
environment allows the software to see a robot that conforms
to the exact physical constraints of the manufacturing robot the
company will use.
“Our software doesn’t know the difference between moving
a robotic arm and playing a game where you have to move a
robot arm,” Pridmore says.
This simulation environment solves one of the problems that
crop up because of the differences between a virtual assembly
line and a real one. In a game, there are only so many controls
– for example, left, right, up, or down. In a factory, a robot arm
can set the angle of any particular joint in a variety of ways.
TWO ARMS ARE BETTER THAN ONE
In the real world “there are nuances as to whether your
control is discretized or continuous,” Pridmore says. “It’s
a much higher dimensional problem, so it requires more
sophistication with the algorithms.”
Over the next few months, Osaro is going to be working with
customers on teaching physical robots how to do things using
deep reinforcement learning.
Pridmore also looks ahead to a future of manufacturing robots,
and sees technology like Osaro’s speeding up deployment and
enabling robots to learn faster. Next, robots will be taught to
perform tasks that are currently impossible for humans, such as
flexible component assembly.
Finally, more maneuverable control policies will allow robots
to use two limbs and pick up a product instead of manipulate
it on a platform. Some collaborative robots, like those from
Re Think Robotics and ABB’s YuMi, already use two limbs.
Combining this with deep learning could allow for more
adaptable, flexible industrial robots.
Osaro is also interested in contributing its deep learning
technology to industries like autonomous vehicles, drones,
software development, healthcare, gaming, advertising, and
“Most machine learning companies are still focused on
exploring deep learning, which addresses only part of the
problem: enabling machines to perceive,” Pridmore says.
“That’s important, but most industries need more. Osaro’s deep
reinforcement learning technology fills this gap – we enable
products that move beyond merely processing images to
automatically take informed, complex actions.”
Get the results that only the right tool can give
When you’re ready to automate your business, choose the automation tool that fts your needs exactly.
Collaborative robot arms from Universal Robots work safely side-by-side with your employees.
These innovative automation tools are lightweight and easy to move and program, so you can optimize
production for processes from gluing, welding and assembly to machine tending and pick-and-place.
And Universal Robots offers the fastest payback in the industry.
Get the results that only the right tool can provide. Book a demo at your local distributor or read more at
Choose the right tool