Robots are getting better at teaching other robots how to do things. Oh.
Robots are getting better at teaching other robots how to
do things. Oh.
BY STAN SCHROEDER May 10, 2017
Teaching a robot how to do something is usually done by
either programming it to perform a specific task, or demonstrating that task
for the robot to observe and imitate. The latter method, however, so far hasn't
been accurate enough for robots to be able to transfer their knowledge to other
robots.
That's changing, however, thanks to researchers from
MIT’s Computer Science and Artificial Intelligence Laboratory (CSAIL) and a
their new teaching method, called C-LEARN. That could have far-reaching
consequences by making it easier for non-programmers to teach robots how to
perform certain tasks. Even better, it allows robots to teach other robots how
to perform the same tasks.
The system does this by giving the robot a knowledge base
with information on how to reach and grab different objects. Then, using a 3D
interface, the robot is shown a single demo on how to, say, pick up a cylinder
or open a door. The task is divided into important moments called
"keyframes" — steps that robot needs to take in order to correctly
perform the task.
The test bed for C-LEARN is a small, two-armed
bomb-disposal robot called Optimus. Once Optimus learns how to perform a task,
it can transfer that knowledge to Atlas, a six-foot-tall, 400-pound robot (we
wrote about Atlas several times in the past, here and here).
“By combining the intuitiveness of learning from
demonstration with the precision of motion-planning algorithms, this approach
can help robots do new types of tasks that they haven’t been able to learn
before, like multi-step assembly using both of their arms,” Claudia
Pérez-D’Arpino, a PhD student who wrote a paper on C-LEARN together with MIT
professor Julie Shah, said in a statement.
The approach, D'Arpino claims, is similar to how humans
learn — they take what they already know and tie that information to a
demonstration of how something is done. “We can’t magically learn from a single
demonstration, so we take new information and match it to previous knowledge
about our environment," she said.
And that's the point — teaching robots in a similar way
we teach humans to do things would make the process faster and easier. Right
now, the C-LEARN method can't handle certain problems, including avoiding
collisions, but the team hopes that the system can be further advanced by
implementing more human-like learning capabilities.
Comments
Post a Comment