Teaching Bert: iCub Robot Learns About the World

This is Bert. Bert is an iCub. iCub is a very sophisticated humanoid robot. It is built by a European consortium. And there are different robot cubs in
various laboratories throughout the world. We happen to have the only one in the
Western Hemisphere. We call our iCub Bert. Bert is short for Bertrand Russell.
We have always named all of our robots after famous people who have contributed to the early
days of cybernetics. My research his about language and getting machines to interact with humans via natural spoken language. Bert has human architecture in many significant ways; particularly dexterous hands…
arms and legs… very sophisticated eye movements…
balance control… and, to all intents and purposes,
Bert can mimic almost any action that a human can do. Onboard, Bert has 13 microcontrollers to control the various motors and
mechanical linkages. Bert also has the human senses, including binocular color vision
and binaural hearing. And we think that all these things are
going to be necessary in order for Bert to ultimately build a
model of the external world that would
support natural language communication. The approach that we’re using is
’embodied cognition.’ The idea is that, the closer you get to
being able to experience the world as a human does, the more likely you will be able to
produce a human-like model of the external world. What’s really going on in natural language communication is that people are making internal mental
models of the external world and then encoding that model in natural language. So let’s take a simple example. How could you possibly know what it
means to balance two ideas or to balance two options against each other if you don’t know how to physically balance your physical being in the world? And natural language is
just simply loaded with words that we use in our everyday language that
have to do with spatial reasoning. And the best way to learn and understand how to reason spatially is by having a
body that moves and lives in a three-dimensional space. A human instructor can go up to Bert and grasp his hand and move it in a circular motion. And then you can say to Bert,
“This is a circle.” And then Bert learns what the circle is. And then, later on, if you say to Bert,
“Draw a circle,” then he can reproduce the behavior. What an engineer might do in order to
produce this capability is to simply take a commercial speech
recognizer, put it inside Bert, and connect it up to Bert’s ability to measure all its joint
angles. And then, every time you said, “Circle,”
Bert would simply play-back the exact recorded motions. That is absolutely not what’s going on
in this experiment. Bert is building a linguistic model of the speech that it hears and it’s
building a mechanical model of the motions that it produces. And, exactly like humans, every time Bert
produces a gesture, it’s slightly different. The models that humans are capable of
making are imperfect. And we are giving Bert the same disadvantage. [BALL HITTING FLOOR] INSTRUCTOR:
“Uhg!” [Laughing] The ultimate goal is to take all the things that Bert
learns about spatial movements and to have him be able to take those
models apart and reassemble them; so that he could
produce novel gestures, novel behaviors, and even novel
linguistic expressions. Now, we think that this is what humans
ultimately learn to do and it seems to us that that their
ought to be mathematical representations of this
process. Nature has already given us a solution in the form of ourselves and our bodies. And, if the robot could learn the same
thing and represent the problems and their solutions in the
same way, then we would learn how to make other
artificial devices that would be able to solve problems in
very specific contexts. And the applications of this kind of
knowledge are almost limitless. [MUSIC]

, , , , , , , , , , , , , , , , , , , , , , , ,

Post navigation

One thought on “Teaching Bert: iCub Robot Learns About the World

Leave a Reply

Your email address will not be published. Required fields are marked *