THE YOUNGSTER eyes me suspiciously as I enter the room, its gaze following as
I cross the floor. Then after a while, it loses interest and turns back to its
toy dinosaur. But then I never was any good with kids.
When Rodney Brooks set out to build a humanoid robot with the intelligence of
a two-year-old child, he didn’t realise what he was letting himself in for. Six
years down the line, he and his team at the Massachusetts Institute of
Technology have transformed themselves from artificial intelligence (AI) experts
into the most unlikely bunch of developmental psychologists and nannies.
Colourful toys litter the labs, and much of their time is spent playing with
and entertaining their charges. This is because Cog and its alter ego Kismet are
the first of a new type of robot designed to behave in the same way as small
children. If you want to create a robot with the intelligence of a two-year-old,
Brooks reasoned, the best approach was to give it the innate abilities of a
newborn and let it develop.
Advertisement
Motors and software may control Cog and Kismet but it’s difficult to believe
that this is all that makes them tick. Cog may be only a head, arms and torso,
but as you watch it explore the world you can’t help feeling that something
profound is going on. It moves smoothly like a creature rather than with the
abrupt, mechanical movements of a machine. And its eyes dart from one object to
the next, with its head slowly bringing up the rear, as though it were
human.
Kismet is even more convincing. It may be just a head, but it has more
interesting facial features than its larger cousin, making its moods easier to
interpret. Complete with eyelids, ears and newly acquired lips, it has the
appearance of something young and cute, and reacts to events with an impressive
repertoire of doe-eyed expressions ranging from surprise and interest to sadness
and anxiety.
One of the unexpected findings from the project so far is that the appearance
of these human-like actions and reactions is not just an optional extra. The
robots, like children, will not develop unless their carers read more into their
behaviour than is actually there.
The idea of building a robot with human intelligence, even that of a child,
is not only ambitious, it’s highly unconventional. Most AI researchers confine
themselves to recreating a single sense, such as vision, or simple
behaviours— not the whole shebang. “The hard-core roboticists are not
comfortable with our work because we are not doing the same sorts of things,”
says Brooks. If it wasn’t that Cog does what it’s supposed to, he says, his team
would have been written off long ago.
But then Brooks has always been controversial. In his work on robot insects
in the 1980s, he rejected the idea of a central “brain” and showed how
intelligent behaviours could emerge from cooperation between a number of simple,
independent systems. Each leg of Genghis, a six-legged robot, for example, had
its own simple controller, and walking “emerged” by timing the actions of the
controllers.
Brooks also pioneered the idea of increasing the complexity of robot
behaviour by building up a hierarchy of simple systems. When “whiskers” fitted
to Genghis detected an obstacle, they modulated the signalling between the legs
in such a way that the robot walked over the obstacle as though it “knew” what
was before it.
His approach broke the traditional mould of AI, which tended to treat
intelligence as a problem that could be encoded in rules, and hence software.
Give a robot a software replica of its world to refer to, and a set of
instructions for negotiating that world, and it would appear to act
intelligently. By contrast, Brooks argued that robots need no internal
description of the world: for him, intelligent behaviours emerge only when a
robot exists in the world and interacts with it.
With the Cog project, Brooks wants to take his ideas further. “We’re trying
to build a Commander Data,” he says with a smile. He knows it’s a lofty aim, the
android from Star Trek: Next Generation is smarter by far than the
average life form.
Four eyes
At present, Cog is still a collection of isolated computer-controlled
systems. For eyes, it has four cameras, two for peripheral vision and two for
high-resolution, narrow viewing, which all move in their sockets in the same way
as human eyes. It also has a movable head, neck, arms and gripper-hands. In
addition, Cog has an auditory system that gives it enough information to know
where sounds come from, a basic sense of balance and the rudiments of a sense of
touch. Most of its motors come with position sensors to let Cog know where the
rest of its “body” is, and strain and temperature gauges so it doesn’t overload
anything.
These are the basic systems—like the legs and whiskers of
Genghis—which when they are connected to each other will allow Cog to
display intelligent behaviours. Already, it can find faces, tell if a person is
looking at it, detect movement, copy head gestures such as nodding, and even
play with a Slinky. “We also do some very basic auditory stream segregation, so
you can get the sound of my voice away from the fans in the background,” says
Brian Scassellati, a cognitive scientist and one of Cog’s principal
architects.
What makes Cog so different from its insect predecessors is that Brooks and
his team are not simply trying to build an increasingly complex system and watch
what emerges: they are trying to produce specific human behaviours.
Consider looking somebody in the eye. Cog does this in a series of steps.
First, it notices that someone is present by detecting motion in its peripheral
vision. Then it checks whether the moving object has a face, using an algorithm
that matches patterns of light and shade to a template made from shadows cast by
faces in different lighting conditions and orientations. Once it finds a face,
Cog’s eyes, then its head, turn towards the face and it matches the peripheral
image to the high-resolution image. This done, a second template picks out the
eyes within the image of the face. The behaviour appears lifelike and the robot
looks you in the eye in real time, without needing massive computing power.
In line with Brooks’s original ideas, all Cog’s actions and abilities are
discrete and built in hierarchies. The result is that new behaviours, such as
detecting a waving hand or recognising individual faces, can be built on top
without redesigning the existing systems.
Still, for the most part, Cog’s abilities are purely reactive. If it is ever
to fulfil Brooks’s dream of learning like a child, Cog is going to need a
memory, and so far the team has not tackled that problem (see “No past, no
future”). The researchers have, however, begun to look at one aspect of learning
that they originally took for granted, but which has proved to be crucial to
Cog’s development—social interaction.
As infants we learn gradually, expanding our abilities as our carers slowly
increase the complexity of the tasks they set us. This incremental approach
relies heavily on social interaction. It had always been part of the plan to
make Cog communicate with people, but its designers hadn’t counted on how
complicated a process this is. The robot must be able not only to understand the
intentions of its carer but also to impart its own intentions. This might be
straightforward but for the fact that Cog doesn’t have any intentions or know
how to express them—which makes it similar to very young babies.
Cynthia Breazeal, principal researcher on Kismet, has wrestled with this
problem. The solution, she believes, is that young babies are great at giving
the impression that there is more going on inside them than there really is.
Developmental psychologists argue that young babies are capable of showing only
a few, innate expressions. Yet parents assume from the beginning that their
babies behave in the way they do for meaningful reasons.
Parents interpret facial and vocal expressions as indicators of how a baby is
feeling—so if a child is content, parents tend to maintain their level of
interaction, but if the child appears uninterested or upset they may intensify
the interaction or change it, to regain the baby’s attention. Children learn
from the consistency of the parents’ reactions how to manipulate their parents,
and so gain attention. This puts them in an ideal position to carry on
learning.
These ideas are embodied in Cog’s cousin. “Kismet takes advantage of the way
we are programmed to interact with small children,” says Breazeal. It was
designed to emotionally blackmail people. If you don’t do what it wants it
scowls or looks sad: if you please it, it rewards you with a smile or a look of
interest.
Kismet is actually a platform for testing behavioural systems before
installing them in Cog. It has a vision system similar to Cog’s, able to detect
motion and faces. But Kismet also has what Breazeal calls a motivational system,
which tries to keep the robot in a happy, interested state. This consists of a
collection of drives, such as the urges to be social and stimulated, and
behaviours that satisfy those drives (see Diagram). The intensities of the
drives, which dictate the expression on Kismet’s face, increase if they are not
satisfied and decrease when the appropriate behaviours are operating.
So, if Kismet is left alone, the intensity of its social drive increases.
This makes it look sad, communicating to anyone passing that it craves
attention, and activating a behaviour called socialise. As soon as it detects a
face, it fixes on it and begins to socialise—sadness changes to happiness
or interest and the intensity of its social drive begins to fall. If, however,
the interaction is too intense and the social drive is pushed too far, the robot
becomes overwhelmed and gives a look of displeasure.
Similar events take place with Kismet’s “stimulation” drive, which is
counterbalanced by a behaviour called play. At present, Kismet’s favourite
object is its toy inchworm. If left alone, the robot looks sad and its play
behaviour is activated, but bounce the worm and it starts to look happy and the
desire to be stimulated begins to fall. If you bounce the worm too fast,
however, Kismet looks disgusted. And if you carry on, then disgust changes to
anger, or it may simply close its eyes and sleep to allow its “brain” to catch
up. In an added refinement, if you repeat the same movement repeatedly then
Kismet gets bored and looks sad again.
So humans watching Kismet’s expressions take them as signs of how it is
“feeling” and modify their behaviour to restore it to a contented
state—just as they would with a child. Eventually, its happy and
interested expressions will signify that it is absorbing information at the
optimum rate. So, Kismet’s expressions will regulate people’s actions to let it
learn at that pace.
The team is now working on a system that will allow Kismet to respond to the
inflection in people’s voices, so that its emotional state can be changed by
auditory as well as visual stimuli. And soon, says Scassellati, they will
introduce a vocal system so it can babble like a baby. This will let Kismet
attract people’s attention even when they’re not looking at it, and let them
know if it is happy or sad by cooing or crying. Eventually, with some form of
software capable of learning, the vocal system could play a part in helping
Kismet to develop language.
To some, Breazeal’s approach might seem pointless: after all, what could
bouncing a toy inchworm teach a robot about the world? The same, however, could
be argued of infants, and yet it clearly works for them. This sort of exercise
teaches infants more about themselves than their environment, such as how to
reach for an object.
It also gives them the opportunity to forge links between different sensory
perceptions. Children might learn, for example, to associate neural signals for
bright colours with those for movement, thereby realising that the bright thing
and the moving thing they are seeing are one and the same. Breazeal and the team
hope to see Cog learning in the same way.
Such basic skills are essential for children because they pave the way for
more complex tasks later. To appreciate this, take the seemingly simple job of
understanding what someone is referring to when they point at an object. Only
other apes and dolphins are able to grasp that there’s something “over there”
worth looking at, and then find the object of interest. This is one form of a
skill, called joint attention, which Cog needs to have because so much social
interaction depends upon it, says Scassellati.
He hopes to give Cog this ability by following the idea that joint attention
is built of simpler skills. For infants to grasp the meaning of pointing, they
first have to develop the skill of knowing when someone is looking at them, and
then learn to follow that person’s gaze and finger. Evidence for this composite
nature comes from observing child development.
While most three-month-olds have developed eye contact, it is not until nine
months that they follow another person’s gaze, and 18 months that they follow
someone’s gaze out of their field of view. And before they can single out the
object being pointed at, it seems that infants first go through a phase of
following the person’s gaze and finger, and fixating on the first object their
eyes see. Cog already has some of the basic skills needed for joint attention,
such as the ability to maintain eye contact. It can also learn to point at an
object. “Once we have enough of these rudimentary social pieces then we can
actually start to learn from people,” says Scassellati.
And this is really the crucial point about the Cog project. It is not about
wiring the robot up and turning it on. But rather about gradually increasing
Cog’s skills, and watching as the richness and complexity of its behaviour
increases, hoping that one day “intelligence” will emerge.
In the next few months the team plans to carry out psychological tests to see
if Kismet can manipulate its carers as well as children manipulate theirs. A
number of new heads are also on the way, including one that will transfer
Kismet’s abilities to Cog, which should make it appear more like a child than
ever.
Already, the robots seem so human that there’s a strong temptation to think
of them as male or female. This is even though the Cog team has worked hard to
keep its charges sexless. Still, if the researchers are ever to achieve their
goal of making a robot that interacts without making people feel uncomfortable,
then perhaps they need to address this issue. We humans, after all, are either
one thing or the other.
OF THE problems still to be faced by Rodney Brooks and his team at MIT, some
of the thorniest involve memory. A robot built using the traditional techniques
of AI would have a model of the world built into its software controls. So when
it saw something red, the robot would “know” it was red because its software
would tell it so. Likewise, if it needed to remember something—the
position of an object, say—an instruction would tell it to do so.
Cog has no such model or instructions because the aim is let it choose what
to remember. So how will it make this decision? Or know that it’s seeing red?
And how will the concept of redness be represented in the robot’s memory? “In
traditional AI you would never think of this as being a problem,” says Brooks’s
colleague, Brian Scassellati. But for Cog these are major stumbling blocks that
the team have yet to address.
Other things that we take for granted become real difficulties for Cog.
Without a sense of time, for example, Cog will not be able to order its thoughts
or know the difference between past and present. “There are so many problems,”
Scassellati sighs.
But then, a few years ago, social interaction seemed a colossal problem. And
today, the team’s approach of breaking behaviours down into simpler steps is
starting to pay dividends. The question is whether storing and retrieving
memories will fall to the same approach.
No past, no future
-
Further reading:
More information about the Cog team, their research papers and the robots is available at
www.ai.mit.edu/projects/cog/ -
Videos of Cog and Kismet in action are at
www.ai.mit.edu/projects/cog/video_index.html -
Birth of a human computer
by Roger Lewin, żěè¶ĚĘÓƵ, 14 May 1994, p 26