¿ìè¶ÌÊÓƵ

Cog has arms, a body and a steady gaze – but no gender. ‘We deliberately
made Cog genderless,’ explains Lynn Stein, of the Artificial Intelligence
Laboratory at the Massachusetts Institute of Technology. ‘It’s called Cog
for cognition, but also for being a cog in the daily workings of the world.
That’s part of the philosophy of the project.’

Cog is a robot, and a very special one. The brainchild of MIT’s Rodney
Brooks, an AI researcher who has already won fame for designing tiny robots
that move like insects, Cog will not only be dexterous and mobile, as any
good robot should be. It will also be intelligent. That, at least, is the
hope of the researchers behind the ambitious, five-year Humanoid Project.

No other team has seriously tried to produce a robot with high-level
cognition and fine manipulative skills. What makes such an attempt possible
now is the emergence of a new and powerful mode of computing, known as massive
parallel processing, and the comparative ease with which computers can now
be linked. ‘The project is a combination of engineering and science,’ says
Brooks. ‘The engineering task is to build a robot that can operate effectively
in its physical environment. And the scientific task is to understand human
³¦´Ç²µ²Ô¾±³Ù¾±´Ç²Ô.’

These sound like lofty aims, and they are. But the real interest is
not so much the end of the journey as the journey itself. If, as Brooks
and Stein hope, Cog becomes intelligent, it will be because of its perceptual
experience of the world, not because it will have been equipped with a brain-in-a-box,
the stuff of classical AI. In other words, Cog will come to know about the
world as a child does – through learning about it.

Alien consciousness

The Humanoid Project is well off the conventional path of AI research.
‘We’re challenging the accepted approach of the science, and there are a
lot of people who expect – or even hope – that we’ll fail,’ says Brooks.
‘If it works, everyone will retreat and find other arguments with which
to criticise us. And if it fails, well, I’m sure I’ll find some excuse why.
But I don’t think it’s going to fail.’

The spectre of intelligent machines has stalked the pages of science
fiction ever since the genre was born. There is something about the creation
of an alien consciousness that seduces humans, siren-like, with the promise
of great discovery and danger. For the founders of the field, the advent
of electronic computers half a century ago seemed to offer a fast route
to turn fiction into fact. But that goal has proved more elusive than most
imagined.

For the past three decades, AI researchers have tended to view intelligence
as a centralised, disembodied function. This approach stemmed partly from
technological constraints of the kind of computation available – high-speed
serial processing – and partly from an intoxication with the early successes
of these serial computers. The Humanoid Project is a major departure from
this route. It returns AI research to the vision of one of computing’s greatest
pioneers, the British mathematician Alan Turing.

In a paper Turing wrote in 1948, but which was not published until 1970,
he argued that for a machine to develop at least some forms of intelligence,
it would need to be embodied. It would need to build intelligence through
perceptual experience, rather than having a model of the world programmed
into it. In 1950, Turing reiterated this view at the end of a much more
famous paper attempting to answer the question: how can you tell if a machine
is intelligent? In Turing’s test, a machine is said to be intelligent if
a person conversing with it via a computer terminal or other such device
cannot distinguish its responses from those of a human.

How would the machine acquire such intelligence? Turing described two
possible paths. One was the disembodied approach in which cognitive skills,
such as chess playing, would be programmed into the machine. The second,
and to Turing more appealing, path was the embodied approach, where an
intelligence would be allowed to develop through the experience of sight,
sound and touch. ‘Artificial intelligence followed the former path, and
has all but ignored the latter approach,’ observes Brooks.

Grandmasters

That was probably inevitable. In the early days of AI, researchers concentrated
on simulating tractable cognitive processes, such as chess playing. Here
serial computers are perfectly adequate. They can search rapidly and repeatedly
through millions of possible moves and pick the best sequence. So successful
has this approach become that one such program, Deep Thought, now plays
at grandmaster level. ‘No one pretends that this brute force, heuristic
search method is the way humans tackle chess problems,’ says Brooks. ‘But
the success of the approach with chess and other similar cognitive tasks
built up a methodological momentum.’

Human cognition, enigmatic though it is, is obviously generated by multiple,
parallel, mental activities that often lead to solutions as though by intuition
rather than through a grinding comparison of all possibilities. Because
the flow of electrical impulses between nerve cells in the brain is so much
slower than that in computers, human cognition must emerge as properties
of the connection of lots of relatively simple networks (otherwise thought
processes would take minutes or hours, rather than seconds). The brain’s
slow, parallel architecture contrasts starkly with the high-speed, narrow-channel
operations of serial computers.

As serial computers became ever more powerful, they were increasingly
capable of delivering the kind of performance that AI researchers identified
as progress towards creating a thinking machine. Speed and performance,
not conceptual insight, became the driving forces of the field. Gradually,
says Brooks, technologies based on serial rather than parallel computers
‘mistakenly became enshrined as principles, long after the original impetus
disappeared’. Intelligence was seen as computing power in a box, completely
disembodied and preprogrammed.

This approach is often referred to as good old-fashioned AI, or GOFAI.
According to GOFAI, intelligence can be broken down into discrete modules
with specific functions, such as perception, planning actions and executing
actions. These modules are equipped with explicit models of the external
world, and they interact to produce intelligent behaviour in a robot, such
as avoiding objects. In GOFAI and traditional cognitive science, there is
a physical and functional separation of ‘high-level’ mental operations,
such as cognition, and ‘low-level’ mental operations, such as the ones involved
in movement and the senses. Moreover, GOFAI assumes that all these mental
operations are performed by specialised centres in the brain. There is a
‘movement centre’, a ‘pain centre’, and so on.

Unfortunately, as research into cognitive psychology and neuroscience
are making increasingly clear, human intelligence is not organised this
way. The brain stores knowledge not only by category – animals, names,
geometric shapes – but also according to how the knowledge was acquired,
through hearing, touch, sight and so on. Contrary to GOFAI orthodoxy, there
is no great division in the brain between structures responsible for ‘high-level’
tasks, such as storing and organising information, and those responsible
for ‘low-level’ tasks, such as gathering information. There are only low-level
brain structures, out of whose combined talents cognition somehow emerges.
This being the case, observes Brooks, ‘traditional artificial intelligence
offers solutions to intelligence which bear almost no resemblance at all
to how the biological systems work.’

In the early 1980s, prompted by the growing gap between the assumptions
of GOFAI and the emerging understanding of human cognition and brain function,
a few AI researchers began to explore a new path. Some pursued the nature
of human intelligence through theoretical work and computer models based
on such things as neural networks. Others developed what has come to be
known as behaviour-based AI, invoking Turing’s notion of embodiment and
involving the construction of simple, mobile robots.

By their nature, robots are embodied. But so far they have not been
capable of high-level cognition. ‘The rationale for embodiment as a route
to cognition in AI is as bold as it is obvious,’ says Stein. ‘The fact that
we have bodies matters. It both constrains and enables the way we interpret
the world. The way the brain evolved and human cognition developed was predicated
on our interaction as individuals with the world.’ If that’s the way it
was for humans, and other animals, then perhaps that’s how it should be
for robots.

But behaviour-based AI cannot be a slave to biology because researchers
still do not know how even the simplest nerve networks generate behaviour.
Take the tiny roundworm Caenorhabditis elegans. While researchers know all
the positions of, and connections between, its 300-odd neurons, they have
only a sketchy picture of how these neurons generate the organism’s behaviour.
And the same is true of much smaller networks of neurons. The rhythmic digestive
system of the spiny lobster, for example, is controlled by a cluster of
28 nerve cells. This network and its neurotransmitters have been studied
in great detail. Yet nobody can say exactly how the rhythms are generated.

Against nature

The notion of building an intelligent machine by mimicking the neural
architecture of the human brain remains in the realm of science fiction.
But even if it were possible to emulate biology completely, it might not
be desirable. This is because natural selection has to work with whatever
starting material is available to it. Often the result is like a cobbled-together
contraption that no engineer would build, given the opportunity to start
from scratch. ‘Perhaps the solutions found for much of intelligence are
terribly suboptimal,’ says Brooks. ‘Their emulation may be a distraction.’

So the MIT researchers back a hybrid approach to AI. Biology is important,
they reason, because minds are shaped by the fact that the brain processes
different types of sensory information – from vision, sound or touch –
in different neural pathways. It is impossible to ignore that biological
fact, says Brooks. But while our pathways are made of networks of neurons,
a robot’s can be made out of microprocessors.

Earlier attempts to construct intelligent robots, in the late 1960s
and early 1970s, achieved some superficial success. There was the ‘Shakey’
project at the Stanford Research Institute and ‘copy-demo’ at MIT. The first
robot navigated around obstacles in a room, the second piled blocks according
to a model it was shown. But both operated in highly structured environments
and encapsulated the traditional AI notion of intelligence: they had a problem-solving
brain linked to some kind of physical interface with the world. By contrast,
one of the goals with the new generation of mobile robots, which began to
be designed in the mid-1980s, was behavioural flexibility, a mark of true
intelligence. A second goal was to abandon the idea that the brain consists
of a two-tier hierarchy of neural structures devoted to low and high-level
thinking.

Two guiding principles emerged. The orthodoxy that sees high-level thought
as the product of specialised ‘cognitive’ centres in the brain was replaced
by a new and revolutionary concept: that thought emerges from the interaction
of scores of humdrum perceptual centres in the brain. As a result, researchers
ditched artificial nervous systems built from specialised cognitive centres
in favour of ones with a ‘distributed architecture’. They also threw away
the idea of programming a robot’s brain with a preformed mental model of
the world. Instead, robots of the future would be able to generate their
own model of the world through their experien-ces of sight, sound and touch.

Brooks describes the artificial nervous systems in these mobile robots
as subsumed, or layered. ‘Each of the layers is a behaviour-producing piece
of network in its own right, although it may implicitly rely on the presence
of earlier pieces of network,’ he explains. ‘For instance, an ‘explore’
layer does not need to explicitly avoid obstacles, as the designer knows
that the existing ‘avoid’ layer will take care of it.’

Building more and more sophisticated robots through progressive layering,
says Brooks, is analogous to the way evolution has worked through time:
‘As with evolution, at every stage of the development the systems are tested.’
By the summer of last year, Brooks and his colleagues had built a zoo of
small robots that explore novel territory, build maps, interact with people
and navigate visually. But traditional AI researchers remained unimpressed.
‘People said that our systems had nothing to do with intelligence, because
they do only what insects do,’ recalls Brooks. ‘They said that real AI systems,
that is the traditional brain-in-a-box systems, are intelligent and would
one day do things like humans do.’

Brooks took that as a challenge, and so the Humanoid Project was born.
Cog is the most sophisticated embodiment yet of behaviour-based AI. In practice,
it will also test theories about human cognition. If the layered, distributed
structure of Cog’s artificial brain fails to generate intelligent behaviour,
then these theories may be wrong. It will take at least five years to find
out.

Cog didn’t have to look like a human for it to develop intelligence
through experience. But its creators wanted the interaction between robot
and people to be as natural as possible. Cog’s humanoid appearance, modelled
on one of the AI laboratory’s graduate students, makes this easier. The
plan, when Cog is built, is to have the robot sit at a table in the AI laboratory
and interact with people, perhaps by playing with toys, by stacking objects,
passing things back and forth and so on – very simple actions, like the
activities of a baby.

Although Cog’s view of the world will be built through experience, it
won’t have been placed in the world with a completely clean mental slate.
To do so would be contrary to what happens in nature. After all, humans
are not clean slates at birth, either. Many simple responses crucial to
early survival are pre-programmed in infants, such as the sucking response
and certain avoidance behaviours. And there is increasing evidence that
many sophisticated behaviours are extensively underwritten genetically.
Language acquisition is a prime example here. ‘Some value systems will be
built in,’ explains Stein. ‘And perhaps some basic emotion responses too,
and positive feedback systems for reinforcing certain behaviours.’

Recognising faces

If, by the end of the five-year project, Cog’s behaviour is comparable
with that of a human two-year-old, says Stein, ‘it will have been wildly
successful’. A two-year-old infant is beginning to talk, of course, but
there is no immediate plan to try to make Cog understand or acquire language.
But Stein hopes that Cog will at least come to recognise individual people’s
faces. ‘I don’t know whether this ability will emerge or will have to be
built in,’ admits Stein. ‘It’s important. I want Cog to know who we are,
or at least that I’m the same person I was yesterday.’

The most difficult problem ahead, however, is consciousness. A two-year-old
has an awareness of itself, but it is different from an adult’s. What about
Cog? ‘Thought and consciousness are epiphenomena of the process of being
in the world,’ argues Brooks. ‘As the complexity of processing to deal with
that complex world rises, we will see the same evidence of thought and consciousness
in our systems as we see in people.’

So, will Cog have thoughts and experience a degree of consciousness?
Will Cog be intelligent? Brooks admits that there is no way to measure this.
‘If people who interact with Cog believe it’s intelligent because of what
it does, then Cog is intelligent. If it seems to be conscious, then it
probably is. But that’s way in the future.’

It might seem bizarre to think of a machine being conscious, but many
see this as entirely reasonable. ‘Humans are machines and we’re conscious,’
says Daniel Dennett, a philosopher at Tufts University. ‘It’ll be important
to see how much simplification can be done, and still build something that’s
³¦´Ç²Ô²õ³¦¾±´Ç³Ü²õ.’

Others believe Cog, or machines like it, will never experience anything
resembling human consciousness. ‘The way we experience the world is the
result of historical accident, not of design,’ says Nicholas Humphrey, a
psychologist at the University of Cambridge. ‘The quality of human experience
is based on our ancestral history. Unless you repeat that evolutionary history
in the construction of an artificial brain, you won’t replicate the quality
of human experience.’

Cog’s consciousness, if such a phenomenon emerges, will have a quality
that reflects its cognitive architecture. Right now the MIT researchers
are developing the massive parallel processing on which Cog’s intelligence
will stand or fall. Unlike conven-tional serial processing, which is usually
aimed at producing an answer to a specific and tractable problem, parallel
processing lends itself to the kind of con-tinuous assess-ment of the world
that must underpin human learning.

‘Our humanoid robot will be situated in a real world over which it has
very little control,’ say Brooks and Stein. ‘There will be people present,
moving about, changing the physical environs of the humanoid, responding
to actions of the humanoid, and generating spontaneous behaviours themselves.
The task for the humanoid will be to interact with these ultimately unpredictable
agents in a coherent way.’

Cog is not the only robot in the world with human aspirations, although
it is the most ambitious. Researchers in other parts of the US, in Europe,
and in Japan are working with systems that combine dexterity with some kind
of intelligence. Probably the most impressive is WABOT-2, which, like Cog,
is humanoid in form (‘Can robots come to care for us?’, ¿ìè¶ÌÊÓƵ, 2
October 1993). Built by Ichiro Kato and his colleagues at Waseda University,
Tokyo, WABOT-2 can recognise music, both heard and seen, and play the piano
to concert standard. The robot is a stepping stone in Kato’s planned journey
to create humanoids that not only function in a human world, doing basic
jobs, but also interact on a personal, emotional level with people.

Currently, however, the piano-playing robots are just that: piano players.
Despite their dexterity at the keyboard and their knowledge of musical pieces
and ability to pick them up, the robots can’t do anything else. Acting together,
the robots’ 80 or so microprocessors can solve a specific and tractable
problem – learning music – but not the much more demanding one of learning
about a variable and unpredictibale environment. An artificial intelligence
could just as easily learn to read and understand music without a body as
with one, says Brooks: but the same is not true of learning in general.

In contrast to these robots, Cog’s artificial nervous system will not
so much dictate how its body moves as respond to its movements through a
feedback mechanism that enables learning to take place. Cog may turn out
to be a poor pianist, but that doesn’t matter to the MIT researchers. Their
dream is to see the robot acquire the broader intelligence of a two-year-old
infant.

But even if they succeed, how human will Cog be? Opinions vary. Its
performance may be limited simply because Cog has Cog’s mental wiring, not
the neural circuitry of an immature Homo sapiens. Humphrey says: ‘My guess
is that the robot will look more like a human than behave like one.’ Maybe
one day, if it eventually acquires language, Cog will have something to
say about that.

Further reading: Rodney A. Brooks, ‘New approaches to robotics’, Science,
vol 253, p 1227. Rodney A. Brooks and Lynn Andrea Stein, ‘Building brains
for bodies’, MIT, AI Memo No 1439, August 1993. Daniel C. Dennett, ‘The
practical requirement for making a conscious robot’, Proceedings of the
Royal Society, in press. Lynn Andrea Stein, ‘Imagination and situated cognition’,
Journal of Experimental and Theoretical Artificial Intelligence, in press.

* * *

Cog’s vital statistics

The point of making Cog as human-like as possible, with head, torso
and arms, is so that it can interact easily with real humans. It will need
a sense of balance, to be aware of symmetry and to coordinate the movement
of its head and body for effective vision. Cog will also experience a sense
of relief when it relaxes its body by sitting down.

The first version of Cog is being modelled on a rather small graduate
student in the AI laboratory. The range of movement around the neck and
hip will be human-like, although the speed of movement will be a little
slower. Later versions of Cog will have touch sensors integrated into the
torso.

Initially, Cog’s hands will be relatively simple, and will lack fine
movement. The arms, however, will be much more human-like, with redundant
‘degrees of freedom’, rather than the six seen in standard commercial robot
arms. Sensing in the hands and arms will be sophisticated, with conducting
rubber that allows touch. Strain gauges, heat sensors and current sensors
will allow Cog to have a feel of how its arms are being used and how they
are performing.

Cog will of course have two eyes, each of which consists of two tiny
cameras, one giving a broad field of view, the other a central field. The
images will be in black and white at first. Key challenges include keeping
the two eyes coordinated, especially as the head moves, and maintaining
the back and forth scanning that is essential for forming sharp images in
perspective.

Unlike humans, Cog will have three ears, or high-quality microphones
for ears. This is because the shape of the human ear is effective for determining
the direction of sound. An important part of Cog’s hearing will be to correlate
sounds with the visual events that generate them. This is a step towards
giving Cog the ability to focus on one sound out of a background of many.

Until Cog develops language – which is beyond the remit of the initial
five-year plan – it will be limited to producing simple noises.