MECHANIMALS
*
In the continuing quest for artificial intelligence, researchers around
the
world believe that a little child, or even a lower-order life form, will
lead the
way to a new breed of robot workers and warriors
The little three-wheeled robot
is the size of a cigar box. It has only two
missions in life. When it hears noise, it
hides in a corner of the room.
When all is quiet, it scampers around looking for
light, then races to the
source. It may sound pretty mundane, but this
sound-aversive,
light-attracted creature represents the cutting edge of artificial
intelligence.
In the near term, this little robot may form the basis of
a vacuum cleaner that
stays out of the way and works automatically only
when no one is in the room. In
the long term, it could lead to robot
workers that build cities, fight wars, or
install scientific monitoring
equipment in space. When you think about artificial
intelligence, or AI,
what comes to mind is usually some Hal-like omniscient device
capable of
beating chess masters at their own game, provide insight into the
most
subtle bits of science or history, write poetry or tell jokes, and,
suitably
equipped with the right end-effectors (arms and such), fix a
watch or cook a
souffle--a device that can think, create, and even
feel. In other words, a perfect
imitation of a human being.
Unfortunately, real AI efforts have fallen short of
these lofty goals.
The reason, many experts now say, is that in the race to create
higher life
forms, AI researchers have lost sight of the basics. At a recent
powwow
in Paris, in fact, the critics declared that instead of tackling the
highest
levels of human mentation, AIers should be examining fundamental
questions of
perception, learning, and adaptation. Scientists at the
meeting, the International
Conference on the Simulation of Adaptive
Behavior, even had a name for the lower
AI creatures they hope to
build-animats, tiny artificial animals that will
scamper across laboratory
floors or computer screens. According to researchers,
they will study
these frisky critters in their "natural habitats, " much the same
way an
entomologist, biologist, or psychologist might study an animal or human
being. As AI experts come to understand these artificial animals, they
will
literally help them to evolve into more sophisticated forms. The
result, says
Stewart Wilson, the scientist at the Rowland Institute for
Science in Cambridge,
Massachusetts, who coined the term animat, will be
truly reactive creatures far more
intelligent than any AI systems that
have come before.
The idea for the animat
actually dates back to 1950, when the
legendary Alan Turing laid out two possible
approaches to machine-based
intelligence. In one approach, Turing suggested
providing machines with
the best sense organs that money could buy, then literally
teaching
them to understand and even speak English. "This process," Wilson says,
"follows the normal teaching of a child."
In the early years, both approaches
competed for the brightest minds
and the biggest funding. But by the Sixties,
researchers modeling
specific human abilities had won out over attempts to
understand natural
intelligence at a more primitive level.
During the Seventies,
however, AI efforts hit the wall: A lot of money had
gone into the work, but little
real-world value had come out. The
knowledge-based programs, although powerfully
capable in very specific,
rigidly controlled (and artificial) environments, had big
problems in the
real world. Their emphasis on exact reasoning at the expense of
perception and adaptation made them brittle and arbitrary. In other
words, they
found it impossible to operate in domains even slightly
different from the ones
for which they were programmed, and their
internal reasoning bore no relationship
to the physical world.
The animat-ors are now attempting to escape the trap of
mimicking
high-level mental competence by turning to Turing's second suggestion-the
child machine, situated from the start in a real sensory environment and
programmed to learn through experience. Today's animats, whether
simulated or real,
are of course closer to Pac-Men, bacteria, or, at
their most evolved, simpler
animals like ants. The basic strategy, by
definition, is to work up to
sophisticated intelligence from below.
Toward that end, Wilson says, researchers
take the holistic approach.
"The animats, like animals, exist in realistic
environments," he says.
"And as they evolve, they cope with more and more
problems presented
by the natural world."
To help the critters negotiate worldly
terrain, animat developers like
MIT's Rodney Brooks have endowed them with a
number of task-specific
computerized boxes. Each box produces a straightforward
behavior, from
following walls to avoiding collisions to moving toward light sources.
As
Brooks's creatures go about their business, they seem to be planning
and
learning. But as Brooks points out, "You can't point to one place
in the
code-any of the specific boxes-and say that is where this
higher-level brain
activity takes place. Rather, the more sophisticated
activity emerges from the
interaction of all the simpler parts."
One of the most sophisticated characteristics
being programmed into
animats is the ability to evolve. Federico Cecconi and Domenico
Parisi of
the Institute of Psychology in Rome, for instance, have created a
computer
simulation based on two types of simple organisms-one that can
grasp an object
only at a precise reaching distance, and a second that
can move to approach an
object and extend its arm to grab it. When
the simulation starts, the population
contains individuals with randomly
different levels of grasping proficiency. After
Cecconi and Parisi's
system has run for one generation, only those with the best
performances are allowed to reproduce, making copies of themselves for
the next
generation. Each time a copy is made, the computer alters it
slightly, mimicking the
random genetic mutations of real evolution.
The simulation works like life itself.
After evolving for 65 generations,
the population on-screen has an abundance of
successful graspers. What's
more, the ability to learn is programmed into more
offspring as a result.
To truly perfect the art of grasping, of course, animats will
need to
recognize the object to begin with. That's why one of the most
important
challenges facing future animat builders will be the
development of intelligent
sensors that recognize patterns and even the
meaning of events.
With this goal in
mind, AI student Dave Cliff at the University of
Sussex, England, is developing
a computer model of the hover fly's
eye. The interesting thing about the hover
fly, says Cliff, is that it
focuses on information at the center of its visual
field, paying less
attention to images at the periphery. Cliff's model further
narrows
down visual signals by eliminating redundant information, thus revealing
underlying
patterns in the images.
To the hover fly's eye, Cliff says, "the world looks like
a Mondrian
painting. The eye detects only the edges between blocks of color.
Without all the redundancy, you save valuable resources, such as
processing time,
that you would otherwise waste on things You already
know."
Cliff's hovermat eye,
thus far a simulation on a computer screen, can
see where it's going and correct
its course itself. Once this eye or
one like it has been built and fit onto an
animat, it will help the
creature navigate the world. In fact, because the
hovermat eye is so
intelligent, the animat brain processing the visual input can be
a little
less smart.
Even with such smart sensors, however, animats will have
trouble
negotiating the world without inner motivations of their own. That, at
least,
is the view of Pattie Maes, a researcher at the AI laboratory at
the University
of Brussels in Belgium and the AI laboratory at MIT In
her new computer program,
motivated creatures have a range of drives
from eating, drinking, and sleeping to
fighting, fleeing, and exploring.
These drives help them face a complex
environment that includes food,
water, obstacles, and other animals. According to
Maes, behavior patterns
emerge as the computer animals react to conflicting drives
in the face of
varied and complex situations.
The animals are motivated toward any
given behavior, Maes says, by what
is called an activation level. For instance, the
presence of food will
activate eating; eating, in turn, will activate behaviors such
as drinking
or sleeping. Yet other behaviors may be deactivated, or inhibited. For
instance, fighting will inhibit the drive to flee. In this way, Maes
says, the
computer animals are literally motivated by internal and
external factors to behave
in a variety of complex ways.
In the future, motivated animats like those designed
by Maes will leap
from the computer screen in 3-D, nuts-and-bolts splendor to
sense the
environment, adapt to the world, and travel around at will. These smart
and flexible creatures may eventually clean our homes, deactivate bombs,
and
traverse the endless void of space. Whether their tasks are mundane
or downright
dangerous, the animats are sure to become our trusty
assistants. They will
ultimately perform in ways no conventional robot
could master, releasing us from
the many unpleasant activities we
reluctantly accept today. And judging from their
capabilities, who knows
how far evolution will take them? Some pundits suggest
that, let loose
in space or under the sea, advanced animats might possess the
independence and adaptive ability to develop vibrant colonies, even
civilizations, of
their own.
PHOTO: Inspired by the notion of robots modeled after the myriad creatures in
nature, artist and animal systematician Louis Bec has generated a new body of
work-as he
calls it, a zoology of change. Bec's colorful, futuristic creatures,
shown above, were
spawned by the artist along the border zone of technology and
biology, imagination and art.
The work of artist Louis Bec suggests that the
animat may one day be as honed for survival
as lions, leopards, bacteria. Bec
says that his array of creatures, including those shown
above, illustrate the
variety of creatures scientists may create through the powerful
technique of
technomorphogenesis, when technology and biology join hands.
~~~~~~~~
By TOM
DWORETZKY
Inset Article
OMNI'S CATALOG OF ANIMATS
*
Animats in AI labs around the world
are currently simple creatures. But
in the future, they will find their way into
our everyday lives. To
learn about future animat species, we've polled the experts
and gazed
into our own crystal ball. A collection of potential animats follows
below:
*
PETMATS
*
The first animats may be toy pets that perceive the environment
through
sensors modeled on those of real animals. Their individual
behavior patterns will
evolve during their own lifetimes, as their
neural networks build up memories of
past situations.
*
Catmats, the first petmats to be built, will be designed with
eyes
that are especially acute at night. These animats will perform a double
duty
as both affectionate pets and stalkers of household pests such
as rodents and
roaches. To accomplish the latter task, their neural
networks will help them
identify unwanted animals, as well as each
species' patterns of flight and
avoidance,
*
With different neural circuitry, dogmats will play catch and hunt
animals
such as rabbits. Components of their brain circuits will make
the sense of
smell the most acute of their senses.
*
Birdmats, a possibility for future amusement
parks, will fly in flocks
according to primitive behavior patterns that make them
follow the bird
directly in front of them. To accomplish this, their eyes will
work
much like those of the hover fly.
*
DOZERMATS
*
Some day hordes of tiny,
bulldozerlike autonomous robots may construct
vast habitats for humans. These could
be as small as an inch on a
side and powered by solar cells. They could even
create a moon base,
for example, a suggestion put forward by MIT's Brooks. His
dozermats
would carry out only simple actions, such as digging and picking up
piles of dirt. Together, however, they would proceed, like ants building
their
hills, to construct vast, intricate structures.
*
Dozermats could also find uses on
Earth, building and maintaining roads.
*
Variations on dozermats, equipped with
different attachments, could also
function as robotic loggers and farmers. Their
mental circuitry could
identify trees and plants appropriate for harvesting,
navigate difficult
terrain, plow, and plant new growth.
*
WORKERMATS
*
Applications for
animats would crop up in environments from the home to
the bottom of the sea.
Workermats could carry your drinks from the
kitchen to the living room but would
also have many more important
functions. Consider a contaminated nuclear power
plant. Workermats could
withstand high heat and lethal radiation to perform their
cleanup duties.
Their bottom-up mental logic would allow them to navigate the
unpredictable
terrain of a severely damaged facility with much greater
adaptability than any
robot relying on a preprogrammed floor plan. They
would be equipped with radiation
detectors, vision sensors, and
pressure-sensitive "skin" to alert them to potential
obstructions.
*
The workermat's brain would consist of a variety of modules, each
motivated to survive and accomplish tasks. Specific mental
components-collision
avoidance, power replenishment, and debris collection,
for instance--would run
parallel, each competing for control over the
wheels and arms of the animat.
*
BOOBTUBEMATS
*
These are a favorite fantasy of MIT's Brooks. He envisions tiny
robots
that would live in a corner of the screen when the set was on. When it
was
turned off, they would scurry around on its surface cleaning the
dust. They would
be cheap, perhaps a few dollars for a vial of 50 or 100
of the creatures. For
power, the boobtubemats would have their mental
functions broken down into simple
behavioral units that competed for
control of the entire creature. These behaviors
would include collecting
dust, avoiding collisions, and recharging power packs.
Boobtubemats would
differ from traditional AI robots in that they would have no
internal
representation of the TV screen and would not actually know that their
combined behaviors worked to clean the set.
*
VACUUMMATS
*
A robotic cleaner the size
of a cat might be just the thing to cut
down on human-powered housework. It
would hide in a corner of a room
whenever someone was moving around, then go
around sweeping when alone.
After collecting a sufficient pile of debris, it would
take it to a
receptacle for future pickup by an artificial creature programmed to
move through the house emptying the trash cans. For these tasks, the
vacuummat
would need sensors that could detect sound, motion, and
obstacles. It would also
need behavior modules for avoiding obstacles.
*
AUTOMATS
*
Imagine stepping into
your car, telling it where to go, then sitting back
and letting it drive. Such a
vehicle would be equipped with sufficient
artificial intelligence to plot a course and
follow it efficiently,
avoiding obstacles and other automats. Some versions might
look like
today's cars, but others might have sets of legs to negotiate rough
off-road
terrain. The internal logic of such a device would allow it to
sense velocity and
to maintain safe speeds, using inhibition modules to
monitor traffic, balance, and road
conditions.
*
SOLDIERMATS
*
Take a workermat and give it a gun and you've got a
soldiermat.
Additional mental components would be added to permit the animat to
differentiate between friend and foe and to aim its weapon at the enemy.
Navigation, obstacle avoidance, balance, and other mental components
standard to
animats would make this a formidable-and expendable-adversary.
Such an animat could
fight on modern high-risk atomic, biological, and
chemical battlefields.
*
LABANIMALMATS
*
Living only in the memory banks of computers, these virtual
creatures
would be another breed of animat. They would be carefully modeled to
mimic the precise physiology of various animals and humans. Using these
animats,
scientists could run simulations of drug tests without risking human
life or
sacrificing real creatures. Researchers could also use populations of
labanimalmats to
study evolution, epidemiology, and ecology.