Mobile Robots and General Intelligence

                             Hans Moravec
                      Carnegie Mellon University
                             November 1986


\section{Introduction}
	The significance of mobile robot research may be much greater
than the sum of its applications. There is a parallel between the
evolution of intelligent living organisms and the development of robots.
Many of the real-world constraints that shaped life by favoring one
kind of change over another in the contest for survival also affect the
viability of robot characteristics. To a large extent the incremental
paths of development pioneered by living things are being followed by
their technological imitators. Given this, there are lessons to be learned
from the diversity of life. One is that mobile organisms tend to evolve in
the direction of general intelligence, immobile ones do not.  The plants
are an example of the latter case, vertebrates an example of the former.
An especially dramatic contrast is provided in an invertebrate phylum,
the molluscs. Most are shellfish like clams and oysters that move little
and have tiny nervous systems and behaviors more like plants than like
animals. Yet they have relatives, the cephalopods, like octopus and squid,
that are mobile and have independently developed many of the characteristics
of vertebrates, including imaging eyes, large nervous systems and very
interesting behaviour, including major problem solving abilities.

\section{Mobility and Intelligence in Nature}
	Two billion years ago our unicelled ancestors parted genetic company
with the plants. By accident of energetics and heritage, large plants now
live their lives fixed in place. Awesomely effective in their own right, the
plants have no apparent inclinations towards intelligence; a piece of
negative evidence that supports my thesis that mobility is a parent of this
trait.

\begin{figure}
\vspace{7in}
\caption[Intelligence]{\label{Intelligence}
{\bf Intelligence on Earth - }
 The diagram gives timing, family relationships and significant innovations
in the development of terrestrial intelligence.  It is likely that very early
evolution occurred in an information carrier other than DNA. With the advent
of learned behavior in mammals and birds, DNA lost a significant part of its
job. More than half of what makes a modern human being is passed culturally.
A self reproducing robot economy could end the DNA era altogether - our
culture will have freed itself of its roots.}
\end{figure}

	Animals bolster the argument on the positive side, except for the
immobile minority like sponges and clams that support it on the negative.

	A billion years ago, before brains or eyes were invented, when the
most complicated animals were something like hydras, double layers of cells
with a primitive nerve net, our progenitors split with the invertebrates.
Now both clans have intelligent members.  Cephalopods are the most
intellectual invertebrates. Most mollusks are sessile shellfish, but octopus
and squid are highly mobile, with big brains and excellent eyes. Evolved
independently of us, they are different. The optic nerve connects to the
back of the retina, so there is no blind spot. The brain is annular, a ring
around the esophagus. The green blood is circulated by a systemic heart
oxygenating the tissues and two gill hearts moving depleted blood.
Hemocyanin, a copper doped protein related to hemoglobin and chlorophyll,
carries the oxygen.

	Octopus and their relatives are swimming light shows, their surfaces
covered by a million individually controlled color changing cells. A
cuttlefish placed on a checkerboard can imitate the pattern, a fleeing
octopus can make deceiving seaweed shapes coruscate backward along its body.
Photophores of deep sea squid, some with irises and lenses, generate bright
multicolored light. Since they also have good vision, there is a potential
for rich communication. Martin Moynihan identifies several dozen distinct
symbolic displays, many correlated with clear emotions in {\bf
Communication and Noncommunication by Cephalopods}.

	Their behavior is mammal like. Octopus are reclusive and shy, squid
are occasionally very aggressive. Small octopus can learn to solve problems
like how to open a container of food. Giant squid, with large nervous
systems, have hardly ever been observed except as corpses. They might be as
clever as whales.

	Birds are vertebrates, related to us through a 300 million year old,
probably not very bright, early reptile. Size-limited by the dynamics of
flying, some are intellectually comparable to the highest mammals.

	The intuitive number sense of crows and ravens extends to seven,
compared to three or four for us. Birds outperform all mammals except higher
primates and the whales in ``learning set'' tasks, where the idea is to
generalize from specific instances. In mammals generalization depends on
cerebral cortex size.  In birds forebrain regions called the Wulst and the
hyperstriatum are critical, while the cortex is small and unimportant.

	Our last common ancestor with the whales was a primitive rat-like
mammal alive 100 million years ago. Some dolphin species have body and brain
masses identical to ours, and have had them for more generations. They are
as good as us at many kinds of problem solving, and can grasp and
communicate complex ideas. Killer whales have brains five times human size,
and their ability to formulate plans is better than the dolphins', who they
occasionally eat. Sperm whales, though not the largest animals, have the
world's largest brains.  Intelligence may be an important part of their
struggle with large squid, their main food. Elephant brains are three times
human size. Elephants form matriarchal tribal societies and exhibit complex
behavior. Indian domestic elephants learn over 500 commands, and form
voluntary mutual benefit relationships with their trainers, exchanging labor
for baths. They can solve problems such as how to sneak into a plantation at
night to steal bananas, after having been belled (answer: stuff mud into the
bells). And they do have long memories.

	Apes are our 10 million year cousins. Chimps and gorillas can learn
to use tools and to communicate in human sign languages at a retarded level.
Chimps have one third, and gorillas one half, human brainsize.

	Animals exhibiting near-human behavior have hundred billion neuron
nervous systems. Imaging vision alone requires a billion. The smartest
insects have a million brain cells, while slugs and worms make do with a
thousand, and sessile animals with a hundred. The portions of nervous
systems for which tentative wiring diagrams have been obtained, including
nearly all of the large neuroned sea slug, Aplysia, the flight controller of
the locust and the early stages of vertebrate vision, reveal neurons
configured into efficient, clever, assemblies.  

\section{Mobility and Intelligence around the Lab}
	The twenty year old modern robotics effort can hardly hope to rival
the billion year history of large life on earth in richness of example or
profundity of result. Nevertheless the evolutionary pressures that shaped
life are already palpable in the robotics labs. The following is a thought
experiment that we hope soon to make into a physical one.

	We desire robots able to execute general tasks such as ``go down the
hall to the third door, go in, look for a cup and bring it back''.  This
desire has created a pressing need - a computer language in which to
concisely specify complex tasks for a rover, and a hardware and software
system to embody it. Sequential control languages successfully used with
industrial manipulators might seem a good starting point.

	Paper attempts at defining the structures and primitives required
for the mobile application revealed that the linear control structure of
these state-of-the-art arm languages was inadequate for a rover. The
essential difference is that a rover, in its wanderings, is regularly
``surprised'' by events it cannot anticipate, but with which it must deal.
This requires that contingency routines be activated in arbitrary order, and
run concurrently. One answer is a structure where a number of specialist
programs communicating via a common data structure called a blackboard are
active at the same time, some operating sensors, some controlling effectors,
some integrating the results of other modules, and some providing overall
direction.  As conditions change the priority of the various modules
changes, and control may be passed from one to another.

\section{Character from Motion}
	Suppose we ask our future robot, equipped with a controller based on
the blackboard system mentioned in the last section, to, in fact, go down the
hall to the third door, go in, look for a cup and bring it back. This will
be implemented as a process that looks very much like a program written for
the arm control languages (that in turn look very much like Algol, or
Basic), except that the door recognizer routine would probably be activated
separately. Consider the following caricature of such a program.

\begin{tabbing}
\vspace{.25in}\\
begintab \= begintab \= begintab \= begintab \= begintab \= \+\kill
 {\bf module} {\sf GO-FETCH-CUP} \+ \\
 {\bf wake up}  {\sf DOOR-RECOGNIZER} {\bf with instructions}\+ \\
         ( {\bf on} {\sf FINDING-DOOR} {\bf add} 1 {\bf to} {\sf DOOR-NUMBER}\\
         \ \ \ {\bf record} {\sf DOOR-LOCATION} )\-\-\\
\vspace{.25in}\\
 {\bf record} {\sf START-LOCATION}\\
 {\bf set} {\sf DOOR-NUMBER} {\bf to} 0\\
 {\bf while} {\sf DOOR-NUMBER} $<$ 3 {\sf WALL-FOLLOW}\\
 {\sf FACE-DOOR}\\
 {\bf if} {\sf DOOR-OPEN} {\bf then} {\sf GO-THROUGH-OPENING}\+\\
            {\bf else} {\sf OPEN-DOOR-AND-GO-THROUGH}\-\\
 {\bf set} {\sf CUP-LOCATION} {\bf to result of} {\sf LOOK-FOR-CUP}\\
 {\sf TRAVEL} {\bf to} {\sf CUP-LOCATION}\\
 {\sf PICKUP-CUP} {\bf at} {\sf CUP-LOCATION}\\
 {\sf TRAVEL} {\bf to} {\sf DOOR-LOCATION}\\
 {\sf FACE-DOOR}\\
 {\bf if} {\sf DOOR-OPEN} {\bf then} {\sf GO-THROUGH-OPENING}\+\\
            {\bf else} {\sf OPEN-DOOR-AND-GO-THROUGH}\-\\
 {\sf TRAVEL} {\bf to} {\sf START-LOCATION}\\
 {\bf end}\\
\end{tabbing}

	So far so good. We activate our program and the robot obediently
begins to trundle down the hall counting doors. It correctly recognizes the
first one.  The second door, unfortunately, is decorated with some garish
posters, and the lighting in that part of the corridor is poor, and our
experimental door recognizer fails to detect it. The wall follower, however,
continues to operate properly and the robot continues on down the hall, its
door count short by one. It recognizes door 3, the one we had asked it to go
through, but thinks it is only the second, so continues. The next door is
recognized correctly, and is open.  The program, thinking it is the third
one, faces it and proceeds to go through.  This fourth door, sadly, leads to
the stairwell, and the poor robot, unequipped to travel on stairs, is in
mortal danger.

	Fortunately there is a process in our concurrent programming system
called {\sf DETECT-CLIFF} that is always running and that checks ground position
data posted on the blackboard by the vision processes and also requests
sonar and infrared proximity checks on the ground. It combines these,
perhaps with an a-priori expectation of finding a cliff set high when
operating in dangerous areas, to produce a number that indicates the
likelihood there is a drop-off in the neighborhood.

	A companion process {\sf DEAL-WITH-CLIFF} also running continuously,
but with low priority, regularly checks this number, and adjusts its own
priority on the basis of it. When the cliff probability variable becomes
high enough the priority of {\sf DEAL-WITH-CLIFF} will exceed the priority of
the current process in control, {\sf GO-FETCH-CUP} in our example, and
{\sf DEAL-WITH-CLIFF} takes over control of the robot. A properly written
{\sf DEAL-WITH-CLIFF} will then proceed to stop or greatly slow down the
movement of the robot, to increase the frequency of sensor measurements of
the cliff, and to slowly back away from it when it has been reliably
identified and located.

	Now there's a curious thing about this sequence of actions. A person
seeing them, not knowing about the internal mechanisms of the robot might
offer the interpretation ``First the robot was determined to go through the
door, but then it noticed the stairs and became so frightened and
preoccupied it forgot all about what it had been doing''. Knowing what we do
about what really happened in the robot we might be tempted to berate this
poor person for using such sloppy anthropomorphic concepts as
determinination, fear, preoccupation and forgetfulness in describing the
actions of a machine. We could berate the person, but it would be wrong.

	The robot came by the emotions and foibles indicated as honestly as
any living animal - the observed behavior is the correct course of action
for a being operating with uncertain data in a dangerous and uncertain
world. An octopus in pursuit of a meal can be diverted by hints of danger in
just the way the robot was. An octopus also happens to have a nervous system
that evolved entirely independently of our own vertebrate version. Yet most
of us feel no qualms about ascribing concepts like passion, pleasure, fear
and pain to the actions of the animal.

	We have in the behavior of the vertebrate, the mollusc and the robot
a case of convergent evolution. The needs of the mobile way of life have
conspired in all three instances to create an entity that has modes of
operation for different circumstances, and that changes quickly from mode to
mode on the basis of uncertain and noisy data prone to misinterpretation. As
the complexity of the mobile robots increases their similarity to animals
and humans will become even greater.

\subsection{Deeper}
	Hold on a minute, you say. There may be some resemblance between the
robot's reaction to a dangerous situation and an animal's, but surely there
are differences. Isn't the robot more like a startled spider, or even a
bacterium, than like a frightened human being?  Wouldn't it react over and
over again in exactly the same way, even if the situation turned out not to be
dangerous?  You've caught me.

	I think the spider's nervous system is an excellent match for robot
programs possible today. We passed the bacterial stage in the 1950s with
light seeking electronic turtles. This does not mean that concepts like
thinking and consciousness are ruled out. In the book {\bf Animal Thinking},
animal ethologist D. G. Griffiths reviews evidence that much animal behavior,
including that of insects, can be explained economically in terms of
consciousness: an internal model of the self and surroundings, that, however
crudely, allows consideration of alternative actions.  But there are
differences of degree.

\subsection{Pleasure and Pain}
	Even single sensory neurons have been shown to habituate to over or
under stimulation. Small networks of neurons can adapt in more elaborate
ways, for instance by learning to associate one stimulus with another. Such
mechanisms tune a nervous system to the body it inhabits, and to its
environment.  Vertebrates owe much of their potential to an elaboration of
this arrangement.

	The vertebrate brain has centralized loci for pleasure and pain.
Stimulation of a pleasure center acts to encourage future expressions of the
preceding behavior, while a pain stimulus discourages it.  The archetypical
demonstration involves an electrode implanted in the pleasure center of a
rat.  Allowed to operate a lever that energizes the electrode, the rat
ignores food and water to rapidly and repeatedly press the lever,
interrupted only by total exhaustion.

	Centralization of conditioning sites probably increases long term
adaptability. A new need or danger can be accomodated through a small change
in the neural wiring, by connecting a detector for the condition to a
pleasure or pain site. The standard learning mechanism will then insure that
the animal begins to seek the conditions that meet the need, or to avoid the
danger, even if the required behavior is complex.

\subsection{Love and Hate}
	We are deep in the realm of speculation now, but the general
pleasure/pain learning mechanism may provide an explanation for
abstract emotions.

	Let's suppose that altruism, for instance of a mother towards her
offspring, can enhance the long term survival of the altruist's genes even
though it has a negative effect on the individual altruist.  Feeding the
young may leave the mother exhausted and hungry, and defending them may
involve her in risk or injury.  Shouldn't the conditioning mechanisms we've
just described gradually suppress this kind of behavior?

	As with more immediate concerns, activities that are
multi-generationally beneficial can be encouraged, and ultimately harmful
behaviors suppressed, if detectors for them are wired strongly to pleasure
and pain centers.  For instance, mother love is encouraged if the sight,
feel, sound or smell of the offspring triggers pleasure, and if absence of
the young is painful.

	To the extent that pain or pleasure has a subjective manifestation
such non-immediate causes are likely to {\it feel} different from more obvious
ones like skin pain or hunger.  Most of the immediate causes are associated
with some part of the body, and can be usefully subjectively mapped there.
Multi-generational imperatives, on the other hand, cannot be so simply
related to the physically apparent world.  This may help explain the
etherial or spiritual quality that is often associated with such
transcendent motivations.  Certainly they deserve respect, being the
distillation of perhaps tens of millions of years of life or death trials.

\subsection{What If?}
	Elaboration of the internal world model in higher animals made
possible another twist.  A rich world model allows its possessor to examine
in detail alternative situations, past, future or merely hypothetical.
Dangers avoided can yet be brooded over, and what {\it might} have happened
can be imagined.  If the mental simulation is accurate enough, such brooding
can produce useful warnings, or point out missed opportunities. These
lessons of the imagination are most effective if their consequences are tied
to the conditioning mechanism, just as with real events. Such a connection
is particularly easy to explain if, as we elaborate below, the most powerful
aspects of reason are due to world knowledge powerfully encoded in the
sensory and motor systems. The same wiring that conditions on real situations
would be activated for imaginary ones.

	The ability to imagine must be a key component in communication
among higher animals (not to mention between you and me).  Messages trigger
mental scenarios that then provide conditioning (i.e. learning).

	Communication that fails to engage the emotions is not very
educational in this sense (yes, yes, we're talking about people now), and a
waste of time.  Imagine detectors for time well spent and time wasted,
themselves wired to the conditioning centers. It's not too far fetched to
think that these correspond to the subjective emotions of ``interesting''
and ``boring''.  Humans seem to have cross wiring that allows elaborate
imagining, for instance about future rewards, to deem interesting activities
that might normally be boring.  How else could there be intellectuals!
Indeed, the conventional view of intelligence, and the bulk of work in
artificial intelligence, centers on this final twist.  While I believe
that it is important, it is only a tiny part of the whole story, and often
overrated.

\section{Coming Soon}

	Some facets of the above have been explored, somewhat haphazardly,
in machines.  British psychologist W. Gray Walter built electronic turtles
that demonstrated learning by association, represented as charges on a
matrix of capacitors.  Arthur Samuel (then of IBM, now Stanford) wrote a
checker playing program that adjusted evaluation parameters to improve its
play, and was able to learn simply by playing game after game against itself
overnight.  Frank Rosenblatt of Cornell invented networks of elements that
resembled neurons that could be trained to do simple tasks by properly timed
punish and reward signals that adjusted the thresholds of synapses between
``neurons'' that had recently fired. These approaches of the 1950s and 60s
fell out of fashion in the last decade, but modern variations are again
in vogue.

	Among the natural traits in the immediate roving robot horizon is
parameter adjustment learning. A precision mechanical arm in a rigid
environment can usually have its kinematic self-model and its dynamic
control parameters adjusted once, permanently. A mobile robot bouncing
around in the muddy world is likely to continuously suffer insults like dirt
buildup, tire wear, frame bends and small mounting bracket slips that mess
up accurate a-priori models. Our obstacle course software, for instance, has
a camera calibration phase. The robot is parked precisely in front of an
painted grid of spots. A program notes how the camera images the spots and
figures a correction for camera distortions, so that later programs can make
precise visual angle measurements.  The present code is very sensitive to
mis-calibrations, and we are working on a method that will continuously
calibrate the cameras just from the images perceived on normal trips through
clutter. With such a procedure in place, a bump that slightly shifts one of
the robot's cameras will no longer cause systematic errors in its
navigation. Animals seem to tune most of their nervous systems with
processes of this kind, and such accomodation may be a precursor to more
general kinds of learning.

	Perhaps more controversially, the begininnings of self awareness can
be seen in the robots. All of the control programs of the more advanced
mobile robots have internal representations, at varying levels of
abstraction and precision, of the world around the robot, and of the robot's
position within that world. The motion planners work with these world models
in considering alternative future actions for the robot.  If the programs
had verbal interfaces one could ask questions that receive answers such as
``I turned right because I didn't think I could fit through the opening on
the left ''. As it is the same information is often presented in the form of
pictures drawn by the programs.

\section{When?}

	How does computer speed compare with human thought?  The answer has 
been changing.

	The first electronic computers were constructed in the mid 1940s to
solve problems too large for unaided humans.  {\it Colossus}, one of a
series of ultrasecret British machines, broke the German {\it Enigma} code,
greatly influencing the course of the European war, by scanning through code
keys tens of thousands of times faster than humanly possible.  In the US
{\it Eniac} computed antiaircraft artillery tables for the Army, and later
did calculations for the atomic bomb, at similar speeds.  Such feats earned
the early machines the popular appellation {\it Giant Brains}.

	In the mid 1950s computers more than ten times faster than Eniac
appeared in many larger Universities. They did numerical scientific
calculations nearly a million times faster than humans. A few
visionaries took the Giant Brains metaphor seriously and began to write
programs for them to solve intellectual problems going beyond mere
calculation.  The first such programs were encouragingly successful.
Computers were soon solving logic problems, proving theorems in Euclidean
geometry, playing checkers, even doing well in IQ test analogy problems. The
performance level and the speed in each of these narrow areas was roughly
equivalent to that of a college freshman who had recently learned the
subject.  The automation of thought had made a great leap, but paradoxically
the term ``Giant Brains'' seemed less appropriate.

	In the mid 1960s a few centers working in this new area of {\it
Artificial Intelligence} added another twist: mechanical eyes, hands and
ears to provide real world interaction for the thinking programs.  By then
computers were a thousand times faster than Eniac, but programs to do even
simple things like clearing white blocks from a black tabletop turned out to
be very difficult to write, and performed hundreds of times more slowly, and 
much less reliably, than a human.  Slightly more complicated tasks took much 
longer, and many seemingly trivial things, like identifying a few simple
objects in a jumble, still cannot be done acceptably at all twenty years
later, even given hours of computer time.  Forty years of research and a
millionfold increase in computer power has reduced the image of computers
from Giant Brains to mental midgets.  Is this silly, or what?

\section{Easy and Hard}
	The human evolutionary record provides a clue to the paradox. While
our sensory and muscle control systems have been in development for almost a
billion years, and common sense reasoning has been honed for perhaps a 
million, really high level, deep, thinking is little more than a parlor
trick, culturally developed over a few thousand years, which a few humans,
operating largely against their natures, can learn. As with Samuel Johnson's
dancing dog, what is amazing is not how well it is done, but that it is done
at all.

	Computers can challenge humans in intellectual areas, where humans
are evolutionary novices, because they can be programmed to carry on much less
wastefully.  Arithmetic is an extreme example, a function learned by humans
with great difficulty, but instinctive to computers. A 1987 home computer
can add a million large numbers in a second, astronomically faster
than a person, and with no errors. Yet the 100 billion neurons in a human
brain, if reorganized by a mad neurosurgeon into adders using switching logic 
design principles, could sum one hundred thousand times faster than
the computer.

	Computers do not challenge humans in perceptual and control areas
because these ancient functions are carried out by large fractions of the
nervous system wired for those jobs as cleanly as the hypothetical neuron
adder above.  Present day computers, however efficiently programmed, are
simply too puny to keep up. Evidence comes from the classic program of
reverse engineering of some of the visual system of vertebrates initiated by
David Hubel and Thorsten Weisel in the 1960s.  They elucidated some of its
operation by microscopically examining the structure of the retina and the
vision-related parts of the brains of cats and monkeys and using electrodes
to monitor signals there as test patterns were presented to the eyes. 

\begin{figure}
\vspace{3.25in}
\caption[Retina]{\label{Retina}
{\bf Human Retina - } This cross section shows a tiny portion
of each of the ten layers of neurons that form the retina. There are about
400 million efficiently organized neurons in all. (From page 185 of
Photoprocesses, Photoreceptors and Evolution by Jerome J. Wolken, Academic
Press, 1975.)}
\end{figure}

	The vertebrate retina consists of a highly organized, ten layered,
structure of densely packed neurons fed by about one hundred million light
sensors (figure \ref{Retina}).  The sensors are merged into clusters giving an
effective resolution of one million picture elements, or {\it pixels}, to
use the computer jargon.  The other neurons combine their outputs in various
ways to detect such things as edges, corners, curvature and motion.  Each of
these simple {\it operators} employs about 10 to 100 neurons per pixel.
Thus processed, the image goes via the optic nerve to the much bigger visual
cortex in the brain.

	Assuming the visual cortex does as much computing for its size as
the retina (perhaps an overestimate - the retina is a small, old and highly
optimized structure; larger and more recent regions may use neurons less
efficiently), we can estimate the total capability of the system.  The
visual cortex has about $10$ billion neurons, a thousand times the number in
a modest retinal operation.  The eye can process ten images a second, so the
cortex may do the computational equivalent of $10,000$ small retinal
operations a second.

	Operations similar to the retinal ones have been found very useful for
robot vision. An efficient program running on a typical (1 million
instruction per second) computer can do the equivalent work of one small
retinal operation in about $10$ seconds. Thus, seeing programs on present day
computers seem to be $100,000$ times slower than human vision.

	The whole brain is about ten times larger than the visual system, so
it should be possible to write real-time human equivalent programs for a
machine one million times more powerful than today's medium sized computer.
In 1987 the largest supercomputers are about $1000$ times slower than this 
desiratum.

\begin{figure}
\vspace{6.5in}
\caption[Think]{\label{Think}
{\bf Think Power - } Computing speed and memory of some animals
and machines. The animal figures are for the nervous system only, calculated
at 100 bits per second and 100 bits of storage per neuron.  These are
speculative estimates, but note that a factor of 100 one way or the other
would change the appearance of the graph only slightly.}
\end{figure}

\section{Intellectual Voyages}

	Interesting computation and thought requires a processing engine of
sufficient computational {\it power} and {\it capacity}. Roughly, power is
the speed of the machine, and capacity is its memory size.

	Here's a helpful metaphor. Computing is like a sea voyage in a 
motorboat. How fast a given journey can be completed depends on the power of 
the boat's engine.  The maximum length of any journey is limited by the 
capacity of its fuel tank.  The effective speed is decreased, in general, if 
the course of the boat is constrained, for instance to major compass 
directions.

	Some computations are like a trip to a known location on a distant
shore, others resemble a mapless search for a lost island.  Parallel
computing is like having a fleet of small boats - it helps in searches, and
in reaching multiple goals, but not very much in problems that require a
distant sprint. Special purpose machines trade a larger engine for less
rudder control.

	Attaching disks and tapes to a computer is like adding secondary
fuel tanks to the boat. The capacity, and thus the range, is increased, but
if the connecting plumbing is too thin, it will limit the fuel flow rate and
thus the effective power of the engine.

	Extending the metaphor, input/output devices are like boat sails.
They capture power and capacity in the environment.  Outside information is
a source of variability, and thus power, by our definition. More concretely,
it may contain answers that would otherwise have to be computed. The external 
medium can also function as extra memory, increasing capacity.

	Figure \ref{Think} shows the power and capacity of some interesting
natural and artificial thinking engines.  At its best, a computer
instruction has a few tens of bits of information, and a million instruction
per second computer represents a few tens of millions of bits/second of
power.  The power ratio between nervous systems and computers is as
calculated in the last section:  a million instructions per second is worth
about a hundred thousand neurons.  I also assume that a neuron represents
about 100 bits of storage, suggested by recent evidence of synaptic learning
in simple neurvous systems by Eric Kandel and others. Note that change of a
factor of ten or even one hundred in these ratios would hardly change the
graph qualitatively.  (My forthcoming book {\bf Mind Children}, from which
this paper is drawn, offers more detailed technical justifications for these
numbers).

	The figure shows that current laboratory computers are equal
in power approximately to the nervous systems of insects.  It is these
machines that support essentially all the research in artificial
intelligence. No wonder the results to date are so sparse! The largest
supercomputers of the mid 1980s are a match for the 1 gram brain of a mouse,
but at ten million dollars or more apiece they are reserved for serious work.

\begin{figure}
\vspace{5.75in}
\caption[Compute]{\label{Compute}
{\bf A Century of Computing - } The cost of calculation has
dropped a thousandfold every twenty years (or halved every two years) since
the late nineteenth century. Before then mechanical calculation was an
unreliable and expensive novelty with no particular edge over hand
calculation. The graph shows a mind boggling {\it trillionfold} decrease in
the cost since then. The pace has actually picked up a little since the
beginning of the century. It once took 30 years to accumulate a thousandfold
improvement; in recent decades it takes only 19.  Human equivalence should
be affordable very early in the 21st century.}
\end{figure}

\section{The Growth of Processing Power}
	How long before the research medium is rich enough for full 
intelligence?

	Although a number of mechanical digital calculators were devised and 
built during the seventeenth and eighteenth centuries, only with the 
mechanical advances of the industrial revolution did they become reliable and 
inexpensive enough to routinely rival manual calculation. By the late 
nineteenth century their edge was clear, and the continuing progress dramatic.
Since then the cost of computing has dropped a thousandfold every twenty years
(figure \ref{Compute}).

	The early improvements in speed and reliability came with advances in 
mechanics - precision mass produced gears and cams, for instance, improved 
springs and lubricants, as well as increasing design experience and 
competition among the calculator manufacturers.  Powering calculators by 
electric motors provided a boost in both speed and automation in the 1920s, as
did incorporating electromagnets and special switches in the innards in the
1930s.  Telephone relay methods were used to make fully automatic computers 
during World War II, but these were quickly eclipsed by electronic tube 
computers using radio, and ultrafast radar, techniques.  By the 1950s 
computers were an industry that itself spurred further major component 
improvements.

	The curve in figure \ref{Compute} is not leveling off, and the
technological pipeline is full of developments that can sustain the pace for
the foreseeable future. Success in this enterprise, as in others, breeds
success. Not only is an increasing fraction of the best human talent engaged
in the research, but the ever more powerful computers themselves feed the
process. Electronics is riding this curve so quickly that it is likely to be
the main occupation of the human race by the end of the century.

	The price decline is fueled by miniaturization, which supplies a
double whammy. Small components both cost less and operate more quickly.
Charles Babbage, who in 1834 was the first person to conceive the idea of an
automatic computer, realized this. He wrote that the speed of his design,
which called for hundreds of thousands of mechanical components, could be
increased in proportion if ``as the mechanical art achieved higher states of
perfection'' his palm sized gears could be reduced to the scale of
clockwork, or further to watchwork. (I fantasize an electricityless world
where the best minds continued on Babbage's course.  By now there would be
desk and pocket sized mechanical computers containing millions of
microscopic gears, computing at thousands of revolutions per second.)

	To a remarkable extent the cost per pound of machinery has remained 
constant as its intricacy increased.  This is as true of consumer electronics 
as of computers (merging categories in the 1980s). The radios of the 1930s 
were as large and as expensive as the televisions of the 1950s, the color 
televisions of the 1970s, and the home computers of the 1980s. The volume 
required to amplify or switch a single signal dropped from the size of a fist 
in 1940, to that of a thumb in 1950, to a pencil eraser in 1960, to a salt 
grain in 1970, to a small bacterium in 1980. In the same period the basic 
switching speed rose a millionfold, and the cost declined by the same huge 
amount.

	Predicting the detailed future course is impossible for many reasons. 
Entirely new and unexpected possibilities are encountered in the course of 
basic research. Even among the known, many techniques are in competition, and 
a promising line of development may be abandoned simply because some other 
approach has a slight edge. I'll content myself with a short list of some of 
what looks promising today.

	In recent years the widths of the connections within integrated
circuits have shrunk to less than one micron, perilously close to the 
wavelength of the light used to ``print'' the circuitry.  The manufacturers 
have switched from visible light to shorter wavelength ultraviolet, but this 
gives them only a short respite.  X-rays, with much shorter wavelengths, would
serve longer, but conventional X-ray sources are so weak and diffuse that they
need uneconomically long exposure times. High energy particle physicists have 
an answer.  Speeding electrons curve in magnetic fields, and spray photons 
like mud from a spinning wheel.  Called synchotron radiation for the class of 
particle accelerator where it became a nuisance, the effect can be harnessed 
to produce powerful beamed X-rays.  The stronger the magnets, the smaller can 
be the synchotron. With liquid helium cooled superconductiong magnets an 
adequate machine can fit into a truck, otherwise it is the size of a small 
building. Either way, synchotrons are now an area of hot interest, and promise
to shrink mass-produced circuitry into the sub-micron region. Electron and ion
beams are also being used to write submicron circuits, but present systems 
affect only small regions at a time, and must be scanned slowly across a chip.
The scanned nature makes computer controlled electron beams ideal, however, 
for manufacturing the ``masks'' that act like photographic negatives in 
circuit printing.

	Smaller circuits have less electronic ``inertia'' and switch both
faster and with less power. On the negative side, as the number of electrons 
in a signal drops it becomes more prone to thermal jostling. This effect can 
be countered by cooling, and indeed very fast experimental circuits can now be
found in many labs running in supercold liquid nitrogen, and one supercomputer
is being designed this way. Liquid nitrogen is produced in huge amounts in the
manufacture of liquid oxygen from air, and it is very cheap (unlike the much 
colder liquid helium).  

	The smaller the circuit, the smaller the regions across which voltages
appear, calling for lower voltages.  Clumping of the substances in the crystal
that make the circuit becomes more of a problem as they get smaller, so more 
uniform ``doping'' methods are being developed. As the circuits become smaller
quantum effects become more pronounced, creating new problems and new 
opportunities. Superlattices, mutiple layers of atoms-thick regions of 
differently doped silicon made with molecular beams, are such an opportunity. 
They allow the electronic characteristics of the material to be tuned, and 
permit entirely new switching methods, often giving tenfold improvements.

	The first transistors were made of germanium; they could not stand
high temperatures and tended to be unreliable. Improved understanding of
semiconductor physics and ways of growing silicon crystals made possible
faster and more reliable silicon transistors and integrated circuits.  New
materials are now coming into their own. The most immediate is gallium
arsenide. Its lattice impedes electrons less than silicon, and makes
circuits up to ten times faster. The Cray 3 supercomputer due in 1989 will
use gallium arsenide integrated circuits, packed into a one cubic foot
volume, to top the Cray 2's speed tenfold. Other compounds like indium
phosphide and silicon carbide wait in the wings. Pure carbon in diamond form
is an obvious possibility - it should be as much an improvement over Gallium
Arsenide as that crystal is over Silicon. Among its many superlatives,
perfect diamond is the best solid conductor of heat, an important property
in densely packed circuitry. The vision of an utradense three dimensional
circuit in a gem quality diamond is compelling. As yet no working circuits
of diamond have been reported, but excitment is mounting as reports of
diamond layers up to a millimeter thick grown from hot methane come from the
Soviet Union, Japan and, belatedly, the United States.

	Farther off the beaten track are optical circuits that use lasers and 
non-linear optical effects to switch light instead of electricity. Switching 
times of a few picoseconds, a hundred times faster than conventional circuits,
have been demonstrated, but many practical problems remain. Finely tuned laser
has also been used with light sensitive crystals and organic molecules in 
demonstration memories that store up to a trillion bits per square centimeter.

	The ultimate circuits may be superconducting quantum devices, which 
are not only extremely fast, but extremely efficient. Various superconducting 
devices have been in and out of fashion several times over the past twenty 
years. They've had a tough time because the liquid helium environment they 
require is expensive, the heating/cooling cycles are stressful, and especially
because rapidly improving semiconductors have offered such tough competition. 

	Underlying these technical advances, and preceding them, are equally
amazing advances in the methods of basic physics. One recent, unexpected and
somewhat unlikely, device is the inexpensive tunnelling microscope that can
reliably see, identify and soon manipulate single atoms on surfaces by
scanning them with a very sharp needle. The tip is positioned by three
piezoelectric crystals microscopically moved by small voltages. It maintains
a gap a few atoms in size by monitoring a current that jumps across it. The
trickiest part is isolating the system from vibrations. It provides our
first solid toehold on the atomic scale.

	A new approach to miniaturization is being pursued by enthusiasts in
the laboratories of both semiconductor and biotechnology companies, and
elswhere. Living organisms are clearly machines when viewed at the molecular
scale.  Information encoded in RNA ``tapes'' directs protein assembly
devices called ribosomes to pluck particular sequences of amino acids from
their environment and attach them to the ends of growing chains. Proteins,
in turn, fold up in certain ways, depending on their sequence, to do their
jobs. Some have moving parts acting like hinges, springs, latches triggered
by templates.  Others are primarily structural, like bricks or ropes or
wires.  The proteins of muscle tissue work like ratcheting pistons. Minor
modifications of this existing machinery are the core of today's
biotechnology industry. The visionaries see much greater possibilities.

	Proteins to do specific jobs can be engineered even without a
perfect model of their physics.  Design guidelines, with safety margins to
cover the uncertainties, can substitute.  The first generation of artificial
molecular machinery would be made of protein by mechanisms recruited from
living cells. Early products would be simple, like tailored medicines, and
experimental, like little computer circuits.  Gradually a bag of tricks, and
computer design aids, would accumulate to build more complicated machines.
Eventually it may be possible to build tiny robot arms, and equally tiny
computers to control them, able to grab molecules and hold them, thermally
wriggling, in place.  The protein apparatus could then be used as machine
tools to build a second generation of molecular devices by assembling atoms
and molecules of all kinds. For instance, carbon atoms might be laid,
bricklike, into ultra strong fibers of perfect diamond. The smaller, harder,
tougher machines so produced would be the second generation molecular
machinery.

	The book {\bf Engines of Creation} by Eric Drexler, and a forthcoming
book by Conrad Schneiker, call the entire scheme
{\it nanotechnology}, for the nanometer scale of its parts. By contrast
today's integrated circuit microtechnology has micrometer features, a
thousand times bigger.  Some things are easier at the nanometer scale.
Atoms are perfectly uniform in size and shape, if somewhat fuzzy, and behave
predictably, unlike the nicked, warped and cracked parts in larger
machinery.  

\section{A Stumble}
	It seemed to me throughout the 1970s (I was serving an extended 
sentence as a graduate student at the time) that the processing power 
available to AI programs was not increasing very rapidly. In 1970 most of my 
work was done on a Digital Equipment Corp. PDP-10 serving a community of 
perhaps thirty people. In 1980 my computer was a DEC KL-10, five times as fast
and with five times the memory of the old machine, but with twice as many 
users. Worse, the little remaining speedup seemed to have been absorbed in 
computationally expensive convenience features: fancier time sharing and high 
level languages, graphics, screen editors, mail systems, computer networking 
and other luxuries that soon became necessities.

	Several effects together produced this state of affairs. Support for
university science in general had wound down in the aftermath of the Apollo
moon landings and politics of the Vietnam war, leaving the universities to
limp along with aging equipment. The same conditions caused a recession in
the technical industries - unemployed engineers opened fast food restaurants
instead of designing computers (the rate of change in figure \ref{Compute}
does slacken slightly in the mid 1970s). The initially successful ``problem
solving'' thrust in AI had not yet run its course, and it still seemed to
many that existing machines were powerful enough - if only the right
programs could be found. While spectacular progress in the research became
increasingly difficult, a pleasant synergism among the growing number of
information utilities on the computers created an attractive diversion for
the best programmers - creating more utilities.

	If the 1970s were the doldrums, the 1980s more than compensated. 
Several salvations had been brewing. The Japanese industrial successes focused
attention worldwide on the importance of technology, particularly computers 
and automation, in modern economies - American industries and government 
responded with research dollars. The Japanese stoked the fires, under the 
influence of a small group of senior researchers, by boldly announcing a major
initiative towards future computers, the so called ``Fifth Generation'' 
project, pushing the most promising American and European research directions.
The Americans responded with more money. Besides this, integrated circuitry 
had evolved far enough that an entire computer could fit on a chip. Suddenly 
computers were affordable by individuals, and a new generation of computer 
customers and manufacturers came into being. The market was lucrative, the 
competition fierce, and the evolution swift, and by the mid 1980s the momentum
lost in the previous decade had been regained, with interest. Artificial 
intelligence research is awash in a cornucopia of powerful new ``personal'' 
workstation computers, and there is talk of applying supercomputers to the 
work.

	Even without supercomputers, human equivalence in a research setting
should be possible by around 2010, as suggested by figure \ref{Compute}.
Now, the smallest vertebrates, shrews and hummingbirds, get interesting
behavior from nervous systems one ten thousandth the size of a human's, so I
expect fair motor and perceptual competence, in about a decade.

\section{Faster Yet?}
	Very specialized machines can provide up to one thousand times the
effective performance for a given price in well defined tasks. Some vision
and control problems may be candidates for this approach. Special purpose
machines are not a good solution in the groping research stage, but may
dramatically lower the costs of intelligent machines when the problems and
solutions are well understood. Some principals in the Japanese Fifth
Generation Computer Project have been quoted as planning ``man capable''
systems in ten years. I believe this more optimistic projection is unlikely,
but not impossible.

	As the computers become more powerful and as research in this area
becomes more widespread the rate of visible progress should accelerate. I 
think artificial intelligence via the ``bottom up'' approach of technological 
recapitulation of the evolution of mobile animals is the surest bet because 
the existence of independently evolved intelligent nervous systems indicates 
that there is an incremental route to intelligence. It is also possible, of 
course, that the more traditional ``top down'' approach will achieve its 
goals, growing from the narrow problem solvers of today into the much harder 
areas of learning, common-sense reasoning and perceptual acquisition of 
knowledge as computers become large and powerful enough, and the techniques 
are mastered. Most likely both approaches will make enough progress that they 
can effectively meet somewhere in the middle, for a grand synthesis into a 
true artificial sentience.

	This artificial person will have some interesting properties. Its high
level reasoning abilities should be astonishingly better than a human's - even
today's puny systems are much better in some areas - but its low level 
perceptual and motor abilities will be comparable to ours. Most interestingly 
it will be highly changeable, both on an individual basis and from one of its 
generations to the next. And it will quickly become cheap.

\end{document}