COMMENT ⊗   VALID 00002 PAGES
C REC  PAGE   DESCRIPTION
C00001 00001
C00002 00002	.require "memo.pub[let,jmc]" source
C00018 ENDMK
C⊗;
.require "memo.pub[let,jmc]" source
.cb STATEMENT ON THE STATE OF THE ART IN ROBOTICS

.cb by John McCarthy


	This is a statement on the state of the art of robotics in so
far as it is relevant to the problem of designing and manufacturing
a "domestic android" of the sort Quasar Industries has been promising
to sell for α$4000 in a few years.

	By way of self-introduction, I am Professor of Computer Science
at Stanford University and Director of the Stanford Artificial Intelligence
Laboratory.  I have been involved with robotics since about 1964 and
with artificial intelligence since 1952.
I received the A.M. Turing award of the Association for Computing
Machinery for 1971 for my work in artificial intelligence and other
areas of computer science.  Besides technical publications, I have
written the article on Robots for the Britannica Yearbook of Science
and Technology and the article on artificial intelligence for Collier's
Encyclopedia.  (To be included in the next revision).

	Robotics is the art of making machines that can carry out
motor and sensory activities like those of humans.  While robots
are old hat in science fiction, the main serious work dates from about 1960.
A fully developed robot will require a body with powered arms and
legs, vision, hearing and touch sensors, and a controlling computer
program capable of visual and auditory pattern recognition and
common sense intelligence.  No-one is even close to meeting all
these goals, but some of them have been partially achieved in
the laboratory and also in industry.

There are several lines of development.
.item←0

	#. Industrial robots.  These machines are usually single
arms mounted on a pedestal.  They perform repetitive industrial
tasks like putting a part in a furnace or press or welding parts
of automobile bodies.  They are mainly used for jobs that are
dangerous or unpleasant for humans, because they have been too
expensive and inflexible for tasks humans do well and comfortably.
Nevertheless, buying and programming an industrial robot is often
cheaper than building a special purpose machine for the job.
The simplest industrial robots carry out fixed sequences of actions
stored on a paper tape, i.e. the action doesn't depend on events
that occur during the action and is always the same once the robot
has been "programmed".  More recent robots have somewhat flexible
programs that can take into account information from sensors giving
the positions of objects.  For example, the robots that weld car
bodies need to sense the position of the car as it moves down
the assembly line, because its position is not sufficiently predictable
in advance.
Some very recent industrial robots, e.g. at General Motors and at
Hitachi in Japan, use television cameras to recognize the positions
of objects to be manipulated.  This requires that the robot be controlled
by a computer that can find the objects in a television image.

	#. Academic projects.  The Stanford Artificial Intelligence
Laboratory has been involved in robotics since 1965.  For several
years, the National Science Foundation has supported a project
aimed at using robots to assemble objects from parts.  An arm
designed in our laboratory is now the basis of a joint project
by Unimation Inc. (the biggest manufacturer of industrial robots)
and General Motors.  This project is designing an arm that can
be used for general purpose light assembly work.

	Perhaps our largest activity has been computer vision
aimed at finding objects in scenes.  We have programs that pick
up objects and assemble them, programs that navigate a vehicle
avoiding obstacles, and programs that analyze pictures returned
from Mars.

	Projects with similar accomplishments exist at M.I.T., at
Edinburgh University in Scotland, at at least one laboratory in
the Soviet Union, and at several industrial and university
laboratories in Japan.  The Jet Propulsion Laboratory of N.A.S.A.
in Pasadena, California has a project aimed at a Mars Rover.  This
is would be landed on Mars and would be able to travel on the
Martian surface on command from the Earth, take pictures, and
pick up samples and analyze them.  Because of the long transmission
time for signals between the Earth and Mars, the rover must
avoid obstacles under its own control.  It would take to long
if each small movement had to be based on a picture transmitted
to Earth.

	The Defense Advanced Research Projects Agency supported
a five year study of speech recognition which achieved considerable
advances in this art.  Vocabularies of several hundred words can
be handled in very restricted domains of discourse.

	The state of the arts required for a domestic robot may be
summarized as follows:

.item←0
	#. Building a robot.  No-one has built a robot with two
arms that can manipulate objects and can also move about on legs
or wheels.  It is undoubtedly possible to build it, but the development
costs might be several million dollars if the weight were to
be kept below 500 pounds and human-level strength achieved.  Since
the single robot arms now sell for more than α$20,000, it seems
unlikely that a α$4,000 price could be achieved for the physical
part of the robot.

	#. Vision.  Television cameras can be provided, but a program that
could recognize the parts of a guest well enough to help him off with his coat
without injuring him seems beyond the present state of the art.
Certainly, nothing that difficult has been achieved.
This is an advertised Quasar function.  An adequate television camera
would cost about α$1,000, but it isn't clear that this would be
adequate.

	#. Speech.  Something useful could be achieved at moderate cost,
but not something that could be useful in teaching the children to
speak French.  The vocabulary would have to be limited to that required
for carrying out domestic tasks.

	#. Control of motion and manipulation.  Control of motion
and recognition of obstacles sufficient to avoid running into children,
animals and furniture is somewhat beyond present achievement.
Even after successful preliminary experiments, it would probably take
several years to certify a robot as safe enough for use in a home.

	#. Common sense.  A useful domestic robot must have a common
sense understanding of the world.  For example, it must know what to
expect if a child starts to pull a heavy object towards the head of
a table or it hears thunder and the windows are open.  This area is
very backward and requires major breakthroughs.

	#. Achievement of goals.  In order to obey a command, a
robot must divide a task into subtasks and overcome the obstacles
required to accomplish the subtasks and make sure that in accomplishing
one of the subtasks it doesn't spoil the preconditions for accomplishing
another.  This has been a major area of research in artificial
intelligence and much has been achieve, but again it isn't close
to ready for domestic use.


.bb The Quasar Promises

	My attention, that of Professor Marvin Minsky of M.I.T.,
and that of other members of the artificial intelligence community
were drawn to Quasar Industries by newspaper articles appearing in
July 1977.  These articles reported the claim to deliver domestic
robots for α$4,000 in four years, and accepted as genuine the
conversations of Klatu alias Sam Strugglegear.  At first, we were
not inclined to do anything, because these claims don't interfere
with our work.  However, Professor Minsky pointed out that if anyone
has a responsibility to expose hoaxes in a given area of science, it
is the practitioners of that science.

	While we were trying to decide what to do about matters distant
from either M.I.T. or Boston, we received a report from Brian Reid and
other graduate students at Carnegie-Mellon University of a Klatu
demonstration at Pittsburgh department stores, and their detection
of remote control.  A similar report soon came from George Dodd of
General Motors about a Detroit demonstration.  It became clear that
Quasar Industries claims that Klatu's speech recognition is genuine
when they can get away with it and admits remote control when
pressed by knowledgeable people.

	The Stanford University press release was based on this
information.  The letter to the Justice Department was based on the
further information that Quasar answers inquiries by suggesting
an interview to determine compatibility with the robot.  This suggested
an opportunity for fraud that required investigation by legal authorities.

	Our knowledge of robotics, as outlined above, told us that there
is no present basis for marketing a robot domestic servant.  It is
conceivable but unlikely that a small group could be that far ahead of a
dozen research laboratories all over the world.
It is especially unlikely, because they would have to be ahead not merely
in one scientific area but in four or five.
It becomes very hard to believe when we see that they are demonstrating
a robot that is far behind laboratory and industrial
robots in mechanical sophistication, i.e. it can only wave its arms whereas
industrial and laboratory robots have been constructing complex objects
out of parts, e.g. an IBM robot can partially assemble a typewriter.
Finally, when we discover that Quasar is faking speech recognition
and motion in carnival-like demonstrations in which the attention of
the audience is distracted by having the robot pretend to try to
make dates with girls, we are left with one conclusion - it's a fake.

	The only remaining question is whether any laws have been
broken, and we leave that to the courts.  However, I am somewhat doubtful
that we are entitled to legal protection merely from their making false
claims about their future plans.  The government should not be involved
in deciding what people can say they expect to be able to do.  Perhaps
consumers would be adequately protected if Quasar were obliged to
prove its claims or else to agree to take deposits only into an escrow
account if at all.