Hans Moravec is a Principal Research Scientist at the Carnegie Mellon
University Robotics Institute. Moravec, whose interest in robots extends back
to his childhood, discusses his intriguing and personal views on robots—from
the current state of technology, to today's bomb defusing machines, to the
capabilities of robots in the next century. NOVA spoke to Dr. Moravec in
NOVA: Can you give me a good working definition of what a robot is and how it
differs, say, from a machine tool or a computer?
HM: Robots are machines that are able to do things that previously had been
associated only with human beings or animals—such as the ability to
understand their surroundings and plan their actions.
NOVA: Most hazardous duty and bomb disposal robots seem to be tethered in some
way to a human operator—either by an actual cable or by a radio signal. Do
they qualify as robots in your mind?
HM: Not really. We call remote control devices "robots" because I guess they
have some of the characteristics of autonomous machines, at least physically.
The research going towards those machines does contribute to the research
towards autonomous machines, but they're lacking the brains. They have the
bodies of robots, but not the brains.
NOVA: I read an interview in which you said, "Today's best robots can think at
HM: A few years ago that was correct. In fact, we're probably a little above
insects now. I make a connection between nervous systems and computer power
that involves the retina of the vertebrate eye. The human eye has four layers
of cells which detect boundaries of light and dark and motion. The "results"
are sent at the rate of ten results per second down each of those millions of
fibers. It takes about a hundred computer instructions to do a single edge
motion detection. So you've got a million times ten per second, times a
hundred instructions; that's equivalent to a billion calculations per second in
the retina. That is a thousand times as much computer power as we had for most
of the history of robotics and artificial intelligence. So we were working
with one-thousandth the power of the retina, or roughly what you might find in
an insect. What happened by the end of the '80s is the cost of one MIPS [the
standard by which computing power is measured] dropped down to $10,000 or
below. And at that point the power available to individual research projects
started climbing. The climbing rate is pretty amazing because computers have
been doubling in power. It was once about every 24 months. In the '80s it was
once about every 18 months. Recently it's been closer to every 12 months. So
we now have in our research projects machines that can do 300 MIPS, and soon
we're going to have a thousand. So we're at the stage of small vertebrates.
Over the next few decades, the power is going to take us through small mammals
and large mammals. I have a detailed scenario that suggests that we get to
human level, not just in processing power, but in the techniques by the middle
of the next century.
NOVA: Can you walk us through what you think robot evolution will look like?
HM: Well, I imagine four stages. I think we're just on the verge of being
able to see machines that work well enough that they'll become the predecessors
to the first generation of mass-produced robots—that are not toys. I have a
particular one in mind: a small machine that could be a robot vacuum cleaner
which, with a thousand MIPS of computing, is able to maintain a very dense
three-dimensional map or image of its surroundings. It will be able to both
plan its actions and to navigate, so that it knows at every moment where it is
and is even able to identify major pieces of furniture and important items
around it. So—a small machine, small enough to get under things and to find
its own re-charging station and to empty out its accumulated dust from time to
time. That's the research we're doing and I think sometime within the next five
to ten years we'll have something like that—and its successors will become a
little more capable. They'll have a few more devices and be programmable for a
broader range of jobs until, eventually, you get a first generation universal
robot, which has mobility and the ability to understand and manipulate what's
going on around it.
NOVA: What do you mean by universal robot?
HM: It's a machine which can be programmed to do many different jobs. It's
analogous to a computer, which is a universal information processor, except
that its abilities extend to the physical world.
NOVA: Okay, so we've got the first generation of universal robot.
HM: Right, the time schedule is around 2010 now. Every single job a robot
needs to know has to be built into the application program and when you run the
program, the robot acts in a pretty inflexible way. Still, it's perceptual in
motor intelligence. It's comparable to maybe a small lizard.
NOVA: What types of tasks might we expect these robots to do?
HM: Well, things like floor cleaning and perhaps other kinds of dusting—delivery. The kinds of factory tasks that robots are now doing should be
possible for universal robots, like the assembly of things. But because this
kind of robot should be mass-produced, the range of tasks will be probably
larger than anything that exists today, and the machines will be cheap enough
to be used in places that you can't use robots today. I imagine car cleaning
tasks and bathroom cleaning and lots of other things that will depend on the
ingenuity of the programmers.
NOVA: What happens next?
HM: All right, so now we come to a second generation. The second generation
machines will have a computer that is maybe 50 times as powerful as the first
generation and is able to host programs that are written with alternatives.
For example, picking up an object, which might be part of some big job, could
be done with one hand or with another hand of the robot. Each of the
alternatives has associated with it a number, which is the desirability of
doing the step that way, as opposed to doing it an alternative way. Those
desirability numbers are adjusted based on the robot's experience. And the
robot's experience is defined by a set of independent programs that are called
conditioning modules, which detect whether good things or bad things happen.
For instance, you might have one module that responds to collisions that the
robot undergoes, and produces a signal that says, "Something bad happened."
Another one detects if the task the robot was doing was finished, or finished
particularly quickly, and signals that something good happened. If the
batteries are discharged—that's bad. Perhaps if they're kept in a good
state—that's good. Gradually the robot adapts, because of this internal
conditioning, to do things in ways that work out particularly well and avoid
ways that have caused trouble in the past.
NOVA: So it's capable of rudimentary learning.
HM: Yeah, it's conditioning. And, you can even imagine these conditioning
modules being tied to external advice. For instance, one module might respond
to your saying, "good." And another one to your saying, "bad." And so you can
direct the robot to act a certain way, as opposed to another. You could, if
you wanted to, train it in the way that you might train a dog—by repeatedly
saying "good." It's Skinnerian training basically.
NOVA: What sorts of tasks would the second generation robots be able to do,
that the first generation ones wouldn't be able to do?
HM: They'll be used presumably in the same kinds of situations but they'll be
much more reliable there. They'll be much more flexible. Take a first
generation robot that's putting away the dishes. Maybe it's programmed to
always grab certain objects in a certain way. And it has motion planning and
collision avoidance. But still, some things may have been overlooked in that
program. And perhaps in your particular circumstance, whenever it goes into a
certain cabinet, it ends up always catching its elbow on the door. The first
generation robot will never learn. It will just keep making that same mistake
over and over and over again. The second generation robot will gradually learn
to do things a different way, maybe use a different arm or reach in a different
manner. Essentially the robot will tune itself. And it will be much more
pleasant to have around, because it won't be making lots of little mistakes
that the first generation robot will be.
NOVA: What year is it now?
HM: Approximately 2020. Each one of these generations is about a decade. The
first generation robot may be comparable to a small lizard. The second
generation robot, with it's limited trainability, may be something comparable
to a small mammal. The third generation robot is the first really interesting
one. It's predicated on there already being quite a large industry, based on
these earlier generations, which is able to support the development of a major
module for these machines—the ability to model the world, a "world
simulator." This simulator allows the third generation robot to make many
mistakes in its mind, running through scenarios in simulation rather than
physically. The second generation robot learns, but quite slowly. It has to
make a mistake many times before it learns to avoid it. And when it's tuning
up, when it's getting good at something, it has to do it many times before it
really really gets good at it. The third generation robot runs through the
task many times, mentally, and tunes it up there. So when it first goes to do
something physically, it has a good chance of doing it right.
NOVA: It thinks before it acts.
HM: Right. And, the simulator is a big deal, because it can't be strict
physical simulation. That's why it's still computationally out of reach, even
with the kind of computer power, I imagine for then. What it needs is
something that's closer to folk physics: basically rules of thumb for every
different kind of object that it's likely to encounter—because one of the
things the robot will have to do is roll into a new room and make an inventory
of everything that it sees around it, so that it can build a pretty accurate
simulation of that room, so it can then do its mental rehearsals. It will have
to identify the objects that it sees and then call up generic descriptions of
what those objects are and how they behave when they're interacted with, and
how to use them. Building this generic database is a major effort.
NOVA: What sorts of things would it need to know about, say, cutlery?
HM: Well, first of all, where cutlery might probably be located, how to pick
up the individual pieces, roughly how heavy to expect them to be, you know, how
hard they can be gripped. Because, for instance, if it sees an egg and wants
to pick it up, it has to know to be gentle with it. If it sees a knife, it has
to know in advance to pick it up a little harder because, if it picks it up too
lightly, the weight will cause it to fall down. And, of course, it can learn
by making actual mistakes, but the whole idea of the simulation is to avoid
those mistakes, in the first place, whenever possible. Now, what's really
interesting about the third generation robot is that, besides having this
physical model for things in the world, it will also have to have a
psychological model for actors in the world, particularly human beings. It
should know that poking a sharp stick at a human being will produce a change in
state of the human being. They will probably become angry and if they're
angry, they're likely to do certain things which will probably interfere with
the robot's tasks. And, since these robots will probably be used, among other
things, as servants working among people, it will be useful for the robot to
have an idea of whether its owners are happy or unhappy and choose actions that
improve the happiness. Basically, machines that make their owners happier are
likely to sell better, so ultimately there'll be market pressure. Third
generation robots should be able to deduce something about the internal state
of the human beings around it—if a person seems to be in a hurry or if this
person seems to be tired. You can probably deduce that from a modest
observation of body language.
NOVA: They're doing some mood recognition already at the M.I.T. Media Lab,
HM: Yeah, that's right. Another interesting thing that a third generation
robots is able to do, is to provide a description of things. You should be
able, with a small additional amount of programming, to generate some kind of a
narrative. You may ask, "Why did you avoid going into that room?"—"Because
Bob's in there and I know he's upset and my moving around him will probably
irritate him further." Now, here's a funny twist. You can have conversations
with the third generation robot where it seems to believe that it's conscious;
it talks about its own internal mental life in the same ways that the people
do. And so, I think for practical purposes, it is. So the third generation
robot can analyze. It's comparable to maybe a monkey. When it simulates the
world, it's always in terms of particular objects or particular sizes and
particular locations. It doesn't really have any ability to generalize. Its
ability to understand the world is very literal.
NOVA: It sounds sort of sweet.
HM: It is. That's right. You wouldn't expect any deviousness at all.
NOVA: Take us to the fourth generation robot.
HM: All right. Basically, if the third generation robot is something like a
monkey, the fourth generation robot becomes something like a human being—actually more powerful in some ways. The fourth generation robot basically
marries the third generation robot's ability to simulate the world with an
extremely powerful reasoning program. Even today, we have reasoning programs
that are superior to human beings in various areas. Deep Blue plays chess
better than just about everybody—and various expert systems can do their
chains of deductions better than just about anybody. And of course, for a long
time, computers have been able to do arithmetic better than everybody, for
sure. But there is a certain limitation that these programs embodying
intelligence have had, which is they really haven't been able to interact with
the physical world. When a medical diagnosis program talks about the symptoms
of a patient, it's only processing words. And when it comes up with a
recommendation, again, it's more words.
NOVA: But now the reasoning will be connected to physical experience or
HM: Right. The reasoning that the fourth generation does is greatly enhanced
by the third generation robot's ability to model the world. So physical
situations that the robot thinks about in its simulation can now be abstracted
into statements about the world. And then inferences can be drawn from the
statements, so that the robot can come to non-obvious conclusions. For
example, it might be able to figure out from running several examples that if
it takes any container of liquid without a lid and turns it upside down, the
liquid will spill out. The third generation robot would need to figure out not
to turn this glass over, not to turn this jar over, not to turn this pitcher
over etc. whereas the fourth generation robot would be able to infer it.
That's just a example. Fourth generation robots would be able to do much more
complicated tasks—and do them probably better than humans, because really
deep reasoning involves long deductive chains and keeping track of a lot of
details. Human memory is not that powerful.
NOVA: Can you envision, in the future, a robot being better than a human at
finding and disarming a bomb?
HM: Sure. You can imagine that for the near future. The sensors that a robot
can bring can be tuned for the task. For example, radar can penetrate various
kinds of materials—depending on the frequency you use, you can see through
walls. So, simply the ability to see into a package would certainly make a
robot a better bomb detector.
NOVA: Can you envision a robot understanding the psychology of a terrorist
better than a human being?
HM: Well, ultimately. Now we're talking 40 or 50 years from now, when we have
these fourth generation machines and their successors, which I think ultimately
will be better than human beings, in every possible way. But, the two
abilities that are probably the hardest for robots to match, because they're
the things that we do the best, that have been life or death matters for us for
most of our evolution, are, one, interacting with the physical world. You
know, we've had to find our food and avoid our predators and deal with things
on a moment to moment basis. So manipulation, perception, mobility - that's
one area. And the other area is social interaction. Because we've lived in
tribes forever and we've had to be able to judge the intent and probable
behavior of the other members of our tribe to get along. So the kind of social
intuition we have is very powerful and probably uses close to the full
processing power of our brain—the equivalent of a hundred trillion
calculations per second—plus a lot of very special knowledge, some of which
is hard-wired, some of which we learned growing up. This is probably where
robots catch up last. But, once they do catch up, then they keep on going. I
think there will come a time when robots will understand us better than we
understand ourselves, or understand each other. And, you can even imagine the
time in the more distant future when robots will be able to host a very
detailed simulation of what's going on in our brains and be able to manipulate
HM: I see these robots as essentially our off-spring, by unconventional
means. Ultimately, I think they're on their own and they'll do things that we
can't imagine or understand—you know, just the way children do.
Photos: (1) Gene Puskar; (2-3) Jesse Easudes.
Future Robots |
Hazardous Duty |
Bomb Squad Home
Editor's Picks |
Previous Sites |
Join Us/E-mail |
About NOVA |
Site Map |
PBS Online |
NOVA Online |
© | Updated November 2000