Rodney Brooks, "Artificial Humanity” - God and Computers: Minds, Machines, and Metaphysics (A.I. Lab Lecture Series)

ANNA: Welcome to the fourth lecture of the lecture series, "God and Computers: Minds, Machines, and Metaphysics." The first lecture was about the human factor, given by Paul Penfield. Then we heard about animals by Mark Hauser-- animals and consciousness, animal and ethics. Last week, we heard about brains, and brain function, and where spirituality might be localized in them or might be able to localize them. And today, now, we hear about robotics.

Before we can start about that, just briefly, one thing-- if you happen to have this brochure-- I have also some left. The discussion meetings at Harvard Divinity School-- the first one is on Monday, at 12 o'clock, at Harvard Divinity School-- 45 Francis Avenue-- run by Harvey Cox and me. So you are very welcome to bring your questions and discussions you cannot address today to this discussion group on Monday. The other dates you will all find in that little brochure.

And now I'm very pleased and happy to welcome to Rodney Brooks. And with the introduction, it's always a little bit difficult if you actually have to introduce your own boss. So I am not quite sure how to do that. But Rod is currently the director of the Artificial Intelligence Lab here, at MIT. He holds a PhD in Computer Science from Stanford, where he was mostly interested in vision.

And when he came to MIT, in 1981, he was still mostly interested in classical AI problems, but, slowly, his focus began to shift. Since AI had failed to reach its goal of human-like, intelligent machines, Rod broke with most assumptions of his field and started anew. Based on the work of Maturana and Varela-- Varela, by the way, will give his talk in two weeks here-- Rod started to perceive intelligence as being in the world, as interaction, as embodiment.

He then built creatures-- insect-like robots-- which were far better than anything else which had been built before in navigating an unstructured world and natural environments. For his work, he received numerous awards, became, in 1996, a fellow in the American Association of the Advancement of Science, and is, today, accepted as the so-called father of embodied AI. And recently, he even became a movie star.

His movie, Fast, Cheap, and Out of Control-- named after a research paper of his-- is currently running at Kendall Theatre and I can only highly recommend you all to go there. Rod's latest attempt to rebuild intelligent creatures goes a little bit beyond insects, hence the title of his paper, "Artificial Humanity." And so I'm very happy, and glad, and pleased to welcome you here.

BROOKS: Thank you. Well, this talk is a little different from the ones I normally give-- even less technical detail than normal. And the title is "Artificial Humanity," but it's a non-linear talk because, to be honest, I couldn't figure out how to fill up a whole hour. So I've got some little diversions along the way, which aren't necessarily straight along the line here. But the question is, to me, whether we can ever have humanity in robots-- whether we can think of them in the same way we think of other humans.

And that's all tied up with lots of beliefs we have and lots of insecurities we have. I know that, in the last few hundred years, we've seen mankind's retreat from specialness. Galileo and others around there let us know that the Earth was no longer the center of the universe. And so this piece of rock that we're standing on wasn't particularly special. And then Darwin came along and we saw that humans and animals have common ancestors. Their origins are not that different.

Crick and Watson, with DNA and the mechanism of life, means that humans and yeast are pretty similar. And you see, sort of, that we're retreating here from all these special things. And each of these retreats has been met by a lot of argument and a lot of fear. And as you get further and further down here, and as you get further and further south in the US, not everyone has accepted all these things.

Then if you look back in the middle of '50s, you have [INAUDIBLE], McCarthy, [? Neil Simon, ?] et cetera-- and Turing, of course. Human thought is the same as computation and, hence, fits on machines. So the thought process, the logic process, was sort of taken away from us. And as we've understood biochemistry better and better, we find that we're collections of tiny machines. And where the essence of humanity is, is harder and harder to find.

And very recently, just this year, we started to see that human flesh and body plans are subject to technological manipulation. Was it over the weekend? Yeah. I think, over the weekend, there were the stories about the headless frogs that were grown in England. And then a whole raft of new stories-- that Lewis Wolpert, in London, responded to-- about whether we could now start building headless humans as a way of growing new organs and all the fears that that produced.

So we're sort of retreating to be less and less special. And each of those retreats comes with quite a bit of consternation, quite a bit of argument. And the one I want to talk about today is whether robots can feel, love-- and since God is in the title of this-- worship, and have souls. So as Anna pointed out, I've been playing with humanoid robots recently. This is a version of our robot, Cog, from a little while back. It was built here at the AI Lab.

And our goal was to build a robot that develops and acts in the world in the same way that humans develop and act in the world. And why human? Well, we wanted it to have similar sensory motor experiences, very much influenced by Mark Johnson and-- who's the guy at Berkeley? Women, Fire, and Dangerous Things.

AUDIENCE: Lakoff.

BROOKS: Lakoff, et cetera. But I have to admit, this particular motivation is the one that worries me most, the one that makes me think that, well, maybe we're engaged in this exercise and a bit of cargo cult science. Do people know about the cargo cults in New Guinea? This is one of the little diversions. Anyone not know about the cargo cults? OK. So during the Second World War, both the Allied troops and the Japanese troops had lots of fighting going on in New Guinea.

And they would calm, and they would clear out some land, and build a control tower. And then these silver birds would come down and disgorge lots of supplies. And so after all the troops left, a lot of the native tribes started flattening out areas of land, building bamboo, control towers, and sitting up there, waiting for these silver birds to come down from the sky.

And so I do worry a little bit that in building this robot to have similar sensory motor experiences, maybe we're engaging in that sort of cargo cult, as it's called, because it doesn't really have a lot of the stuff that humans have.

But one thing that has certainly turned out to be true is that humans interact with these robots in natural sorts of ways. They can't help themselves. And this is not something we've studied in great detail, so it's rather anecdotal. And I'll only give anecdotal incidents today.

But humans just can't help but interact with these machines that look like humans, in human like ways. And that, I think, is the key to the question of answering whether robots will eventually have souls. And that's what I'm going to try and get to today.

People just find themselves interacting with it, like a human. Let me show you the first video, just to show you Cog doing some stuff. And compared to two Star Wars movies, it's a little disappointing. But it took us a long time. So I'm supposed to press-- do I have something up there? Yeah. So here, you see there's an old version of the robot head up there. It keeps putting things in its hand. It's attracted by the motion. It's looking at things.

We built the arms with a series of elastic actuators developed by Gill Pratt and Matt Williamson, which enabled us to interact with this robot. This is the head system. You see the eyes saccading, from place to place, very rapidly, at about human speeds. There's a wide angle lens and a narrow angle lens in each eye. This is saccading to motion. You see the eye is rapidly moving from place to place. It's a smooth pursuit, where the eyes are smoothly following something.

These are all the basic sorts of visual operations that people do. You can't move your eyes smoothly from side to side unless you've got something to track, it turns out. This is the the vestibular system simulating the inner ears. Here, it's not switched on. As we move the head around, the eyes just waggle around all over the place. When we switch on the inner ears, now the eyes are stabilized, as we move the head unpredictably and we take locked onto one position. And that's very important to you, to stabilize your vision.

Now here, you see the eyes saccade somewhere and then the neck tries to get the eyes back into, roughly, the center of their range of motion. And an efference copy signal is sent to the eyes to compensate for the motion. When we put this in, the robot started to feel much more human-- the way it looked at us. Here's Matt Williamson-- designer. This is with his old arm. He's just showing some basic, infant-like reflexes. This is withdrawal reflex. It feels the touch and pulls its arm back.

Infants have the grasp reflex, withdraw reflex, et cetera. And in order for Matt to get his Master's thesis, I made him prove that the arm was safe to interact with. And this is some new arms. We've got a bunch of new stuff with this, but I'm not going to show that today. It's not relevant. This is a tape we made up for AAAI a couple of months ago. This shows you some physical coupling through the system. The body is important, as with physical bodies of humans.

And recently, Matt has had Cog playing with Slinkys and [INAUDIBLE] feeling pendulums swinging back and forth. This is Cog learning to reach-- learning hand-eye coordination. It saccades its eyes to some point, and then reaches out its hand, moves its hand to see how well it got to the center of the location of its eyes, and then learns how to move its hand. So it sits there for a few hours, reaching out to places it looks. This is its point of view.

You'll see the saccade happen in a second. And during the saccade, we have to suppress the motion detection, just like happens in the human visual system. And then it puts its hand out and it moves its hand around to see where its arm ended up, so it can learn how to operate. And after about three hours of learning, here, Brian [INAUDIBLE] gives it a motion cue. It reaches out. The hand wasn't working at this point.

Now he's going wave. Watch the eyes up there. And you'll see the eyes, in a second or two, saccade over to that motion. There, they saccaded. Now it's going to try and reach out to that. This is the result of its learning for a few hours with no prior knowledge of the kinematics, or dynamics, or whatever. It's all learned. And here's an interesting case where Cynthia Farrell, one of the designers of the robot, was just trying to engage some motions of the arm.

And when we looked at the videotape, we saw that the robot and she were playing-- taking turns. And we weren't planning on putting turn taking into the robot for quite well. But she, even as a non-naive observer, couldn't help but get into that dynamic of playing-- oh, what happened? Getting into that dynamic of playing a game with the robot. And that's going to be an important point later on.

Now here, I'll do another aside. We're building this humanoid robot. There are lots of other people building humanoid robots, particularly in Japan, right now. It's become quite an industry. And there's been some pretty surprising developments recently for robotics people in the world. Honda Motor Corporation has had a secret project-- it's just like a James Bond movie-- for 10 years, developing this robot called P2. P1 was just a pair of legs with a box on it. It looks pretty weird.

But they were completely secretive about this. No one knew what was going on. Last October, I started hearing a couple whispers. And then, in December, they announced. It's 250 kilogram, the first biped that can walk without a tether, but it's mostly mechatronics rather than AI research, at this point. But MITI-- the Ministry of International Trade and Industry-- was talking about a plan to build lots of humanoid robots and give them to universities.

And Honda didn't want to be left out of the action, given that they had put $100 million into this over 10 years. They announced that, last December, by about three or four weeks ago, they had their new model out-- the P3-- which is smaller, weighs about 110 kilograms with a 30 kilogram payload. And so this was done in just nine or 10 months. They've now got 100 people working on this project, building humanoid robots.

And Honda has suddenly become a major robot company in the eyes of the Japanese public, through this project. There's some talk about having the P2 or P3 at the opening of the world-- what do you call it? What's the world soccer tournament thing?

AUDIENCE: The World Cup.

BROOKS: The World Cup. Having it out there on the field, kicking the ball. It won't work, if the ground is soggy, by the way, because the algorithms that they use our rather fragile. But lots of other people in Japan are also building humanoid robots. This is Waseda University with a couple of their humanoids. They now have 100 people working in their humanoid research laboratory. And there are other humanoid research laboratories. There's the ones at ETL, there's the ones at Tokyo University, the HARP Lab, et cetera.

So it's become a big industry. And there's going to be lots and lots more humanoid robots. Most of these Japanese robots, at this point, don't have much in the way of intelligence or emotional models, et cetera, although one of the robots at Waseda does have a model of the amygdala, the hippocampus, and a whole bunch of other inner, emotional sections of the brain. Perhaps they're trying to put that together. That was an aside.

Let's get back to whether we're going to be able to build robots with humanity in them, whether we're going to respect them. You just saw a videotape of our robot. But I need to come clean a little bit, I think, in this talk.

I haven't been to any of the talks previously in the series because I've been out of town. But Anna tells me that, at the first talk, when Paul Penfield spoke, someone complained that they wanted to hear what an atheist had to say. And she pointed that I was going to be the atheist coming along to talk. So my assumptions are that there is a completely mechanistic explanation for everything in the universe.

I believe we're all made up of gazillions-- and I chose that word because it just seemed large-- mindless, soulless robots-- molecular machines. The little, biochemical molecules are these machines. We are made up of robots, as [INAUDIBLE] pointed out. And I believe that every human interaction, in principle, can be reduced to such mindless explanations. Now is this how I live my life? No, I don't live it this way at all.

I don't think of my kids as, oh, they're just those little, mindless, gazillions of machines. I think of them as one big, mindless machine. That's not true. I operate in a totally different level. And I must admit, I was perplexed, for a long time, by scientists who are also religious. For a long time, it just didn't seem, to me, to make sense that someone could be religious and, at the same time, be a scientist. I couldn't see that duality.

And it was only fairly recently that I realized that that's exactly what I do-- I do live in this dual world, where I have this level of explanation, but, for my everyday life, I operate at a very different level. And although they may deny it, I think all atheists operate this way and I think all scientists operate this way-- having different parallel belief systems that they operate with. So I'm atheist, but I'm not-- I've got to be careful to choose my words here. No, I won't use them. I'm one of the good atheists.

Now related to this, it seems to me that most scientists engage in a certain cultural constructivism. All researchers base their work on some, usually unstated, dogmatic beliefs. And this gets a lot of people upset because a lot of people believe that they're in search of truth and there is one truth. But I think that when you push people who speak that way, many of those beliefs are in the scientific method, which has a certain religious aura about it when a lot of people talk about it.

Or whether they believe, as I do, that humans, ultimately, have mechanistic explanations, when it gets down to it, we can't explain it beyond that in any logical way. We base our scientific lives on some set of assumptions. And I've observed the challenges to these implicit beliefs are often met with hostility. And for those of you who've been on the mailing lists around here, you certainly saw that with the announcement of God and Computers, the course that goes along with this series of talks.

So those two slides are just sort of saying where I come from in this. Oh, boy. That sounds awful '70s, doesn't it? So can we have artificial humanity? Can we ever build robots where we, people-- our sort of machines, as distinct from those sort of machines-- will all agree that these robots are things to be empathized, whether they're things we should pity, when appropriate-- whether there will ever be an appropriate time to pity a robot-- actually I did.

Do you remember-- was it RoboCop? Before RoboCop sort of got invented, they had this legged machine that sort of blew everyone away. Do you remember that? And they chased it out, and they chased it down the stairs, and it fell down, and its legs were quivering. I was building walking robots at the time. I really felt for that robot. So whether we should protect these robots, when necessary-- when the bad graduate student is going to switch it off, we should stop them.

And then, ultimately, whether they should be equal before the law. And, as we all know, being equal before the law is very different from being equal before our hearts. And that's an even bigger step to get to. So can we have artificial humanity? From a religious point of view, I would think, if we had artificial humanity, we would like to say that these things had souls. I can talk about this, as an atheist.

So I believe that, in the same sense that humans have souls, it would be impossible to build humanoid robots that have human-level interactions without them having souls. But I use my definition of a soul, which we'll get to a little later. And we won't even have to try to give them souls, they'll just have them. And they'll have them because of the way we feel about them. That's a sort of a tricky question. So a simpler question is, can a robot be afraid?

Now I think people are willing to say we can make robots act as if they're afraid. We've seen robot actors, in movies, act as if they're afraid. We can make robots that seem to be afraid. We can make robots that simulate having fear. And a while back, people used the same sorts of caveats when they talked about reasoning systems in computers. But I think, today, most AI researchers are willing to say that robots or programs can reason about facts, they can make decisions, and they can have goals.

This simulate, act as if, and seem have been replaced by can. But I'm not sure that AI researchers, in general, are willing to say that robots can be afraid, because I think AI researchers are still-- they've got that specialness. They've been pushed back, by their own work, into giving up a lot of stuff.

But this visceral fear that we all know and feel, are we willing to say that a machine is going to have that sort of real, visceral sort of fear? Or is that something it will just seem to have? It will just be a bunch of, heaven forbid, C programs doing stuff. Well, what I want to claim is that, ultimately, the robots will be viscerally afraid. And it's a matter of us accepting that rather than any great technological breakthroughs necessary.

So in this sense, this is a little disappointing talk because I'm going to say we don't need technological breakthroughs. It's sort of ultimate cop out talk. So I want to examine a few other people who talk about some of these similar sorts of issues, although not exactly this issue. And I think there are some generic ways in which they go wrong. One way they go wrong is through an implicit rejection of mechanistic explanations, where that rejection is often denied, but, nevertheless, it's there.

And I'll show you an example in a minute. And they conserve the specialness of humankind that way and they rationalize it as a scientific argument, but, in fact, they're doing something sneaky. They're getting rid of a mechanistic explanation by dressing it up as though it is a mechanistic explanation. And I'll give you an example. And the other thing, which is much more common, is amongst AI researchers. It's an adoption of a higher mechanism.

Researchers want to maintain a purely mechanistic explanation, but they can't face reductionism to current models of the universe, so they invent or wish for-- and I'll show you examples of both-- some super mechanism that would be rather special. And we don't know about it yet, so that sort of maintains the specialness of humans. First case-- implicit rejection of mechanistic explanation. And the great example is John Searle, from Berkeley. So here's John Searle.

And the web is great. I just went out, and looked, and I found John Searle-- a picture of him. And he even fit the clip art. So Searle makes this argument-- suppose he's in a room, and he's got some instructions to follow, and someone feeds him a piece of paper with some Chinese symbols on it. And he doesn't know Chinese. I don't know Chinese. I found these on the web, too. I have no idea what this says, but that fits the story.

And I have no idea what the output says. It should be related, but who knows. Anyway, so Chinese symbols come in. And this is a question-- he follows the rules, and he outputs some answer by following these rules, mechanistically, and he says, see, John Searle still doesn't know Chinese. He can act as if he knows Chinese, but it's not the same as knowing Chinese.

But what he misses is that the whole system does know Chinese-- John Searle, and his pencil, and his paper, and his book of rules, and his procedure that he's running does no Chinese. But he wants the Chinese to be in his head or, otherwise, the system doesn't know Chinese. And I think his argument maps pretty well to the argument that, well, horses can transport things, but a car can't transport things because, when you look inside a car, there's no place where the transportation is happening.

There's wheels, and there's gasoline, and some engines, and there's no horses there either. So it can't be transporting stuff like a horse transports stuff. And this is the same argument he makes, because he wants to find the understanding in a component. But he doesn't insist that the same thing happens for humans or for animals because he says, well, animals are animals.

Recently, there was a TV program where they interspersed something I said with something Searle said. They interviewed me. I didn't know Searle was going to be in the program. And they got me to say that if it walks like a duck, talks like a duck, smells like a duck, it's a duck. And Searle came on and said, if it walks like a duck, talks like a duck, smells like a duck, it ain't a duck, because a duck's a duck.

And he says the same thing about intelligence-- intelligence is intelligence. It's only in humans, therefore it cannot be in other machines. See? I proved it. Because he sort of flushed away that mechanistic explanation, but he denies that he flushes it away. Roger Penrose, on the other hand, at Oxford, he wants everything to have a mechanistic explanation. He actually misunderstands Godel's theorem and Turing computability.

And I think he does that because he wants to maintain the specialness of humans-- that humans can prove theorems that mere machines can't. And if you read his description of this, there is a misunderstanding there. Gerald Edelman, by the way, makes the same mistake in his analysis of the human kind. So this is not uncommon-- looking at Godel's theorem and saying we're better than Godel's theorem.

So his conclusion is that people can't be computers, but he wants everything to be mechanistic. So people are more complicated than ordering machines and consciousness is mysterious, but he really wants mechanistic explanations. So what's he to do? Well, he's a physicist, so he says, well, quantum mechanics is more complicated than ordinary physics and is mysterious-- hey, they must be the same thing. And really, I don't think there's anything more in his argument. It's this wishing for some other explanation.

David Chalmers, now at UC Santa Cruz, a philosopher, comes at it a slightly different way. He talks about consciousness. And he's quite seriously. He did organize that consciousness extravaganza in Arizona last year, but he is serious-- besides that. And he talks about mass force, et cetera, and natural physical kinds. You can't reduce them simple things. They're just things in the universe. They're natural kinds. They're stuff.

And his argument is consciousness is yet another natural kind. It can't be reduced something simpler. Therefore, there's no need or way to explain it. So he's maintained the specialness by having consciousness as being some natural kind. Well, that may be, but I sort of doubt it because all of other natural kinds display some sort of observable interaction. So we would expect if consciousness was a natural kind-- telekinesis or some interaction between consciousness and other stuff, which didn't quite fit. We don't see that.

So I suspect that that's not going to work. Now, of course, I think everyone's guilty of this in other ways. I'll tell you my own version of this, my own folly. If you look at an engineered system and you look at a biological system, people don't make mistakes. Even young kids, pretty much, don't make mistakes. They get fooled by some things, when they're very young, but as they get a little older, they get pretty sophisticated. They don't make mistakes and think that's a live animal, when it's really a robot or something like that.

They can make that distinction. Robustness, the generality, the adaptability, the domain of the performance-- it all gives it away. So, it seems to me, biological systems are still fundamentally different from almost all our engineered systems. So my version of looking for the higher thing is something I call the juice. I really believe this, by the way. Is there something different in life? Not an essence of life, in the normal sense. My belief is we're not going to have to go outside of current day physics or chemistry.

But in the same way the idea of computation changed what we could think about-- before Turing came along and formalized computation, you could think about certain things. And after, you could think about a whole bunch more things. That notion of computation wasn't a change in the universe, but it enabled us to think about new things. If you took a late 19th century mathematician, you could teach them the fundamental ideas of computation in a few days.

And they wouldn't be holding their heads, saying, oh my god, this can't be, this is so foreign to me. It wasn't very foreign. It was just another idea that fit on top of 19th century mathematics. But it completely enabled thinking "how to" knowledge. So my version of what we're missing is there's some, conceptual juice-- I call it-- waiting for us to discover it-- some different way of thinking about organization and processes of complex systems that are in all these biological systems, at all sorts of different levels.

Either they're in at the molecular level-- in the maintenance of a cell-- they're in at the neural level, they're in at the genetic level, they're in at the immunological level. Some sort of organizing stuff that's in there-- sort of like computers inside mechatronic devices. That's there, but we just haven't got a way of describing it yet or a way of thinking about it, so we never think to put it in our artificial systems. So this is my version of this.

By the way, when I first talked about this-- two years ago, in Switzerland, at a workshop-- that night, a graduate student from Oxford was sitting at the dinner table with me. And he said, oh, yeah, I wasn't surprised to see you talk that way today, in the talk. I think that's the sort of ideas a lot of scientists have when they're in the subset of their career. So that was an aside, that last one.

Where am I going? Let's go back a couple slides. Where am I getting to? I'm getting to-- what will it take for us to be willing to say that robots can be afraid-- they can be viscerally afraid? And I gave those examples of what I think is wrongheaded thinking to see where we might go wrong in thinking about that. And one thing, it seems to me, is that, if we're really going to be thinking about robots being afraid, we're going to have to have some sort of models of emotions.

If we're to empathize with robots, we may need to be able to identify, physically, with them. And that's why I built this human shaped robot. But we may need to be able to identify emotionally with them, too. So can robots have emotions? And there's been quite a bit of work-- many people in this room have done work on this-- in putting emotions into robots or at least into software agents.

And I just want to go through and briefly recap what I think are sort of three different versions of emotional models-- surface level emotions, subsurface emotions, and emergent motions are three different ways of putting emotions into systems. The surface emotional models seem less satisfying, somehow, in terms of things having visceral emotions, visceral feelings. The primitive types are directly the emotions-- happiness is a number.

Jim Albus even had love as a 4-bit number in a paper on system, man, and cybernetics, only a couple of years ago. And then external events excite or depress certain emotional levels. These sorts of models have been around since the '60s, by the way. And there may be lateral inhibition mechanisms, so that you can't be happy and sad at the same time. If you're really happy, it pushes sadness down. If you're really sad, it pushes happiness down, et cetera.

And then there's some external reflection of the surface emotions. And, in fact, that's maybe all there is. And that's what this emotional model is-- it's just a surface model. Then the subsurface models-- here, there are some inner drives and needs that are the fundamental types, and the emotions sort of come out of those inner drives and needs being satisfied or not. So the level of satisfaction of these drives or needs determines the mood or emotional state of the system.

And then the emotional state may excite or inhibit certain classes of behavior of the robots. And it may change the way it operates. And then there might also be some explicit, external reflections of motions designed just to exhibit that to the external world. So you know what sort of mood robot is in, you know what sort of behaviors it's probably going to engage in, and so you know how to interact with it.

A robot or a software agent-- if it's busy, it's harried, it's looking over the net for some dumb question you asked it, don't ask it another one right now. So that's the subsurface model. And then there are emergent emotional models, where there are still internal drives and needs-- by the way, evolution put these internal drives and needs into us. We're putting them in our robots. And changes in these are reflected in changes in excitation or inhibition of behaviors of the robot.

And this is not aimed at showing an emotional response, it's aimed at getting the robot to satisfy the drives or needs, which are directing the robot to do what it was designed to do. So these behaviors that get excited or inhibited, they're all associated with the primary mission of the robot rather than as some explicit, external, emotional display. So there are no emotional variables, in the sense of having happiness as a 4-bit number or whatever.

There are no explicit, external displays, but the observer attributes the emotions to the systems. And these are deeper sorts of models. They are much harder to do. And no one's really, I don't think, done a successful one at that level. Here, in the lab and around the place, various people are working on versions of this. We have a pet robot. This is Cynthia Ferrell's child robot. It's got eyebrows, and ears, and stuff-- now it has lips, but I don't have a photo of that yet-- to give emotional responses.

This is over at my company, IS Robotics. It's a Robot called IT, which had a very surface level, emotional model. And I want to show you another emotional robot from IS Robotics, which is sort of a subsurface, almost emergent level, emotional model. This is a doll. The neat part about this is we think we can manufacture for very little money. But let me let me show you this one.

And this is an advertisement tape for the company, so there's a little bit of hype here, which you'll excuse, please.

[VIDEO PLAYBACK]

[MUSIC PLAYING]

- This is Bit, our interactive baby doll. Like [INAUDIBLE], Bit has invested intelligence which allows for interactive play. He understands bouncing, sucking his thumb, tickling, hugging, paddy cake, and much, much more. Like a real baby, Bit not only has his moods, but expresses them very well, using his face and voice. Bit understands the difference between play, soothing activity, and things he doesn't like, which all contribute to his mood state. The more you play, the happier he gets.

But if you keep playing for too long, he will get tired and cranky.

[BABY CRYING]

Keeping Bit happy is not always easy. But to solve all your problems, just give him his favorite thing in the world-- his bottle.

[BURPING]

[END PLAYBACK]

BROOKS: That's all been done on an 8-bit processor. It's not a lot of computation behind there, but it's engaging, at some level, of having this emotional model. So why have these emotional models? It gives humans feedback and, perhaps, other animals, it gives them feedback. Maybe the dog wants to know when the garbage robot is in a bad mood and should keep out of the way. It's intuitive, high level feedback on what the machine is trying to do currently, what its current set of goals are likely to be.

That's the surface level. But it also, inside, provides a mechanism for focusing the behavior of the machine in some, coherent way, when it's got lots of competing pressures it's trying to do. And this is not too dissimilar from Damasio's talk about from the frontal lobes, et cetera. Let me show you another videotape. The important message here is that the observer can really get sucked in fairly easily, I think.

This is Cog again. And this is the work of Robert [INAUDIBLE], who's somewhere in the audience. And what we've got-- just a second. One of the other things we're doing in Cog is trying to localize sound. So we see motions, we saccade the motions, we learn that sort of mapping. And as in the superior colliculus in the human, besides the visual motion map that's coming in, the ocular motor map, there's also a sound, oral map.

And we learned that coordination, between being able to saccade to where a sound is and from hearing the sound. When you put that on this robot, suddenly it starts to appear to be a lot more engaged. So it's saccading to where it hears sound. Now, hopefully, there's going to be-- okay. Now watch its eyes. You can watch the videos in the back. As they talk, it sits there, looking back and forth between the two people having a conversation.

And to a third observer, seeing that-- look, it's there, back and forth-- it seems to be engaged in understanding what's going on in that conversation, even though it's not understanding, in any deep sense. But we, as human observers, are willing to grant that licence that it is doing this higher level thing. Emotional content may often be in the eye of the observer. And the level of engagement may often be in the eye of the observer.

And the observer, as we saw with the earlier videotape-- where Cynthia Farrell was playing with the eraser with the robot, back and forth-- becomes a component of the dynamics of the behavior of the robot. And we, as humans, seem to be built to-- we sort of can't but help give the benefit of the doubt to systems that looked biological and to assume that they're engaged much more than they maybe are. So as our systems become more complex, the engagement with be longer term and the illusion will be shattered less often.

Let me give you a quote from Sherry Turkle's latest book, Life on the Screen. Sherry came over to our lab. And this is, actually, a couple of years ago. And when she walked in-- this is from her book-- she said, Cog noticed me soon after I entered its room. It's head turned to follow me and I was embarrassed to note that this made me happy. I found myself competing with another visitor for its attention.

At one point, I felt sure that Cog's eyes had caught my own. By the way, the colors here are mine, not hers, in the book. My visit left me shaken-- not by anything that Cog was able to accomplish-- because it was just a few, fairly simple, feedback loops-- but by my own reaction to him. For years, whenever she'd heard Rodney Brooks speak about his robotic creatures, I'd always been careful to mentally put quotation marks around the word. But now, with Cog, I'd found the quotation marks had disappeared.

Despite myself-- and this is important thing. Despite myself-- she's a skeptic-- and despite my continuing skepticism about this research project, I'd behave as though in the presence of another being. So, it seems to me, that as we build these systems, whether or not we put juice in them-- because that's just my personal belief that they need this juice-- they will appear, more and more to us, to have emotions. Maybe they've got this emergent emotional thing going on, maybe that's been designed into them, to force that to happen to us.

And we will get ourselves to the point where we want to attribute to the systems things such as fear. When viewed in its complete context, a robot can be just as afraid as a person. Well, we're all willing to say a person can be afraid. A chimpanzee-- can they be afraid? Yeah. A dog? They certainly seem to be afraid. Birds? Yeah. Lizards? Beetles? What about dust mites? Do dust mites get afraid? And they don't have the same stuff as us.

I think it gets down to the level of empathy that we're willing to give these systems. So it's going to take us another intellectual leap to get beyond our own fears of this lack of specialness in order that we can finally admit to such possibilities.

Some people will claim that even a chimpanzee can have no feelings, no fear, but I think the most common belief these days is somewhere around here-- that's where fear, and pain, and things go away. So once machines can be afraid, they will certainly have souls.

Once we're willing to admit they're afraid, then I think we'll have to be willing to admit they have souls, in just the same way people have souls. See, this is my little trick here, because I don't really believe people have souls. So in the same way you all believe that people have souls, you'll have to believe that about machines. But here's a question--

AUDIENCE: I think you were getting emotionally involved there.

BROOKS: I was. But I do worry. I do worry. What can machines legitimately be afraid of? Biological systems have their information content locked into their cells. Or they have, until recently. This may change for us, too. But in robots, that information content can be located offboard, at least the equivalent of the genetic material. Darwinian evolution had this urgency to make us be self-preserving. Because there wasn't a data bank over there which had our stuff.

If we got killed, we died, then our biological gene material died. So Darwinian evolution had to invent the need for fear and, incidentally, souls. But our robots are not necessarily based on Darwinian evolution. It's Von Neumann evolution. I made that up. But it seems to me that we can reproduce the robot from this master copy. It's a different sort of evolution than Darwinian evolution. So robots may not have to be afraid, in the long term. And they could be nasty.

If they don't have to be afraid, if they don't have to be self-preserving, they could be rather nasty, different aliens. But if we need to nurture them, if they're complex systems, and if we build robots that need nurturing to develop-- as a lot of people are starting to work on. Because if we think that we're the embodiment, that's the only way we're going to get to the right information content inside the head of the current robot, which will be very dependent on that robot, its physical embodiment, and the world it interacted in to get there.

We'll feel we have some investment in them. We'll instill them with fear, so they don't go off, and do silly things, and get hurt. And their culture that they develop will, irrationally, just as ours-- as I was talking about earlier, this dual belief system-- continuing to instill that fear in future generations. And finally, we will have built worthy successors to ourselves. And I'll take questions, I guess.

AUDIENCE: I'd like to bring up two examples in the history of physics that [INAUDIBLE] you would be replaced by a few of you. One was Descartes' cave cosmology. He thought the planets could be explained by interactive gears. And then Newton came along with his gravitational theorem, so it blew that out of the water.

And then you also have the ether hypothesis of electromagnetic waves, where people understood acoustics-- that had to have a medium. So electromagnetic had a medium. And, as a matter of fact, that was replaced by a field. And so I was just wondering if whether, maybe, your idea of juice might be equivalent to replacing a mechanistic viewpoint by something more--

BROOKS: Yeah. Maybe. Although I think David Chalmers fundamental type of consciousness is also a field-type replacement. But I just don't see any evidence for it. The only evidence I have for wanting this juice, which I believe is some mathematical construction, is that I don't see the current explanatory power of the gears being enough to do it. Yeah. So maybe.

AUDIENCE: This is very interesting. But the robots seemed to lack one thing-- and that's subjectivity-- to make it human.

BROOKS: What do you mean by that?

AUDIENCE: The thing that's going on inside of us right now.

BROOKS: So why can't the robots have that? Is it because it's special? Because only humans can have subjectivity? If we assume a mechanistic explanation, then we can put that mechanistical in there. We may be missing a few technical details, like the juice, but I don't see anything, in principle--

AUDIENCE: What's going on in the mind of the robot [INAUDIBLE].

BROOKS: That's the John Searle argument.

AUDIENCE: What?

BROOKS: That's a John Searle argument. That's the special stuff that makes people special. And since the robots don't have that, they cannot be people. It's that same, circular argument that John Searle makes. You want it to be special for humans, and you're defining it to be special, and including it special.

AUDIENCE: The most wonderful thing about myself is that I'm so [INAUDIBLE].

BROOKS: You're just a bunch of little molecules messing about. And it's only because I've got this other, dualistic notion that I don't just come and squash you. And I don't see why we can't feel the same way about our machines and have that dual nature. You want to have the specialness be a little box in there. It's the homunculus argument again.

I don't buy it. But I don't expect us to be able agree on that. I don't think John Searle and I can ever agree on that, because he's stuck in his set of dogmatic beliefs and I'm stuck in mine. Yeah?

AUDIENCE: I'm wondering if one of the subtle, implicit-- maybe explicit-- values of a computer is things such as efficiency, validation of what they're doing.

BROOKS: Do you use Microsoft software? It's too cheap. I can't.

AUDIENCE: And finding explanations. The mechanism is put together to find explanations, which, if so, would mean that--

BROOKS: I think when we're building such complex systems, we can't, any longer, look for the explanatory level that we used to be able to.

AUDIENCE: Then it would be that, if not defined, an answer would then mean that it's no longer needed. At such time, no further explanations are needed. But then a machine that just says, now that we have all explanations, our only goal is to self preserve. That is it. So either it is just self preserve or explain. There are no values that can be attached [INAUDIBLE].

BROOKS: Oh, yeah. We could we could build in that it should be nice to people and clean up their trash. We can build that in as a drive.

AUDIENCE: Oh, okay. Then you're saying that one of the things that you built into the [INAUDIBLE] thing is that the necessity of human beings-- which then would mean that they wouldn't be replaced.

BROOKS: Yeah. They might change things, after a while, but we could make ourselves feel good for a while by doing that. Up there.

AUDIENCE: Yeah. The thing is that the process of evolution-- for individuals in evolution-- is to reproduce, carry on your DNA.

BROOKS: I think that's an emergent property in the system.