Stork: So could you describe Kismet in a little more detail?
Brooks: Kismet is a robot with a head and a face. It can make facial expressions that people can understand. It doesn't have a body. Its got a good neck so that it can do some of the things that we use our shoulders for, so it can be a bit more expressive than without that. And Kismet has eyes with models in front of them like a human eye so that a person looking at Kismet can tell where Kismet is looking and Kismet can steer its eyes around using high resolution cameras and at the same time the person understands what Kismet is paying attention to which is a great and important cue in understanding how to interact with the person or with a robot.

Stork: So you're training not only Kismet, but you're training human conversation partners to get used to dealing with a machine this way?
Brooks: Any interaction between a robot and a person needs to be an interaction. The person controls the robot in some respect but the robot may need to control the person. If the person is moving around too fast the robot can't keep track of them, of where their eyes are looking, so the robot in Kismet's case gives the same sort of cues a person would give, looking unhappy, looking confused, if someone is moving around too quickly or too close to their face, and that controls the person into a realm where Kismet can understand what is going on. So it is a two way street. Robot controlling person, person controlling robot, but we make it so we don't have to teach the person anything ahead of time, its just the natural cues that they are used to using when they talk to another person.

Stork: So why is it important for robots to have emotions?
Brooks: Ultimately it may not be important for all our machines to have emotions. I expect my refrigerator to work 24 7 and if I start feeling sorry for it working that hard it isn't going to be a good refrigerator. The ATM machine at the bank, I don't want to have an emotional relationship with that, but some machines I do want to have some sort of emotional relationship. I want it to give me the sort of feedback which lets me know how it's understanding me and is a little more personal in its interaction with me. So we're exploring the extreme. We're exploring a robot which is based on emotions. There'll be many other robots in the world in the future which don't have much emotion or don't have any emotion. But I think there is gong to be a place for some robots with quite a bit of emotion and we happen to be pushing in that direction to find out what is possible.

Stork: Do you think the emotion itself is crucial towards cognition?
Brooks: If we look at ourselves, other mammals, all their intelligence whatever that means for some mammals, is built on top of a more primitive emotional system. So every instance of intelligence is based on top of an emotional system. It may not be entirely necessary, but it may be that any system that doesn't have that emotional system may always feel like an alien to us. We may never really deeply understand it. So perhaps for a system that we can understand in a deep way, perhaps that emotion is important.

Stork: So that leads us to how you would define intelligence itself?
Brooks: Defining what intelligence is is a very difficult question. In the early days of AI, The early practitioners, the first researchers took intelligence to mean the stuff that they had trouble doing. So that meant proving theorems, playing chess. We now understand that there is a lot more to intelligence than that. The early researchers thought that vision must be pretty easy because anyone can do vision. A kid can do vision. Animals can see things. That turns out to be incredibly hard and is still an enormous area of research. So what exactly is intelligence has changed over time. And partially we get into a trap of whenever we understand how to do something, 'well that's no longer intelligence,' that's not part of the equation anymore.

Stork: Suppose we build AI to everyone's satisfaction, sometime in the very distant future. What would that mean for us in understanding ourselves and us as humans?
Brooks: As we are trying to build these intelligent systems, especially the ones we model after the human system, it helps us sharpen our understanding of how the human system works, because we can't just circumscribe a piece of the intelligence and talk about that, we have to connect it to the other pieces so it extends the theories so we end up with more and more of an understanding of ourselves. If we are able to build systems with the intelligence level of HAL say, ultimately we get to the point of saying well what rights does HAL have? What makes something a living creature, we certainly give rights to our pets, its illegal to do all sorts of things to a pet that it is perfectly legal to do to a chair. At what point do our created intelligences start to get some of those rights?

Stork: And your suggestion would be?
Brooks: I think this is going to be a real controversial issue over time. We've seen such controversy over history with the tribalism that we've engaged in. When do we give rights to that other tribe of people who are different from us? What rights do we give and what respect do we have for animals, for different sorts of animals. Some people may forever feel that a machine is a machine is a machine. But I think over time we will come to empathize with them more and more until gradually we will change our views of the emotional, intelligent machines. We will still have other machines, like chairs where it is ok not to switch them off at night. You're not being cruel to them. But to some machines that would be extreme cruelty.

Stork: So what do you think some of our biggest problems are in a path towards a HAL?
Brooks: We've made a lot of progress in the past few years in perceptual systems. Now part of that has come from Moore's Law finally giving us enough computing power and individual cheap computers that we can start to do some aspects of real time vision. We can recognize faces, track moving objects, but there are some other parts of vision that we just can't do well at all. Vision systems are not good at recognizing objects and understanding what objects are for. So there are a lot of things that are very easy for us which are still very difficult for our perceptual systems. So I think we've got a big gap still in computer perception. There is a lot more to be done. And I don't think it's just a matter of computer power. I think we need some more ideas. We don't quite know how to do it yet. Since I think that intelligence through evolution was built on top of the ability to do perception and simple motor actions, I think once we get those sorts of perceptual abilities the higher level aspects of intelligence will follow on relatively easy, but not without more brilliant minds thinking about the problem. So I think that perception is still the critical issue for intelligence.

Stork: You suggested that we need more smart people, smart ideas, but what about more data as necessary for training these systems. Is that crucial or just a by-product?
Brooks: I think we are in a very interesting place right now. We have vast amounts of data online. We could, can, do build disembodied agents that can gobble up more gigabits of data than the whole history of humanity has had to deal with. Also because we are starting to have real time vision systems we can put our robots in an environment with people where they interact with people over days, weeks, and months and get a lot of data from which they can learn from. So I think we are not probably at the mercy of not getting enough data to train our systems, we are at the mercy of not knowing how to learn from that data.

Stork: What do you think of HAL?
Brooks: HAL of course turns out to be a psychopath but he's still the dearest intelligence to me. There was a small flaw in his character. He was an intelligent, artificial entity which to me was just so inspiring.

MOC: Could you describe Cog?
Brooks: Cog was our first humanoid robot that we built, although I think its undergone three head transplants and two arm transplants since we started working with Cog. And Cog is a waist-up robot with torso and arms able to interact with the world with its arms. When we started building Cog we didn't quite realize the importance of social interaction so Cog is not as expressive as some of our other robots. This was something we learned form our early interactions with Cog, how much people tended to treat Cog as a social being, which we were not thinking of when we started the research.

MOC: Why did you shift from building insect like robots to humanoid robots?
Brooks: When I first started building insect robots, my ultimate goal was to get to HAL, to get to an intelligent being. I thought I'd work on insects, then maybe I'd work on simple reptiles, a simple mammal, then maybe a cat, and maybe primates. But I spent ten years on insects, and then when I did the arithmetic I realized I'd be long dead before I got to the really important stuff. So I decided to go for it and make a jump straight to humanoids and see what I could do there, rather than being remembered as the guy who built the best artificial cat.