FAQ

 

Home
Up

--------------------------------------------------------------------------------

Why is Lucy being built?

As a test bed for some ideas about how the cerebral cortex of our brains might work, and more especially, how we might build future artificial intelligence. Lucy herself has no applications - she's really a toy (but a rather expensive and delicate one, so don't expect to find her being sold in toy shops just yet!). We just needed a complex organism - something with many biologically plausible sensors and actuators, interacting with a complex and messy world. Most neuroscience theories are tested on highly simplified "toy problems", which misses a number of perhaps critical points: for example, how intelligent would you be if you lived in an empty box and interacted with it via a bump sensor and a couple of wheels?!?

But why an orangutan?

We felt that a more application-oriented platform would always be pulling us in the direction of solving the application problem (making it work, no matter how much we had to cheat), when what we really want to solve is the technology problem - how to define a new artificial brain architecture that can ultimately be used for many different tasks. Lucy fits the bill - after all, what use is a baby? Yet look how useful and adaptable grown-up humans are.

Yes, but why an orangutan?

Oh. Well, it seemed like a good idea at the time...

I need to give you a bit of background here: Good Old-Fashioned Artificial Intelligence (GOFAI) set out explicitly to create human-like intelligence, but it got it all horribly wrong and wasted the best part of fifty years. As a reaction to this, a new kind of AI emerged; one that eschewed symbolic representations of thought processes and instead concentrated on more biologically founded ideas. This New AI (and the loosely associated field of A-Life) focuses heavily on understanding real nervous systems that are very much simpler than the human mind. But this backlash against symbolic AI leaves a huge gap in the market, as it were. Between the biologically tractable but not really very intelligent world of insects and sea slugs, and the far-too-hard, conscious, language-using world of humans, lies an interesting area that very few people are exploring. In this area lie most of the animals that we normally think about - dogs, cats, birds. They are less sophisticated than humans but there's something profoundly different (in my view) about their nervous systems, compared with those of ants and beetles. Such animals are highly intelligent, very adaptable and (many would say) conscious of themselves and their environment. They're still extremely complex compared to an ant, but at least they don't have to use tools, form complex societies and write research papers on artificial intelligence like humans do, so they're a more tractable problem. This type and level of intelligence is what interests us. Maybe we should be building a dog (but Sony got there first!). On the other hand, for various reasons, early language is a useful thing for us to study and gives us important insights. Dogs can't speak, but primates like chimps and gorillas have been taught simple forms of communication, and so building an artificial primate seemed like a good idea. We didn't want Lucy to be compared too closely with primate geniuses like Koko the gorilla or Panbanisha the bonobo, so we picked a slightly more distant relative - the orangutan. Plus orangs look cute. Plus when we went to Toys R Us to look for some soft toys that would give us anatomical inspiration, all they had was an orangutan...

Why's she called Lucy?

We actually named her after the famous Australopithecus skeleton. Although Lucy is ostensibly an orangutan she is, after all, a robot, and in many ways quite unlike any living primate. We called her Lucy in honour of our relatives from the past. The name also reminds us of Lucy the chimpanzee, who was raised as if she were a human baby by Richard Temerlin. Our Lucy will also be raised in a family, and our very long term objective is to get her successfully through kindergarten.

So what happened to the autonomous glider project? Too hard, huh?

Before Lucy was conceived, we were attempting to build an eagle - a two metre wingspan model glider, which we were planning to teach how to soar like a bird. This was an earlier instantiation of similar brain ideas to those used in Lucy. A glider seemed like a good choice - it has to learn to fend for itself in a complex and noisy environment, where its actions needed to be analogue, not digital (i.e. it had to think in terms of "left hand down a bit more", rather than discrete decisions like "turn left", which makes the problem much more realistic than many robots can handle). Unfortunately, we found we were spending too much time worrying about solving glider-specific problems like GPS positioning and weight saving, and not enough time working on the brain. Also, the British weather kept the glider grounded more often than not, and anyway, our thoughts on the brain were moving on and the simplified architecture we needed for the glider's brain was no longer appropriate. So we decided to change species. The glider is still hanging from the lab ceiling and the project will resume one day. After all, flying it is good fun!

What senses and actuators does Lucy have?

There are two Lucys, or rather the one Lucy has had two bodies. Lucy MkI was built on a shoestring budget and very quickly, in order to get a feel for the problems. This body is now pretty much worn out and unuseable. In 2003 NESTA provided some funding for Lucy MkII, which is still in development (and driving me mad). This will be a lot more sophisticated and significantly more like a biological organism.

Regarding Lucy MkI:

She has moving arms (each with four degrees of freedom) and head (two degrees in the neck, plus moving eyes, eyelids and jaw). She has no legs, because this first version uses standard servomotors for her muscles and these aren't strong enough to support her weight. Lucy 2 won't have legs either but will at least be mobile. Lucy's motors are configured as "virtual muscles." This means that they can behave almost like real pairs of antagonistic muscles - becoming limp, responding to changes in applied force, and detecting external forces such as touch and collisions with objects.

Besides muscles, Lucy has a voice, which allows her to make arbitrary calls and speech sounds (she has to learn to do this for herself - there is no built-in speech synthesiser).

Her senses include:

  • An eye (only one in Lucy MkI, because I couldn't afford the computer power for two). Her eye has low resolution peripheral vision and higher resolution foveal vision, just like ours, and contains a simplified model of the processing present in the cells of the retina.
  • Ears, with which to hear our voices and her own. The auditory parts of her brain can also detect the direction that a sound is coming from, and help her to isolate sounds from each other (auditory attention).
  • A sense of balance and head motion (accelerometers).
  • Senses for temperature, battery power, current drain and assorted other things.
Lucy MkII differs in the following ways:
  • Two eyes, with higher resolution, colour vision and a fine-grained control over their movement
  • More powerful hardware for voice and hearing
  • True muscle groups, unlike the simulated muscle pairs in Lucy 1. This means that she can genuinely let her limbs go limp (so that I can guide them as one would to teach a baby, for example). It also means her brain has a much harder but more realistic task leraning to control them. For example, feel your own neck and you'll find that you have four main tendons crossed over. By contracting them in various combinations you can nod and tlit your head and also rotate it. The relationship between the contractions and the movement aren't straightforward, but our brains learn how to do them, so Lucy's must as well.
  • Some mobility (probably in a wheelchair)
  • A lot more computer power
  • A generally more professional quality, now that I have some decent tools!

What computing power is inside her?

Lucy MkI has:
  1. Five small, custom-built, 16-bit microcontroller boards (Hitachi H8S) for controlling her sensors and actuators, and supporting inter-organ communication. the boards handle vision, hearing, muscles, voice and PC-comms respectively. All five boards are connected together as a parallel computer, with some shared memory for intercommunication. These computers don't do any of Lucy's "thinking" - they are mostly there to make her human-built technology behave in a biologically valid way.
  2. A couple of 8-bit microcontrollers (PIC) to help with muscles control and proprioception.
  3. A link to an external PC, which contained her brain during research and development.

Lucy MkII has any number of 8-bit microcontrollers (PIC16F876) handling stepper motors, muscle sensors, vision chips and so on. Her brain is a network of three main PCs, plus a fourth for monitoring her systems (acting like a brain scanner, essentially).

What will Lucy be able to do?

We really have no idea, yet. "As much as possible" is the short answer, but don't expect to be dazzled by her intellect! We hope to help Lucy's brain learn many of the things that young babies have to learn - how to move one arm independently of the other limbs; how to reach out and grab; how to recognise simple spoken words and the mood of the speaker's voice; how to copy those sounds; how to play (and enjoy) pat-a-cake...

The important thing is to test out our theories of how the imagination works and how we build mental models of the world inside our heads. At this stage most of the mental modelling involves body image (how Lucy's limbs are arranged in space; which bits belong to her and which are parts of the outside world; knowing whether a sound she hears is coming from someone else or is the sound of her own voice, and so on).

The limiting factors are the number of neurons we can fit inside her brain, the capacity of her senses to perceive the world and the limited power of her body to interact with it.

At the end of the Lucy MkI phase, she could learn to recognise simple objects (such as bananas and apples) regardless of which way up they were or how far away they were, and then she could point to them. The recognition of objects at any angle or location is a genuinely hard problem that has no satisfactory traditional solution. Work on Lucy has made some significant progress towards the baffling nature of visual invariance. But apart from that Lucy's actual abilities seem really unimpressive. This is because any idiot can program a robot to recognise bananas and point to them; it's not what she can do but the way she does it that matters. Conventional AI makes really impressive progress at first but quickly comes up against a brick wall. Lucy is about doing things the hard way in order to find a route around that wall.

What stage is the project at?

Lucy was conceived in May 2000, and Lucy MkII was begun in November 2002.

Lucy MkI reached the point where she could point at bananas, which confirmed the power of the approach and the potential of the theories behind her brain. However, she was physically so limited that I decided to replace her with Lucy MkII. This turned out to be a longer and more difficult job than I'd hoped! At the time of writing (May 2004) Lucy still has a long way to go, because I got stuck trying to find suitable ways to emulate human muscle. Our NESTA funding ran out before this problem could be solved, and at the moment the project is on hold while I focus on working for a living. This is no bad thing as far as you are concerned, because it will lead to products that you can use for yourself, instead of just reading about. It's a pity, though, and if there are any billionaires reading this I'm open to offers!

Even though Lucy's hardware is shelved for the moment, the theoretical work is still progressing nicely and I'm very excited about how things are going. Lucy MkI demonstrated that many of my ideas are on the right track and can be implemented in the real world, and since then I've had a good many new ones that I'm itching to try out. Some of these ideas will make it into real products eventually, but hopefully Lucy will still be the first beneficiary.

Meanwhile I've written a book about my adventures with Lucy MkI - click here for more information.

How long will it take?

Who knows? Our long-term plan is to help Lucy to work her way through nursery school. To do this we'll have to build several Lucys over a period of a decade or two. This is not a short-term project! Other people may seem to be making much faster progress, and if you read some people's books you'd be forgiven for thinking that superhuman artificial intelligence is just around the corner. But as I've already said, the classical approach always hits a brick wall very quickly, and for very good reasons. The Lucy project is about taking the long route, by trying to understand the clever tricks that animal brains use. Remember, even if Lucy had the brainpower of a newborn human being right now, and all of the theoretical and technical problems had already been solved, it would still take her eighteen years to learn to keep her bedroom tidy! Intelligence implies learning, and learning takes time.

Who's paying for all this?

We are. During 2003 NESTA, the National Endowment for Science, Technology and the Arts very kindly picked up the tab by giving me a Dream Time fellowship. But apart from that we are funding the work ourselves, supported by writing books and software. It's tough. Really tough. But it's the only way.

It's not that we haven't had offers of help, but they always come with strings attached. Commercial funding, no matter how principled and long term it tries to be, soon ends up with short-term goals. I know this from past experience. If we were under pressure to develop applications from Lucy this would soon compromise the integrity of the research and lead to dead-end pragmatism.

Meanwhile, established academic funding has its own problems. For a start we would have to have a clear-cut goal that could be achieved within a standard funding period of a year or two. It wouldn't be in our interests to make ambitious promises, and the short period would require us to set our sights really low. The result would be a watered down project that led nowhere. Frankly, I referee enough science funding applications to know that this is what happens to 99% of academic research - the system militates against risk-taking, long-term research. The accountants would rather see a guaranteed failure than risk an uncertain success.

The other problem with both academic and commercial funding is that it is predicated on collaboration. It's generally held that no single person could possibly do something this ambitious and complex alone. But history suggests otherwise - almost all major breakthroughs come from individuals, not teams. Teams are good for consolidating ideas that are already largely proven. Individuals may not be as competent as a team of experts, but the fact that all the thoughts reside in one person's head allows all sorts of cross-fertilisation and creative hunch-building that are simply not possible when the problem is distributed amongst several people with different mindsets, specialisms and motives. Working alone like this may well be a disaster, but it's doing no-one any harm and costing the taxpayer nothing, so it's worth a try!

How will her brain work?

The key to Lucy's brain is imagination. Many people still think of the brain as a passive receptor of information - as if raw data comes in from our eyes and other senses and gradually gets refined into more symbolic form (e.g. recognition of a face or a word) until finally all the streams come together somewhere and somehow give rise to an action in response. I think of perception as a very much more active process. As conscious beings we don't live in the real world - we live in a virtual world inside our heads. Most of the time this world is closely synchronised to the external world - our model matches reality, tracks it and predicts it. When we dream or when we imagine things (including making plans and rehearsing scenarios) we disconnect from the real world and let the model run on its own. The same mechanisms are at work in both cases, but there's no synchronisation with reality going on when we dream or think. The model is the crucial thing, and perception is an active process of using this model to predict, hypothesise about and correct for the data coming in from our senses - "filling in" when the data is incomplete and "being surprised" when reality fails to live up to the model. We are only ever conscious of the modelled world; not the real one. Being a child is not about growing a better brain, but about growing a better model - the only difference between a young child and an adult is that the child's mental world is less able to differentiate between fantasy and realism (not reality - none of us is aware of reality, only a realistic model of reality), and hence the child is still capable of believing in fairies.

Lucy's brain is designed around a key set of hunches about how such a mechanism can be made using (simulated) neurons and biochemicals, and how something similar might have evolved in nature.  More details are available here.

How does Lucy compare with other robots, like Cog?

Cog (at MIT's AI Lab) and Lucy are physically similar, in that they are upper torso, humanoid (primatoid?) robots with multiple sensory systems, virtual muscles and so forth. Cog is a good deal bigger and heavier, and is much more professionally engineered. Lucy meanwhile is taking much less time to build and is several orders of magnitude cheaper!

The main distinctions between Cog (including its baby sibling Kismet) and Lucy are ideological. In relation to the field of AI as a whole, Cog and Lucy are very much "on the same side", but there are important differences. Cog's intelligence is meant to arise from the interactions of a collection of "behavioural modules". Each module is designed explicitly and there's an implied sense that most of what people normally call intelligence can be explained (and perhaps therefore is explained) solely in terms of these modules - that intelligent behaviour arises from a hierarchy of simple unintelligent behaviours.

Lucy is based on a belief that, although this logic works well in "lower" animals such as insects, the higher vertebrates possess a rather different architecture. Evolving new neural circuits every time a new behaviour is required works only up to a point. Beyond that point, modules need to be much more generalised, and capable of learning whole new functions. Moreover, I think that there is no simple mapping between behaviours and the physical modularity of the system (many neuroscientists claim that modern brain scanners prove that the brain is made of discrete functional modules, but I think this is a misreading of the truth, based on a false paradigm). Undoubtedly the brain is modular to a degree - the hippocampus, reticular formation, etc. are all functionally distinct. But in the cortex at least, I think the behaviour of any given region is determined almost entirely by what it is (currently) connected to. So the Lucy project is a search for a single, general-purpose neural "module", which, when thousands or millions of them are placed side by side and connected together in suitable ways, can learn to exhibit a wide range of behaviours, none of which are pre-programmed or explicitly represented. It is a search for the basic building block of a mind.

There are many other robot projects, but none compare quite so closely with Lucy. Many are designed to test or develop theories of social interaction, physical coordination (such as walking or swimming) or other "piecemeal" concepts. Lucy and Cog (and Creatures) are pretty much unique in attempting to put together all the aspects of life and intelligence into a single functioning entity.

Where will it all lead?

My personal objective (apart from satisfying my own curiosity about the nature of life and mind) is not scientific but technological. I want to build machines with minds of their own because these machines will share many of the attributes of natural living systems:
  • They will be intelligent and adaptable, and can therefore be applied to many tasks or adapt to subtle variations in a task. They can also figure out how to solve tasks by themselves, without needing to be told.
  • They will be robust and will withstand changes and stresses in their environment far better than today's technology, which is beginning to creak under the strain of its own complexity. Moreover, when they go wrong, these new machines will do so gracefully (becoming "ill", rather than simply breaking down) and will suffer fewer cataclysmic failures.
  • They will be personable and user-friendly. Instead of us having to adapt to our machines, our machines will adapt to us, and we will be able to communicate with them in natural and intuitive ways.
We're still a very long way from robot butlers and android pilots, but simpler machines with imaginations and minds of their own will find many applications when embedded in equipment such as prostheses, spacecraft, traffic management systems, domestic appliances, vehicles and many other places.

People say that robots will take over the world. Is what you're doing safe?

People who say that are just scaremongers. There are several good reasons why this doom and gloom argument is wrong. You'll have to read my books to see my counterarguments in full. But suffice it to say that most of the "humans will become slaves" arguments are based on false logic, a misunderstanding of the nature of intelligence, misguided views about how rapidly artificial intelligence can develop (even in principle) and a strangely Victorian attitude to machinery.

Lucy will help us to see how to design machines that are rather less liable to cataclysmic failure than present technology, less irritating to work with and more able to avoid stupid mistakes. She will also (I hope) help us to understand ourselves better. She will not "explain away" all that we hold dear (our lives our minds and our souls) - in fact she will help us to exalt them as they deserve to be exalted. She will undoubtedly raise questions in people's minds and by doing so will help to destroy some of our more simplistic assumptions about morality and ethics. This is a good thing.

 
Copyright © 2004 Cyberlife Research Ltd.
Last modified: 06/04/04