Slashdot Mirror
Flesh and Machines: How Robots Will Change Us
In a way, robots are portraits of humans. Machines are just machines and assembly lines are just assembly lines. The buckets of bolts don't become robots until they start to take on some of the characteristics and a few of the jobs of humans. A drill for tightening a bolt may replace a biceps, but it's just a motor until it's on the end of a fancy mechanical arm that positions it automatically. Then it's a robot ready for a call from central casting.
Defining just what is and is not a robot is not an easy job for technologists because the replicants and androids are a touchstone and a benchmark for measuring our progress toward the future. It's 2002 and everyone is asking: Where's mad Hal steering a space craft to oblivion? Or more importantly: Why am I still vacuuming the floors and mowing the lawn by myself?
If you are asking these questions, then you might want to read the answers Rod Brooks, the director of MIT's Artificial Intelligence Laboratory, offers in his charming book, Flesh and Machines: How Robots will Change Us. The book is half a thoughtful biography of the various robots created by his graduate students and half a philosophical explanation of what to expect from the gradual emergence of robot butlers.
The biographical part is probably the most enjoyable. He and his students have produced more than a dozen memorable robots who've crawled, rolled and paced their way around MIT. One searched for Coke cans to recycle, one tried to give tours to visitors, and another just tried to hold a conversation. Brooks spends time outlining how and why each machine can into being. The successes and more importantly the failures become the basis for creating a new benchmark for what machines can and can't do.
An ideal version of this book should include a DVD or a video cassette with pictures of the robots in action because the movement is surprisingly lifelike. Brooks is something of a celebrity because a film maker named Errol Morris made a droll, deadpan documentary that cut between four eccentric geniuses talking about their work. One guy sculpted topiary, one tamed lions, one studied naked mole rats, and the fourth was Rod Brooks, the man who made robots. Brooks minted the title for the film, Fast, Cheap and Out of Control, a phrase he uses to describe his philosophy for creating robots. The movie tried to suss out the essence of genius, but it makes a perfect counterpoint for the book by providing some visual evidence of Brooks' success.
One of the stars of the movie was a six-legged robot called Genghis, a collection of high-torque RC airplane servo motors that Brooks feels is the best or most fully-realized embodiment of this fast and cheap approach. The robot marches along with a surprisingly life-like gait chasing after the right kind of radiation to tickle the IR and pyro-electric sensors mounted on whiskers. If you've seen the film, it's hard to forget his gait.
Brooks says that the secret to the success of Genghis is that there is no secret. The book's appendix provides an essential exploration of the design, which is short and very simple. The soul of the machine has 57 neuron-like subroutines, or "augmented finite state machines" in academic speak. For instance, one of the AFSMs responsible for balance constantly checks the force on a motor. If it is less than 7, the AFSM does nothing and if it is greater than 11, the AFSM reduces the force by three. That's doesn't seem like very much intelligence be it artificial or real, but 57 neuron-like subroutines like this are all it takes to create a fairly good imitation of a cockroach.
Brooks calls this a "subsumption architecture" and the book is most successful describing the days that he spent with his graduate students building robots and seeing what the architecture and a handful of AFSMs could do. He half mocks the roboticists who load up their machines with big computers trying to compute complex models of the world and all that is in it. In his eyes, the lumbering old-school machines just move a few inches and then devote a gazillion cycles to creating a detailed, digital description of every plant, brick or wayward child in the field of view. After a few more gazillion cycles, the machine chooses a path and moves a few more inches. Even when they find their way, time passes them by.
There are no complex control mechanisms sucking down cycles on the machines from Brooks' lab, the source of the claim that they're "out of control". It's just AFSMs wired together. One of the robots fakes human interaction by tracking fast motion and flesh colored pixels. Brooks marvels at how a few simple rules can produce a machine that is remarkably life-like. If you're not sure, they have video tapes of lab visitors holding conversations with the machine, who apparently takes part in the conversation with the patient interest of a well-bred host. As if by magic, the AFSMs are creating enough human-like movement and visitor in the tape begins treating the robot like a human!
If you're still not sure, you might buy a "My Real Baby" doll designed by Brooks with the help of the adept mechanical geniuses in Taiwan. The story of taking a highbrow concept from MIT to the local toy store is a great part of the book. The so-called toy is filled with AFSMs that tell it when to gurgle, when to pout, when to sleep, and when to demand sustenance. Alas, the toy makers tell Brooks that the market can't stomach so much innovation. One new thing at a time.
So are these machines truly successful simulacra? Are they infused with enough of the human condition to qualify as the science-fiction-grade robots or are they just cute parlor tricks? Some readers will probably point to the AFSMs and scoff. Seeing the code is like learning the secret to a magic trick.
Brooks, on the other hand, is sure that these machines are on the right track. In a sense, he makes it easier for his robots to catch up with humans by lowering the bar. On the back of the book, Brooks ladles out the schmaltz and proclaims, "We are machines, as are our spouses, our children and our dogs... I believe myself and my children all to be mere machines." That is, we're all just a slightly more involved collection of simple neurons that don't do much more than the balance mechanism of Genghis. You may think that you're deeply in love with the City of Florence, the ideal of democratic discourse, that raven-haired beauty three rows up, puppy dogs, or rainy nights cuddled under warm blankets, but according to the Brooks paradigm, you're just a bunch of AFSMs passing numbers back and forth.
If you think this extreme position means he's a few AFSMs short of a robot professor though, don't worry. Brooks backs away from this characterization when he takes on some of the bigger questions of what it means to be a human and what it means to be a machine. The latter part of the book focuses on what we can and can't do with artificial intelligence. He is very much a realist with the ability to admit what is working and what is failing. His machines definitely capture a spark, he notes, but they also fall short.
He notes with some chagrin that his robot lawnmower leaves behind tufts of uncut grass. Why? It uses a subsumption-like algorithm that doesn't bother creating a model of the yard. The robot just bounces around until the battery runs out. Eventually the laws of random chance mean that every blade should be snipped, but the batteries aren't strong enough to reach that point at infinity. A model might help prevent random lapses, but that still won't solve the problem. Alas, the machines themselves are limited by the lack of precision. One degree of error quickly turns into several feet by the other end of the yard. A robot wouldn't be able to follow a plan, even if it could compute one.
What's missing, Brooks decides, is some secret sauce he calls "the juice". Computation and AFSMs may work with cockroaches, but we need something more to get to the next level. Faster computers can do much more, but eventually we see through the mechanism. Genghis looks cool, but learning about the 57 AFSMs spoils the trick.
The standard criticism of Brooks' machines is that they don't scale. There is no superglue juice that can save a scaffolding built of toothpicks. The AFSM may produce good cockroaches, but that's just the beginning of the game. Humans are more than that. Eventually, the AFSMs become too unwieldy to be a stable programming paradigm. In fact, Brooks sort of agrees with this premise when he suggests that Genghis is his "most satisfying robot." It was also one of the first. The later models with more AFSMs just don't rank.
But humans and other living creatures don't scale either. We may be able to run 20 miles per hour, but only for 100 yards. We may be able to troll for flames on five bulletin boards, but eventually we get our pseudonyms confused. Limits are part of life and we only survive by forgiving them. To some extent, the lifelike qualities of his robots are direct results of the self-imposed limits of the AFSMs.
Your reaction to these machines will largely depend upon how many of the limits you are willing to forgive. Stern taskmasters may never be happy with a so-called robot, but a relaxed fellow traveller may ignore enough of the glitches to interface successfully. Some will see enough of themselves to be happy with the whirring gizmos as a portrait of human and others may never find what they're looking for. That's just the nature of portraits. For me, this book is an excellent portrait of a research program and the collection of questions it tried to answer. You may look in the mirror and want something different, but it's worth taking a look at these machines.
Peter Wayner is the author of two books appearing this spring: the second edition of Disappearing Cryptography , a book about steganography, and Translucent Databases , a book about adding extra security to databases. You can purchase Flesh and Machines from Barnes & Noble. Want to see your own review here? Just read the book review guidelines, then use Slashdot's handy submission form.
Biological neurons have been shown in the laboratory to grow new connections based on information learned. In a robot, what possible mechanism could guide such growth? Programming is the only answer, but keep in mind that "programming" is just shorthand for "the intelligence of the programmer". In other words, the AI itself isn't self-contained, as it were.
There is no other way for "mental" activity to be guided, thus AI will always be as unattainable as the Philosopher's Stone.
Combine everything thats going on into the soup of the future: robotics, quantum technology, biotechnology, high speed wireless internet, satellite communications...
I believe the robots are going to be us, except for advanced machinery in manufacturing the "happening" thing will be integration and interfacing of electronics and biocircuitry with ourselves. You will think and your interface will retrieve data from storage attached to you.
Electronics can monitor your bloodstream for diseases, lack of resources, and the like, and synthesize whatever is required. Good for anyone with a genetic defect or an illness. Good for your general health & wellbeing.
The advantages are so enormous these technologies will be used in that manner. You will probably want to have it. But you'll also realize that at that moment you are not only vulnerable to hackers that try to access your biosystems, also those that create the hardware and software within you are potentially able to upgrade software and firmware that has essentially become a part of your being.
So who will controls that, us, intimately? Open Source at least insures that we will have insight into if not control over who we are developing into...
Of those to whom much is given, much is required.
Intelligence:
The capacity to acquire and apply knowledge.
The faculty of thought and reason.
According to the above AFSM's are the exact principle behind intelligence. Think about how any analysis of the world happens. We don't consider the entire world when we try to catch a ball, we consider the position of the ball and where it will be. We don't take the position of a bird in relation to the ball, or something far away, all that matters is the ball.
Slightly more complex would be hit detection, is there anything close to me? Yes or no...that easy, you'd have a range that it's ok for an object to be in, a range where we should slow down and a range where we fire thrusters to stop.
Simple actions put together equal complex life form.
internet like monkeys'
I listened to Brooks present the semi-academic version of his talk at Duke. The really fascinating thing about this robot/experiment is that making the robot react to simple cues from the human makes the robot act much more intelligent than it actually is. It may be easier to make a robot that behaves intelligently around humans than it is to make one that intelligently explores mars.
By giving the robot the ability to recognize eyes and where the human is looking, it can pick up cues as to what aspects of the environment are important. By making it maintain a proper conversational distance from the human, it prevents collisions and makes talking to it much more comfortable.
Because the robot responds to its environment, the environment shapes the robot's behavior. If that enviroment is alive and intelligent, the robot's behavior becomes more intelligent than it would normally be. We give off hundreds of little cues that allow us to respond intelligently to each other, and Brooks' work has opened the door to letting robots bootstrap themselves to a higher level of interaction.
Penrose's lawn mower robot doesn't mow his lawn properly because he forgot to design it to WANT to mow his lawn properly.
Seriously! To properly want something, you need a means to know that that desire is or is not satisfied, and a means to move closer to achieving your desire - just like Genghis' leg muscles.
His mower robot needs a laser scanner to light up stalks that stick up too high, a sensor to detect stalks being lit up within maybe 10 feet, a desire to go to spots where that light is seen, and a desire to wander and seek out lit spots if it doesn't see any nearby.
A bit more is needed to handle edge conditions (literally the edges of the lawn and objects in it). It needs the ability to learn where it can't go, and the ability to slowly forget that learning so if it makes a mistake about not being able to get somewhere it can eventually correct itself.
Take an assembly-line robot, for example. It so happens that a human configuration for an arm (A fairly mobile shoulder, a somewhat limited elbow, a fully-functional wrist, and some sort of manipulator at the end) is very useful. With a system like that you can reach any part of a design. Could you add another joint and achieve more flexibility? Or perhaps give the elbow more degrees of freedom? Naturally, and people have in fact done these things. However, there are a number of good reasons to mimic human design.
First of all, we are innately familiar with the operation of an arm. We have no trouble visualizing just how an arm like our own would move around something - For those who are good with math, this can translate into an easy understanding of the math involved.
Second, lots more work has gone into human-similar models. This means you can draw upon the accumulated design experience of hundreds and thousands of other people even inside the field of robotics.
Finally: Adding more joints/making more capable joints costs more money. In most systems which need to be versatile, the human-mimic system is the most efficient from a cost:capability standpoint.
Robots are like humans where they need to be. When we can make them identical to humans, no doubt some will, while others will feel that that is some sort of travesty. We all know that the big application in robots is the self-mobile realdoll, though, and that's an attempt to make something as much like a person as possible.
You might as well argue that giving birth is creating a portrait, since there is such variation in humanity - And there is still MORE variation between robots.
"You're right," Fisheye says. "I should have set it on 'whip' or 'chop.'"
Not so fast.
There are some cockroaches, you step on the, and all they do is get mad. You have to splat them with a hammer. Of course, you could always get some as pets. Nevermind the ones in Florida that fly imported from Asia.
"It is a greater offense to steal men's labor, than their clothes"
I studied robotics and Brooks' work for a couple of years in graduate school. I built several robots using some of his ideas with some pretty spectacular results (I was impressed anyway) considering that they were able to navigate around and perform some very simple tasks using less code than your average mouse driver. Brooks turned the whole notion of robotic intelligence upside down and started from the bottom up, keeping things simple.
/. and get back to work. Let me know when you make the breakthrough, I'll buy you a six pack.
Its pretty striking to me how different an engineer's life can be depending on his area of interest. There are some topics where we are essentially on the "right track". Some genius has made the initial breakthrough in thinking. Steady progress can be made by moderately intelligent people such as myself by following the premise to its logical conclusions. While I was studying robotics, the Web was really taking off. Ideas spread like wild fire and advances are still being made fairly rapidly.
Other areas of study stagnate for years with random dispersed periods of growth and euphoria followed by periods of disappointment and disillusionment. In AI/machine intelligence, we have had several small breakthroughs that allow us to progress a little before hitting the brick wall again. We're all waiting for someone to make the leap in thought that will allow us to progress.
My opinion now is that we have some fairly specialized approaches that work well in specific circumstances but we are all essentially still on the wrong track.
Rodney Brooks caused quite a bit of excitement in the early '90's with Ghengis and some of his other robots but it wasn't that breakthrough that
we are all waiting for.
From what I understand, if you have read his papers and publications through the years then this book doesn't offer much new information. If you aren't familiar with his work and are interested in the subject then definitely read the book. Even if Brooks doesn't turn out to be the genius who makes the breakthrough, his work has definitely contributed to the field and brings us a little closer.
In the mean time I guess I'll just have to wait for the big breakthrough by building some more little robots to keep me busy. I've been thinking about a little robot with a single board Linux computer for a controller and a WIFI adapter. That way I can sit at my desk or laptop and watch what is going on a tune code and develop behaviors from the comfort of my couch instead of having to track the little bugger down and stick a serial cable in its ass to upload new programs and download data. I was also thinking that I could then give real time performance feedback and let some genetic algorithms and/or neural networks tune the parameters. That should keep me preoccupied for a while while the geniuses work on the really heady stuff.
If you are one of those geniuses, quit screwing around reading
"Why do you close your eyes?", Sussman asked his teacher.
"So that the room will be empty."
At that moment, Sussman was enlightened.
I may seem a bit foolish here for asking, but what does this mean? I don't understand. Is it that Sussman learned to start with all 0s instead of random inputs? Or that cutting out all preconceptions is only counterproductive?
Remember "Bring 'em on"? *sigh