Made to Order
The Dumbing Down of the Robot Mind
When most people think of robots, they envision walking, talking mechanical humanoids along the lines of The Jetsons' domestic Rosey the Robot, Futurama's grouchy Bender, or Lost in Space's B-9, frantically waving its mechanical arms at impending danger.
But to judge by robot pioneer Rodney Allen Brooks' new book Flesh and Machines, as well as several recently published robot-building how-to's that follow in the path Brooks blazed, robot science is scooting off in a less anthropomorphic direction. Rather than general-purpose sentient machines walking about, offering advice, cracking wise, and helping with the chores, robots have mutated into specialized creatures, dumb machines that undertake complex, useful tasks such as cutting lawns, crushing fellow bots for entertainment, and serving as mechanical pets. And like most people who hire domestic help, their owners don't care how smart they are.
Brooks runs the artificial-intelligence laboratory at the Massachusetts Institute of Technology. Flesh and Machines is a fascinating account of how he got involved in the field and where he thinks it will head. Brooks' speculations have the weight of years of garage-shop-level practical experience behind them. He and his lab have built a number of robots, including the wheeled prototypes for NASA's Sojourner vehicle, which tooled around Mars.
Flesh and Machines revolves around an epiphany that Brooks had early in his career. When he first got involved academically in the robotics field the prevailing wisdom was that a well-designed robot required some sort of brain, a central processor to regulate the mechanical tasks it would undertake, and that this brain would need to build and hold an internal map of everything it could sense about its surroundings. The problem was that rendering a computerized representation of every obstacle in a robot's path, and then executing the orders to the limbs to avoid those obstacles, takes an immense amount of computational power.
Brooks takes delight in describing a slow-moving robot that he helped develop while a research assistant at Stanford University in the late '70s. Tethered to a mainframe computer by a large cord as it tried to make its way around the lab, "the Cart made its forward lurches about once every 15 minutes," he writes. "Once every 15 minutes if [the device's] main computer wasn't busy running jobs for other people."
This struck Brooks as absurd. Why, he asked himself, could insects with roughly the same amount of computational power (in the form of slower interconnected neurons) move about the world at speedy clips? The conclusion he reached is that insects don't employ an internal representation of the world. The world itself is their model; the bugs just bounce off of it using simple reflexes. Instead of designing robots with bulky command-and-control noggins, in the mid-'80s Brooks began building creatures with numerous simple control systems.
When enough of these simple, largely independent circuits were overlaid, the sum total of their combined actions produced remarkably sentientlike effects. One of Brooks' robots was designed to walk visitors about the lab, giving them a guided tour of the facilities. Another, named Kismet, was a disembodied metal skeletal face with two eyes and ears that mimicked the physical cues of a person participating in a conversation, down to the raising of the eyebrows at the appropriate times. With 15 independent computers controlling Kismet's reactions, people would find themselves unwittingly chatting with this thing, even though it was only babbling well-timed nonsense syllables.
Although this decentralized approach may seem obvious now, at the time the academic community greeted Brooks' ideas with shock. "Without quite being explicit," he writes, "I was saying that all previous work in artificial intelligence was misguide."
Brooks' ideas have taken hold. The growing presence of simple novelty robots since the mid-'90s--Sony's mechanical pet Aibo, the fighting robots of TV's BattleBots and Robot Wars--are his legacy, in that they rely more on crafty mechanical engineering than computational power to do what they do.
Brooks' simplified approach serves as the basis for a number of recently released guides on how to build robots. In fact, David Cook's Robot Building for Beginners could be subtitled Robot Building for Liberal Arts Majors Who Flunked Shop Class in High School. It really does assume its readers have no mechanical or electrical experience.
Wisely, Cook devotes the entire 568-page volume to building one robot, a simple four-wheeled device with a small plastic food container for a body that can follow light and dark colored lines. For intelligence, the only microchip Cook's robot uses is a simple preprogrammed "comparitor" that can be purchased at Radio Shack. The chip operates by comparing the voltage of two inputs and sending signals to fire up the wheel motors on either side of the bot based on what it gets from those inputs. In other words, Cook's robot is a mental ninny, one evolutionary step up from a radio-controlled toy car only by the virtue that it doesn't take orders from a remote unit.
Cook doesn't get into more complex robot technologies. Once you start thinking in a robot-building way, he calculates, you can design more complex stuff pretty easily. Besides, he offers examples of other quasi-useful bots that can be built from similar simple designs, such as a slug squisher, a window washer, a houseplant-watering robot, and so on.
Pete Miles and Tom Carroll's Build Your Own Combat Robot doesn't dwell too much with internal logic either. This book was written for those interested in participating in those gladiatorial robot-war tournaments on TV. Unsurprisingly, combat bots don't require much higher-level thinking to find opponents and pulverize them. "Of course, if you're building a BattleBots-style [radio control] machine," the authors note, "you probably won't need any software, and the 'sensors' are your own eyes as you guide it across the floor of the battle arena."
Combat robots, however, are not simple creatures. In fact, Combat Robot presents designs for mechanical marvels that blow away anything in Cook's Robot Building. Unlike Cook's book, however, Combat assumes you already possess considerable mechanical skills and are only seeking wisdom on how to harden your bot against thrashings.
The debt of these how-to's to Brooks' ideas is clear. At one point the authors of Combat Robot even discuss swivel-mounted wheels, referencing the success of early Sojourner prototypes that Brooks' lab worked on. Both books harp on the theme that Brooks devised--off-load as much intelligence as possible in favor of clever mechanical design.
Naturally, neither of these how-to's has any advice regarding the construction of something resembling consciousness--the trick the original AI researchers were after with their brainy bots. Nor do they even venture the idea that it can be done, given enough circuits. Brooks admits near the end of his book that his ideas of decentralization probably won't engender artificial intelligence, arguing that the secret ingredient to producing consciousness is something scientists just haven't discovered yet.
But maybe this is as it should be, given people's attitudes toward bots. The late science-fiction humorist Douglas Adams understood the inherent tension in what is desired from robots--namely to be smart enough to do their owners' bidding, yet not smart enough to think on their own. One of the mechanical characters in Adams' novel The Hitchhiker's Guide to the Galaxy was Marvin, a robot that, because it was programmed with humanlike traits, was relentlessly depressed by the futility of its own existence. "Here I am, brain the size of a planet and they ask me to take you down to the bridge. Call that 'job satisfaction'? 'Cos I don't,'" Marvin would kvetch. It's as if Adams was warning against making robots too smart.