They are already here, quietly at work in our homes, our offices and on our streets. Electrolux, the Swedish appliance manufacturer, makes one that vacuums your carpet while you sleep. Surgeons in Pennsylvania recently employed one to perform heart surgery on a cow. Some blind people in Japan use them instead of guide dogs to navigate city streets. What are they? A new breed of autonomous, intelligent robots that could soon be doing everything from exploring other planets to delivering the mail.

Robots, until recently confined to manual labor on factory assembly lines, are set to become more like people in the future: they will be able to learn, communicate with one another and even "think" for themselves. But this new range of robots won't look anything like the murderous machine depicted in the film The Terminator or the cute humanoids in Star Wars. In fact, they will look very much like, well, machines. "Robots are just computers with senses," says Kevin Warwick, professor of cybernetics at the University of Reading in the U.K. "There is no reason for them to look like human beings." Warwick knows better than most where science fiction parts from science fact. Together with his team at the University of Reading's cybernetics laboratory, he has designed devices that learn about the world in much the same way that human beings do: by trial and error.

The seven dwarves, a group of roller skate-like robots that have been under development at Reading for the past eight years, are scurrying about on the laboratory floor. One of them, Bashful, is acting a bit shy. When a student reaches down to touch it, it quickly withdraws like a frightened puppy. Another group seems to be playing follow-the-leader. While one dwarf leads the way, three other robots trail along in its wake like a string of ducklings drawn through a pond. What's so amazing about a bunch of machines rolling around on a laboratory floor? The seven dwarves have not been programmed to mimic the antics of puppies and ducklings; they've learned to do it by themselves. "Ordinary machines do what we tell them to," explains Warwick. "But intelligent robots can actually learn from experience and then decide for themselves what to do."

The seven dwarves owe their intelligence to the neural network with which they are endowed. A neural network is an electronic approximation of the human brain, in which computer chips replace the web of interconnected neurons inside our heads. Each of the seven dwarves has about 40 chips in its neural net. These are connected to six artificial eyes — ultrasonic sensors mounted across the front of the robot — through which the dwarves "see" their environment. The eyes send out a chirp of ultrasound that bounces off solid objects. The dwarves use the resulting ultrasonic echoes to get around in the world in much the same way that bats use sonar during flight.

But rather than being meticulously programmed to avoid every conceivable obstacle — a virtual impossibility — the seven dwarves are given a simple mission: keep moving and don't bump into anything. How they carry out this mission is up to them. The robots have no built-in behavior pattern. The dwarves playing follow-the-leader were never programmed to do this. They simply saw another robot successfully avoiding objects and discovered that if they followed it, they didn't hit anything either. By repeating this process, the robots learn that a game of follow-the-leader is one good way to fulfill their mission.

And once a dwarf learns a lesson, it can share it with the other dwarves. Each robot sports a ring of infrared transmitters that keep it in touch with all the others. A robot that has figured out how to navigate a particularly difficult terrain, for example, can broadcast this information to its fellows, essentially teaching them how to perform the task. As a result, the other robots don't need to repeat their tutor's mistakes. Using this method, a robot located at the University of Reading taught a repertoire of maneuvers to robots in New York and Tokyo via the Internet.

This kind of emergent behavior — the ability of intelligent robots to do things they aren't programmed to do — opens up a whole new range of potential applications. Similar technology is employed in the ASRAAM air-to-air missile manufactured by Matra BAe Dynamics, an Anglo-French firm that produces guided weapons. ASRAAM, which this year enters service with Britain's Royal Air Force, can be operated via the aircraft's radar or the pilot's helmet-mounted targeting system. But since the speed and precision of smart weapons far outstrips that of their human operators, ASRAAM can also act on its own initiative. Using infrared search and tracking capabilities, it can autonomously locate and destroy hostile targets that the pilot can't even see. Apart from digital dogfights like this, Warwick and his team envision more peaceful duties for their smart machines. Their prototype of a robotic fireman, for example, is equipped with visual, heat and smoke sensors and can roam independently through buildings detecting and extinguishing fires.

Watching the seven dwarves at work, it is tempting to imbue them with human emotions and motivations. After observing one determined dwarf repeatedly throw itself against the same wall, it's easy to regard it as "stubborn." Another hapless robot, lurching around randomly in the middle of the floor, seems thoroughly "confused." Warwick is quick to dispel such notions. "Machine intelligence is different from human intelligence," he says, "just as the way a plane flies is different from the way a bird flies." But he does see today's robotic surgeons and vacuum cleaners as only the beginning: their descendants, Warwick believes, will be faster, more accurate and smarter than we are.

"Machines can learn to do things we never thought of," he says. "In the future, they will be far more intelligent than we are." By then we may just be smart enough to use them wisely.