BOSTON (05/23/2000) - John Dolan says he wants to build a robot that can be flushed down the toilet - literally - to inspect sewer pipes. Tucker Balch proposes building a disposable robot airplane that could be dispatched by a forest ranger to check out a suspicious column of smoke. And Branislav Jaramaz envisions a RoboChef that can move around the kitchen and cook a meal. But he worries, "Who is going to build the robot to clean the mess behind this one?"
Dolan, Balch and Jaramaz were among 40 robotics researchers from Carnegie Mellon University's (CMU) Robotics Institute who met recently at a faculty retreat near Pittsburgh to brainstorm about the future of their craft. Each was asked to outline a vision for a useful robot that could be built within the next five years.
But past prophecies about robots have been notoriously off the mark. Experts in the 1950s predicted that robots would be vacuuming our floors, carrying out the garbage and doing other mundane chores in just a few years.
A half-century later, we're still waiting.
Experts in the 1950s and 1960s assumed that if computers could play chess and prove mathematical theorems - and they could then - it should be fairly easy to get them to help around the house.
"The impression was that computers were already at least as powerful as the human mind," says Hans Moravec, a CMU researcher who has spent 40 years building robots. "But it's much, much harder for a computer to vacuum the rug than to prove theorems. It's only when humans do it that it looks the other way around."
Moravec says the brain in an advanced industrial robot today works at about 10 MIPS - no smarter than an insect. But he says faster processors will allow robots to evolve to reptilian, then mammalian and finally to human intelligence over the next 40 years, when they will compute at 100 trillion operations per second. "By that time, the world economy will be fully automated, and human beings will be retired," he says.
Moravec says he plans to produce by 2003 a 1,000-MIPS robot that will be able to navigate by 3-D images 1,000 times richer than the 2-D images that today's robots use with limited success.
Shortly thereafter, Moravec says, he will produce a commercial product called a "navigational head." The size of a basketball, it will contain stereoscopic cameras, 3-D mapping and image-recognition software and a layer of application software. The heads will be retrofitted onto existing industrial vehicles such as robotic cleaning machines.
Today, those machines must be laboriously trained and calibrated at great expense. But with the big new brains grafted on, they will learn new routes after being led through them just once, Moravec says. At that point, the specialty market for robots will explode, he says.
Meanwhile, the Web could solve another fundamental problem with today's robots.
Although they can extract limited information from their environments via sensors, robots are constrained by the information built into them.
"[A robot] doesn't have any way to go out and learn the information to add to its knowledge," says David Bourne, a principal scientist at the Robotics Institute.
But now there's a huge store of knowledge from which robots can learn, Bourne says. Many of the most important strides in robotics in the next few years will involve teaching robots to access the Web and to interpret and act on what they find there, he says.
But interpreting information extracted from external sources such as the Web is difficult for robots and is likely to remain so for a long time. Bourne says the best robots in the future will be those smart enough to ask for help.
One of CMU's mobile robots does just that, he says. "It goes to the elevator, and if it senses people standing there, it says, Would you push the Up button for me?' because it doesn't have an arm."