IN MARIA GINI'S computer-science lab there are computers, of course. But the more useful laboratory instruments are her LEGOs. She buys them in the standard colors and sizes from toy stores and she keeps them in open trays, from which they spill out onto the floor of her lab. Jumbled in with them are wires, little electric motors, and rubber wheels, all purchased for pocket change at Radio Shack.
Next to the trays sits a large LEGO gadget--a big block of jaunty yellow, red, and blue plastic bricks balanced precariously on four tires; a sort of truck dragging a trailer behind it. A few wires stick out of the truck, and a phone jack. On the back of the trailer, secured by electric tape, is a pointy-looking set of horns. The whole contraption is like something a very ambitious 11-year-old might construct with LEGOs and a broken telephone on a rainy Saturday afternoon.
But really, this rattletrap is a state-of-the-art scientific breakthrough: a robot. A small, disposable robot, costing just a few hundred dollars. A small, disposable robot...with a brain.
Gini is in the business of artificial intelligence, or AI. She and a team of graduate students at the UM are building small robots with intelligent processors capable of learning simple tasks. This particular robot, named Trailer-Backing Mini-Robot (TBMin), learns how to back up its trailer and touch its bumper to a light.
If you've gotten used to the notion of big, fancy robot androids that have human consciousness, you'll probably be disappointed in these little LEGO-bots. "People often want to use big computers, more expensive materials," says Gini. "But you don't need the super computer to make a robot. After all, insects have very small brains and they do very, very well."
In another computer lab a few floors below, Paul Rybski, the student who built and programmed TBMin (it took him about a month), demonstrates his machine. The testing perimeter is marked off by a stiff plastic fence, like garden edging, secured with duct tape in a large circle. At the circle's center stands a bare light bulb. Rybski positions the robot with its front bumper against the plastic fence and the photo-sensor on the back of the trailer facing the light.
Rybski flips a switch on TBMin. The robot buzzes and whines, jerking its front wheels to the left and right like a foal on new legs. TBMin creeps tenuously backward toward the light. Then, after a few feet, it jackknifes the trailer. It sits there for a dejected half second. Then it revs forward to the plastic fence and begins again. This time, it gets much closer to the light before it jackknifes. On the third try, TBMin has figured out how to keep the trailer a little straighter, but then it plows on past the beacon. When the sensor loses the bright light, the robot stops dead and then tries again.
And so it goes. TBMin moves backward and forward, trying to touch the light, but instead jackknifing or sliding past it--failing over and over and over again. And each mistake makes it smarter. "It learns from failure," says Gini.
TBMin's brain, she continues, contains 64 artificial neurons, divided equally between the robot's two functions: backing up the trailer and finding the light. "It doesn't have real neurons," she says, "but it tries to copy the brain." It's a very rudimentary insect, but pretty complicated nonetheless.
Consider what it takes for the robot to back up the trailer. Each neuron is activated by a particular angle of the trailer. So if the trailer sits at a 90-degree angle to the robot, then the "90-degree neuron" is active. It sends a simple numeric message to the robot's wheels: "Each neuron says, 'Turn left,' or 'Turn right,'" Gini explains. Positive numbers mean "turn right," negative, "turn left."
When TBMin is first turned on, the decision of whether to turn left or right is completely random: The neurons fire any old way, spraying positive and negative numbers all over, and steering the wheels back and forth. As soon as the robot begins to move, the numbers transmitted, positive or negative, begin to change. If they get larger, that tells TBMin he's on the right track. But as soon as the robot fails, when the trailer jackknifes or he sails past the light, the numbers given to a particular neuron fall. In other words, the robot is punished.
"Whenever it fails or succeeds," Gini says, "it updates all the values and then dumps them into the computer." And when the robot hits the light, "the bumper controls a switch that says, 'Success!' It's his reward, he gets more numbers."
By running thousands of trials on TBMin and a dozen other minirobots built in her lab, Gini has learned a few interesting facts about how robots learn. For example, if TBMin succeeds too quickly, it learns badly. "If it sees a lot of different situations it learns better," she says. In that case, it moves more and more slowly while the robot calculates all the variables from the failures. But once it succeeds under these conditions, it can handle any variety of starting positions and variables.