Sunday, September 10, 2000

Professor foresees super-robot

By Michael Wentzel
Rochester Democrat and Chronicle

        ROCHESTER, N.Y. — With inspiration from Pablo Picasso, University of Rochester researcher Randal Nelson has created a computer system that recognizes and remembers objects.

        Using a camera as eyes, his computer can correctly identify a model airplane, for example, even if half of the plane is covered.

[photo] University of Rochester researcher Randal Nelson has created a computer system that recognizes and remembers objects.
(Rochester Democrat and Chronicle photo)
| ZOOM |
        Now he wants to find a way to advance the system so it can talk, learn and move — a robot that performs more than programmed tasks.

        If it works, Mr. Nelson's “intelligent” computer eventually could help build a space station, perform surgery or, perhaps, just keep track of his car keys.

        “Computers in the future will help people act in the physical world as opposed to just finding information and what's in cyberspace,” said Mr. Nelson, an associate professor of computer science.

        Mr. Nelson and a team that includes 10 faculty members and about 30 graduate students received a $1.85 million grant from the National Science Foundation for the powerful computers and equipment needed for the research.

        “No one has looked at doing this because computer perception and recognition never worked well enough before to think about it,” he said. “This is very practical research. And as we talk about ways to make it work, we will also learn about the ways human intelligence works.”

        Although talking and thinking robots appear routinely in movies, it is only recently that fast and powerful computers have made even the most simple efforts possible in reality.

        Sandy Pentland, academic director of the Massachusetts Institute of Technology Media Lab, called Mr. Nelson's object recognition system “definitely state of the art.

        “It has a firmer intellectual base because it is based on the way biological entities perceive and learn,” he said. “His concept has legs.”

        Mr. Nelson had already developed a working system that could recognize a few objects when, about two years ago, he set off in a new direction inspired by the paintings of cubist artists such as Picasso.

        In the early 20th century, cubists played with perceptions. They painted a landscape or portrait in fragments, often taken from several points of view. A woman's face, for example, would be shown in angles or from many sides. Mr. Nelson constructed a computer system that adopted a cubist approach to objects.

        “After taking a whole bunch of pictures, the computer does something reminiscent of Picasso: It generates fragmented representations of the image, taking in all the prominent features,” Mr. Nelson said.

        The computer retains the fragmented images in its memory, where they can be used easily to match and recognize an object in another environment or in another position.

        “It figures out parts and how they are related in order to identify an object,” Mr. Nelson said. “Kids learn about things by watching them from different angles, by moving them. The robot would learn the same way. You'd play with objects like you'd play with them with a child.”

        The system recognizes six objects. For research purposes, it probably will learn about two dozen. With the current state of computer power and the system design, Mr. Nelson doubts a robot could learn more than a couple of hundred objects.

        With a link to a robotic arm, the system now can pick up the correct object when someone points to a picture of it: It can move the arm, for example, to grasp the model airplane.

        The team's goal is a system that communicates with humans to get information and accept instructions to work together on tasks.

        Mr. Nelson does not minimize the problems ahead. Robots are bulky, clumsy and expensive. Because of robot problems, much of the research on computer perception and manipulation will be conducted in virtual reality, where the objects and the actions will appear in computer-generated space.

        The computer system that drives object recognition and one that operates natural language are very different, Mr. Nelson said.

        “How do you tie these together? How do you give words to a computer, give them meaning and connect them to the real, physical world?” he said. “This is difficult, or at least, we think it is.”

        Within five years, Mr. Nelson predicts a system that knows simple objects, responds to voice commands and grasps objects.

        He envisions an intelligent system that could learn the objects needed to construct a space station and how to help build it. He also sees a home system linked to multiple cameras that would remember designated objects. It could identify a favorite coffee mug or car keys, for example, and find them when they get lost.

        Randal Nelson's Web site, which has research and illustrations, can be found at:


Ohio electric consumers have the power
California-type crisis not expected in Ohio
Frequently asked questions about power restructuring
Ready, set . . . .PUCO chairman lays plans for electric deregulation
Tiger may be boosting even Kroger Senior Classic
A big fish in little ponds
E-mmediate results?
- Professor foresees super-robot
Small-business diary
Small business
What's the Buzz?