|
December 1, 2000
|
A New Race of Robots
| Evolution in Action |
|
Robot technology has come a long way, but robots still can't think for themselves. Or can they? How has the Golem Project combined computer science, physics, and biological theory to build a better robot? A small tangle of plastic pipes inching along the ground in a Brandeis University laboratory may not look like the future of technology. These primitive robots scarcely resemble the sophisticated versions NASA is building to explore space. Even the robotic dogs that are so popular this holiday season can do many more tricks—except for one. The robots that are part of Brandeis's Golem Project are able to evolve and change on their own. "If you look at the history of robotics, there have been some very nice achievements," says the Golem Project's Hod Lipson. "But they use a high level of human engineering. It takes thousands of engineers to design the robots for our Mars missions." In contrast, Lipson adds, the Golem Project minimizes the human role in robot development. Rather than receiving instructions on how to perform certain tasks, these robots discover how to behave on their own. In that respect, the Golem robots live up to their namesake. In Jewish legend, a golem was a lump of clay in human form that magically came to life. What becomes an actual robot in the Golem Project starts out as a computer program, similar to the ones in sophisticated computer games. This program contains the designs for the plastic pipes, joints, motors, and electronic circuits that will eventually become the actual robot. The computer also is programmed to understand the physics of movement. The next part is up to the computer, which randomly matches these parts so that the resulting robot can move on its own. This process of trial and error--in which the design of the new robots undergoes continuous changes, or mutations--follows the path of Darwinian evolution. And the Golem Project isn't done with Darwin yet. A computer simulation tests the new robot design to determine if it can move according to the laws of physics. If the design does not work in the real world, the computer discards it and saves only the designs that show the promise of movement. These survivors continue to mutate and improve. Thus, the computer mirrors the process of natural selection, but many times faster than the original development of living creatures. What took nature a million years can take just days for a high-speed computer. |
|
| From Bacterium to Insect |
|
The human scientists of the Golem Project build the robots that have been designed, tested, and selected by the computer. The results often are not what humans would have created. Some of these robots creep. Others move sideways. Still others drag themselves along. But they all move successfully. You don't need to be in a laboratory to see these results in action. The Golem Project has created a screensaver program for personal computers that creates its own moving shapes. "If you run it at home," Lipson says, "in a day or two, on average, you will see crawling robots unique to your computer."
But the activities you see on your screen--and what the Golem Project's researchers are finding in their labs--are pretty basic. "They are little more than toys, with the brains of a bacterium," writes Lipson and his research partner, Brandeis computer science professor Jordan B. Pollack, on the Golem Project Web site. "We hope to achieve insect status in a couple of years. While the bodies and brains are simple now, it is just the beginning of our long-range project." In fact, the only task these robots perform is to move. The next steps might include maneuvering around objects in the way, or reducing the amount of energy while in motion. In each case, the computer program designing the robot would be encountering problems it had not see before and would have to discover the solutions. Lipson sees the problems and solutions becoming more complex. "Let's say you'd like the robot to pick up a Coca Cola can from the floor and throw it into the trash," he says. "The robot would have to learn to recognize the shape of a soda can, to move to its location, to pick the can up, to identify a trash can, and to drop the soda can into the trash." The key, Lipson emphasizes, is that the robot figures out what to do without being told by humans. |
|
|
|
There's more than scientific curiosity driving the Golem Project researchers. They envision a time not far off when these self-learning robots can be used for a variety of commerical purposes. "You don't see robots in daily use because of the high costs," says Lipson. Those costs, he explains, come from the human time and expertise it takes to design them. For instance, the robots fitting and welding parts on automobile assembly lines have had every detail of their movements determined by human engineers. In contrast, robots that are designed by computers to do specific jobs would be much less expensive to build. They could then take the place of humans in certain situations, such as cleaning up the stands in a football stadium after a game, or figuring out how to locate live mines in a mine field. Lipson also foresees how human engineers and "thinking" computers could share responsibilities. "Let's say you were building a bridge over a canal with a very particular slope on its banks," he theorizes. "You would enter these features into the computer and it would generate a blueprint for the bridge." That blueprint would be the computer's own creation, based only on its programmed knowledge of physics. How it solves the problem with the data you've presented is its own business. |
||
|
|
|
|
|