The prospect of robots creating other robots fascinates both science fiction writers, who typically envision mechanized predators running amok, and artificial intelligence researchers, who see them faithfully serving humanity. Despite the fears and hopes, self-replicating robots have remained firmly in the distant future. But now two computer scientists report a tiny step toward that future with a robotic system that designs and builds robots with just a bit of help from a human hand.

The first stage of the new project--using computer simulations to select the most fit robots--had been done before. Computer scientists Hod Lipson and Jordan Pollack of Brandeis University in Waltham, Massachusetts, started with a few virtual building blocks: straight bars for structure, joints to connect the bars, and artificial neurons for control. Then a computer program fiddled with the bars and neurons, selecting models with the most mobility. After 300 to 600 generations, the program had evolved robots with various ways of getting around. The most sophisticated model was designed to crawl on three legs, they report in the 31 August issue of Nature.


Once Lipson and Pollack had simulations of the fittest robots, they took the project one step further than previous attempts. They hooked the simulation computer to a rapid prototyping machine, commonly used by industrial engineers to build models of parts and products that are being designed. The computer instructed the machine to construct the best robots. In a step Pollack admits is "cheating a bit," however, the researchers inserted by hand some computer-evolved motors that are still beyond the capabilities of the prototyping machines. In the end, the real-world robots replicated their evolved virtual movements.


"It's a clever piece of work," says John Koza, consulting professor of bioinformatics at Stanford University, California. But he adds that the robots don't yet fulfill the ideal of a complete evolutionary cycle. "This is an intriguing portion of the cycle; what's missing is the self- reproduction." But Pollack says they are not out to produce a "mythical humanoid robot working in a machine shop making more humanoid robots." Instead, their more modest goal is to make robot design more cost-effective.


In other robot news, a team at the University of Lausanne, Switzerland, programmed groups of small robots to behave like ants. They knew when and how to forage, how to avoid bumping into each other, and how to recruit colleagues into following them to a rich stash of resources. The social robots were more efficient and effective than single robots, Laurent Keller's team reports, suggesting that setting loose swarms of robots might be a robust and flexible strategy for exploring distant planets.