For the first time ever, a computer programmed to follow the rules of evolution has designed and manufactured simple robots with minimal help from people.

The 8-inch automatons did not take over the world or even vacuum the lab. Instead, they crawled across a tabletop, exactly as they were digitally bred to do, said Jordan Pollack, a Brandeis University computer scientist.

“It’s not what our robots do that is so surprising,” he said. “They’re not humanoid robots—they don’t raise their eyebrows and make you giggle. But what they did do was autonomously designed and manufactured.”

By having a computer create designs using natural selection, researchers edged closer to solving two of robotics’ biggest obstacles: robots’ lack of versatility and their high cost of development.

Robots engineered by people typically function only under specific conditions with limited ability to adapt to changing situations.

A simple robot that vacuums a floor, for instance, could cost millions to develop and sell for $5,000 after engineers figured out a way to make sure it doesn’t crash into furniture or fall down stairs, Pollack said.

“Then again, you could just hire a minimum-wage worker with a $100 manual vacuum,” he said. “The cost of building an intelligent humanoid robot is so high, we just can’t get the economics going.”

Ultimately, the Darwinian approach could revolutionize everything from manufacturing to space exploration.

“Down the road, if we could have a thing like this in space, you could send the building blocks and let them evolve themselves,” said Yoseph Bar-Cohen, director of a robotics lab at NASA’s Jet Propulsion Laboratory. “That would be fascinating.”

Pollack and colleague Hod Lipson merged automatic manufacturing techniques with evolutionary computing. Their results appeared Aug. 31 in the journal Nature.

The computer that evolved the designs was told only what parts it would be working with, the physics of the environment in which its offspring would be moving, and the goal of locomotion.

Over several days, the computer thought up different designs and methods of movement, creating traits that worked and failed. Like dinosaurs, woolly mammoths, and dodo birds, the failures were cast into the dustbin of history.

The most promising designs survived and passed their success to future generations. Hundreds of generations later, three robots were manufactured by a prototyping machine.

“It evolved various kinds of locomotive mechanisms—all surprising, given there was no human coming up with how to do it,” Pollack said. “We got ratcheting motions. We got rolling motions. We got swimming motions.”

The little white robots were made of bars, actuators, ball joints, motors, and circuits. People intervened only to insert the motors into the plastic parts spit out by the prototyping machine.

The next step will be to incorporate sensors into the robots so that success or failure in the physical world can be built into future generations.

It could be a difficult project, said Maja J. Mataric, a professor at the University of Southern California and director of the USC Robotics Research Labs.

“The authors very cleverly figured out a way to design a body and then actually manufacture it, which is an amazing feat,” she said. “What is not at all clear is how to come up with a sensor design and manufacture it.”

Brandeis University