Fan of the “Terminator” movies? You've seen successive generations of the android assassins develop increasingly fearsome abilities in their quest to kill future human rebel leader John Connor. But while the movies are entertaining, you might think the idea of robots evolving on their own seems pretty far-fetched — or at least, not something that's going to happen in the foreseeable future.
If that's what you think, you puny meatbag, guess again.
Researchers from the Institute of Robotics and Intelligent Systems at ETH Zurich and the University of Cambridge engineering department have demonstrated that machines can build mechanical offspring, evaluate their performance, and then pass the best-performing characteristics to successive generations.
In the experiment, the team of Luzius Brodbeck, Simon Hauser and Fumiya Iida created a “mother robot” that built children out of a few simple parts that were glued together. The baby robots were then transported to a testing area, where they were activated with wireless controls. Their movements were then observed with a video camera and evaluated by a software program.
The software identified the designs that enabled robots to move the farthest in the least amount of time, and then directed the mother robot to another generation with those characteristics. Over time, some of those successful characteristics were blended together, in the same way that living creatures create hybrids.
Meanwhile, the unsuccessful designs essentially went the way of the Neanderthals.
Based upon computer simulations, scientists already believed that robots could evolve, says Josh Bongard, a University of Vermont computer scientist who has been researching how robotic evolution might work. But the latest experiments mark the first time that actual robots have shown the capability to evolve, says Bongard, an editor of the journal article explaining the process.
Robotic evolution actually emulates the sort of selective breeding that humans have practiced for thousands of years with animals and plants, Bongard points out. “It's actually an old technique, just redone with modern technology,” he says. “Our ancestors evolved wolves into dog breeds, and turned wild grains into wheat by selectively breeding for the characteristics that were most useful to them. But now, what we'll hopefully be doing is breeding more useful and interesting robots.”
Bognard says that what we're doing is essentially boiling down evolution into a computer equation. “If you implement this algorithm, you'll gradually see the robots get better at what you're asking them to do,” he says.
In exploring the possibilities of evolutionary robotics, computer simulations still have some advantages over real robots for the time being. It's possible to breed virtual generations more rapidly than present robotic technology can create more mechanical offspring. But eventually, he envisions that advances in 3-D printing will enable robots to make new and better versions of themselves far more quickly than they can now.
Bongard and his colleagues have found that robotic evolution can produce strange and unexpected results. “Sometimes, we get evolved robots that look like things we might see in nature,” he says. “Other times, they don't have a biological analog.”
For example, in one recent study, his team instructed the computer software to evolve robots who could climb over rough terrain. The result: Generations of virtual robots who evolved arms and hooks, and could push and pull themselves forward. “They looked like crabs and lobsters,” he says.
But in another experiment, the scientists tried to evolve a four-legged robot that they figured would walk and run like a dog or horse. “Instead, it lay on its belly and did the worm, that move from 1980s breakdancing,” he explains with a laugh.