Beware. Scientists are Creating Machines That Can Evolve on Their Own

Forget what you see in movies. Most robots sit in a factory somewhere doing dull, repetitive work. Even if their software does dream of a more interesting job, their physical form remains fixed from the day they’re assembled to the moment they’re junked.

But it doesn’t have to be that way, say pioneers of “evo devo robo”—evolutionary developmental robotics, which applies principles of natural selection and biological development to machine design.

Josh Bongard, a computer scientist at the University of Vermont, is designing robots that go through growth spurts. Like anxious teenagers, they experience an awkward period of physical development before they find their place in the world.

Bongard’s virtual experiments precisely model robot actions on a computer, beginning with a limbless, wormlike machine with several body segments. It has one goal: Make progress across a flat surface toward a light source. But instead of giving the robot a complete set of instructions, Bongard generates a whole population of bots and lets evolution do the work.

Each virtual robot moves randomly—for instance, one segment might move an inch to the left and another might bend 90 degrees. While most of the bots flail hopelessly, a few, by chance, nudge forward. The steps behind all the little advances are stored and combined in a process that mimics the mixing of genes in living, reproducing beings. Then, at some point, Bongard steps in and gives the successful robots legs. And evolution continues.

Bongard has found that he can evolve a contraption that walks upright in just 100 generations. By contrast, a contraption that starts with legs already formed takes 250 generations. “It makes sense to stay close to the ground when you’re young,” he says, “and only gradually grow legs and stand upright, which makes you more unstable.”

A robot that can grow would be highly useful. Imagine a spacecraft landing on an alien world growing an extra pair of legs when it encounters treacherous terrain. Or a search and rescue bot that morphs from stocky to slender to navigate a tight crevice. But will such a machine ever be more than a computer simulation? “It’s the last manufacturing step that has always haunted roboticists,” says Hod Lipson, who directs Cornell University’s Creative Machines Lab. “You can design crazy robots, but in the end you have to make them.”

Lipson’s work might help. He has used a 3-D printer to churn out robot components and has demonstrated a machine capable of assembling copies of itself with premade components. Those feats suggest it might be possible to design a bot that manufactures itself—that is, evolves a body plan and builds it.

For now, Bongard wants others to join the experiments. In August he made his virtual robots available online. He calls them ludobots, ludo from the Latin play. “Primates are evolved to recognize a limp in another animal,” he says. “We would like to crowd-source robotics,” so people can build machines, observe how they move and fix flaws.