A typical robot today is good at simple, repetitive tasks: Picture a robotic arm on an assembly line doing the same thing hour after hour. But some researchers are trying to help robots evolve past these basic limitations.
“We’re trying to break the old paradigm and involve robots in everyday tasks out in the world, interacting with things like animals do and like we do,” says Nick Cheney, a PhD student at Cornell University who’s been leading research in evolutionary computation at the Cornell Creative Machines Lab.
In this video, the lab’s virtual robots become blobby but effective runners. The team gave the robots four basic parts--the red and green parts act like muscles, while the blue parts act like soft tissue and bones. After programming the creatures with one goal--to become faster--the researchers sat back to see what would happen.
After 1,000 simulated generations, the robots had evolved to move in every imaginable way, from trotting like a horse to pushing and pulling along the ground.
“We don’t have to have an engineer sitting down at a computer with a real intuition of exactly how to design something from the ground up,” says Cheney. “It opens things up and takes human intuition out of the loop, which is something that’s a powerful tool but is also holding us back in certain ways. If we don’t understand how something works, like animal behavior, it’s hard to design it.”
The lab is currently working on creating actual robots from some of the simulation. “It’s a great proof of concept to show that we really can transfer these things,” Cheney explains. “Not that robots walking around is an end goal for this work--I think that we can scale up to more complex adaptive behaviors.”
Evolutionary computation isn’t new, and as long as 20 years ago a researcher named Karl Sims had already shown that he could make a program that created somewhat lifelike forms. But there hadn’t been much progress until recently.
Most previous work focused on evolving robot brains rather than their bodies, Cheney says. And research in past was also limited by the sheer amount of computing power it takes to simulate evolution. “Biological evolution is millions and billions of years,” Cheney says. “It’s a very long trial-and-error process. Having a complex environment also takes a lot of simulation, and it’s a drain on resources.”
“It’s only recently that we’ve been able to have the kind of soft-body simulator that I’m using, combined with evolutionary computation to be able to produce more complex and more compliant materials in robot morphology like you see in this video.”