University of Vermont

cems
College of
ENGINEERING AND MATHEMATICAL SCIENCES

Extended Q&A With Josh Bongard

Written by Melissa Pasanen

Josh Bongard

University of Vermont associate professor Josh Bongard talks about what robot engineers can learn from nature, why he won a prestigious presidential award and if we should really worry about a robot uprising. (This is an extended version of the interview that appears on page 72 of the Spring 2014 edition of Vermont Life.)

VL: When did you first make the connection between technology and the natural world?

JB: Well, I played a lot of video games as a kid but I was also interested in building things and I loved the outdoors and animals. I was always fascinated by animals: how they move and how complex their bodies are, but incredibly efficient and fast. Machines, to this day, are still relatively clunky. Hollywood showed us all these fantastically amazing machines like big metal cats and humans but we didn’t actually have them and why not? That question motivated me. I was an undergraduate in computer science when I read a book that opened my eyes to borrowing ideas from nature and using them to build machines. I never looked back.

VL: Now you use computers to evolve robots based on the model of natural selection. Why is that a good way to develop robots?

JB: Humans have figured out how to build better cars, airplanes, the international space station, but autonomous robots are much more complicated than any of those machines for one reason: they need to move and live alongside us in the real world without a human leading them. These kinds of robots are different from industrial robots, which are designed to do the same thing over and over again in a controlled environment. The biggest challenge is not to get a robot to do one thing well, but to get one robot to be able to do a dozen things well enough. An autonomous robot has to be adaptive because its world is always changing around it. Again, humans and animals are good at changing what we do based on circumstances. We don’t yet know how to make machines that do that, but Mother Nature, or natural selection, has produced adaptive machines for billions of years and they’re very good, so why not borrow that idea and teach the computer how to evolve robots for us?

VL: And what does this look like in practice?

JB: In evolutionary robotics, with our input, computers create hordes of robots and then we tell the computer to select for certain traits, like you’d select for milk capacity in a cow. So we might tell the computer to select for speed and agility and avoidance of obstacles and self-damage, or the ability to reach for an object, distinguish between objects and put the blue objects to the left and red to the right. Computers are relentless, tireless; they can test things over and over again. The results aren’t always what we anticipate. We had a robot with four legs and challenged it to learn to walk. We thought it would evolve to walk like a horse or a dog, but it moves more like a break dancer doing the worm.

For the complete interview visit Vermont Life

Posted with permission from Vermont Life magazine.