A study to find the quickest way to evolve walking behaviors in virtual robots was conducted by University of Vermont researcher Josh Bongard. He ran simulations involving genetic algorithms on different types of robots as they tried to reach a virtual light source and evolve a walking gait to get there. One robot slithered along the ground like a snake and gradually grew four legs, while a second robot began the simulation with four legs that moved into the vertical position. The third robot started the study with four vertical legs but could not evolve a more advanced body type.Once each robot was able to walk upright on four legs, Bongard ran the algorithm on an actual robot to study the results.
The robots that had evolved the ability to walk on four legs were much more stable than the robot that was able to walk the entire time, but could not evolve. "This is what human infants do—they progress from crawling to walking gradually, even as the bones in the legs and feet change to accommodate the change in behavior," Bongard says.
View a video of robots changing their body plan to improve their balance.
Other researchers also have been studying changing body plans in robots. Robert Gordon University's Chris MacLeod used a similar algorithm to teach two-legged robots how to adjust their gait to deal with the addition of two extra legs.
From New Scientist
View Full Article
Abstracts Copyright © 2011 Information Inc., Bethesda, Maryland, USA