A University of Vermont researcher has created robots that are capable of evolving, much like tadpoles becoming frogs. 

Josh Bongard, creator of the evolving robots and an assistant professor in the University of Vermont's College of Engineering and Mathematical Sciences, has simulated and created robots that change body performance over time instead of having a fixed body form and behavioral traits like other robots. 

Up until this point, robots have been designed and built one specific way and are programmed directly instead of having to learn certain behaviors. But Bongard argues that this method may not produce the most efficient robots. 

Instead, Bongard has created robots capable of evolving both its body and behavior over a period of time, much like the way humans grow from babies to adults. The goal is to create four-legged, upright robots that can move to a light source without falling.  

Bongard's robots were made in many different ways. Some start out flat on the ground, or like snakes with legs while others may have splayed legs like lizards. The robots may start out in any of these forms, but the end result is that they form upright legs and know how to use them. They have 12 moving parts and are very simple-looking structures with a jointed spine, the look of a mammal's skeleton, and four sticks for legs.

The prototypes for these robots were made of Lego's, showing how the evolution of these robots would work. They were made as four-legged "robots" like in the simulation, and wore braces on its front and back legs that would tilt it. This causes the controller to look for successful movement patterns, which results in the legs going from horizontal to vertical. They would go from a reptile to a quadruped. 

"We built a relatively simple robot out of a couple of Lego Mindstorm kits to demonstrate that you actually could do it," said Bongard. 

To make the real robots, Bongard first ran 5,000 computer simulations -- each taking about 30 hours to complete -- on the University of Vermont's parallel processors to create synthetic models that move around in 3-dimensional space. Each generation of each creature then "runs" a genetic algorithm, which is a software routine that helps the creature learn different body motions such as slithering, crawling or walking. An appropriate motion is applied to each generation of the creature, giving it a proper plan to be able to obtain the goal of moving toward a light source without falling over. 

"The snake and reptilian robots are, in essence, training wheels," said Bongard. "They allow evolution to find motion patterns quicker, because those kinds of robots can't fall over. So evolution only has to solve the movement problem, but not the balance problem, initially. Then gradually over time it's able to tackle the balance problem after already solving the movement problem."

Bongard noted that robots who evolve this way learn to walk more quickly and have a more "robust gait" than those that were built with fixed bodies and behaviors. In tests, Bongard's evolving robots were able to reach the final goal of moving to the light source without falling over faster than non-evolving robots. Also, researchers found that the robots were able to attempt new kinds of challenges that were not previously given to them after reaching the light source. This may be because controllers used in the evolving robots could have maintained a certain behavior over a wider range of sensor-motor-related functions while controllers in fixed robots did not. 

"We're copying nature, we're copying evolution, we're copying neural science when we're building artificial brains into these robots," said Bongard. 

This study was published in Proceedings of the National Academy of Sciences.