Robots that evolved from crawling babies into upright adults could help pave the way for better bots.
Using a computer program, researchers at the University of Vermont simulated a population of naive "baby" robots. The robots had to complete various tasks in their virtual environment, such as finding objects and walking toward them. Those robots that performed poorly got deleted, while the best-performing ones remained "alive."
The robots that changed their body forms (like tadpoles growing into frogs) learned to walk more rapidly and developed the most stable gait, the researchers found.
“We learned that it is easier to breed robots that change shape,” said lead study researcher Joshua Bongard, a professor of computer science.
In addition to creating virtual robots, Bongard and his team built a real robot that could gradually change from a legless "snake" creature to an upright quadruped.
“If you're a snake, it's easier to learn how to move, because you don't have any legs so you can't fall down while walking,” he said. “Once the snake learns how to move, it can adapt its brain to keep moving forward while it grows legs, and learns, in addition, how to maintain balance.” [Read also “Three-Legged Dog Teaches Robots New Tricks.”]
Robot DNA
Before the physical bot, Bongard and his team created simulated robots, each loaded with virtual brains and bodies.
“The brains of our robots are very, very simple: They're made up of only a few dozen simulated synapses and neurons,” Bongard said. (The human brain, by comparison, has 100 billion neurons.)
The virtual robot bodies differed from those of today’s actual robots by being able to change shape as they moved around their virtual world. “Some robots start as snakes slithering toward the target object, but as they approach they begin to grow legs; by the time they reach the target object, they are standing upright on four legs, very much like a dog or horse,” Bongard said.
After 5,000 simulations, the team found the perfect bot — at least perfect for performing the specified tasks of their virtual world. To prove that this wasn’t just theory, the researchers built a physical robot using Lego Mindstorm kits. The Lego creature had four legs and wore a brace that let it gradually move from its belly to an upright position.
"While the brace is bending the legs, the controller is causing the robot to move around, so it's able to move its legs and bend its spine," Bongard said. “It's squirming around like a reptile, flat on the ground, and then it gradually stands up until, at the end of this movement pattern, it's walking like a coyote."
Building better bots
The next step will be to build more actual robots that can rapidly change form, Bongard said.
The researchers want to “better understand why changing shape helps us build better and more-complex robots,” Bongard said. “We'd like to study robots that can move about, pick up different objects and create different piles of different objects.
“Such robots could be useful on a construction site, and might someday greatly reduce the cost of construction, much like industrial robots helped reduce the cost of manufacturing things.”
Bongard and his team detailed their findings online Jan. 10 in the Proceedings of the National Academy of Sciences.
