Artificial neural network guides ‘playful robots’

August 13, 2008

Robot Article Picture
Two humanoids wrestling, controlled by the new artificial neural network.

By controlling computer simulations with an artificial neural network that is roughly modeled on the human brain, researchers hope to create robots that can take on lives of their own, as much as possible.

That’s because the software doesn’t give any explicit instructions to the simulated creatures on how to move or what goals to aspire to.

The network, developed by researchers at the Max Planck Institute and Edinburgh University, works with different kinds of animals, such as dogs, snakes, and humanoids. Each simulated creature has about 15-25 joints that can move in various ways. By experimenting with how they can move their joints, the creatures slowly learn to do things – similar to how a baby learns to move its legs and eventually crawl and walk.

The creatures move by trial and error, and the researchers are working on giving them long-term memories so that they can remember what to do when they encounter familiar situations. One simulation shows a humanoid thrashing its legs and arms around on the ground, learning what it can do and exploring its surroundings. Because their behavior emerges from scratch by way of trial and error, the scientists describe the creatures as “playful robots.”

The network follows the concept of “self-regulation,” which explains human behavior as a balance between blindly following rules and creatively adapting to new situations. Computer scientist Chris Langton uses the term “edge of chaos” to describe this trait, which is unique to living things.

At an artificial life conference in Winchester, UK, Ralf Der of the Max Planck Institute revealed demos of the virtual creatures. In the videos, a dog learns to jump over a fence, a snake figures out how to jump out of a hole, and a humanoid learns to stand up, do push-ups, and perform back flips.

"It doesn't know anything but tries motions that are natural for its body,” Der said. “Half an hour later, it's rolling and jumping.”

One scenario was particularly interesting. When two humanoids come in contact with each other for the first time, their first action – based only on natural joint movements – is to spontaneously begin wrestling with each other.

via: The Telegraph

Robot World News homepage

Search Archives

 

Top Stories of the Week

Coming soon.

Exclusive Stories

Coming soon.