Rise of the Robots?

E-mail Melanie Martella

Yesterday was Charles Darwin's 200th birthday and so it seemed fitting to write about how researchers have created a robot that evolves. There's also the fact that I felt in need of a lighter topic in the midst of continuing dire economic news.

Robots who learn (often with terrible consequences to the surrounding humans) are a common feature in movies and fiction. Researchers working with robots have long wrestled with how to construct robots that can learn. Now, Christopher MacLeod and his colleagues at Robert Gordon University in the U.K. have built a robot that mimics biological evolution to adapt to physical changes.

As explained in the New Scientist article "Unnatural selection: Robots start to evolve", the robots created by the researchers use a neural network. This isn't new and artificial intelligence researchers have been tinkering with neural networks in robots for quite a while. What is new, however, is their use of an incremental evolutionary algorithm (IEA).

Typically, with a neural net and for a given primary command (required/desired behavior), the robot will try a bunch of things until it finds some combination that works. And, over time and trials, it will develop the optimal control method to achieve that primary goal, at which point the neural net is frozen. MacLeod's team, however, added a little something extra. Their IEA pays attention to whether the robot suddenly ceases to meet its primary command.

The researchers started off with simple box-like robot, added peg legs, and told it to walk as fast as possible for 1000 s. Once the robot had learned to do that, they added jointed legs. The robot, faced with the fact that it could no longer achieve its goal, assigned new "neurons" (processing nodes) to figure out how to use these new, wobblier legs, building on its earlier work with the peg legs. Once again, the robot evolved a method of locomotion that met its requirements. This same process kicked in if the researchers added more legs or, interestingly enough, vision sensors.

We're still a long, long way off from the Terminator knocking on our doors and telling us to accompany them if we want to live. For starters, most of our current robots do poorly in uncontrolled (dynamic) environments and are sorely limited in the tasks that they can carry out. However, taking a page from nature's book and enabling complex behavior by building it up, layer by layer, from the basic to the complicated, is a really interesting approach. Let's see if it has legs.