Robots evolve to learn cooperation, hunting

If robots are allowed to evolve through natural selection, they will develop adaptive abilities to hunt prey, cooperate, and even help one another, according to Swiss researchers.

In a series of experiments described in the journal PLoS Biology, Dario Floreano of the Ecole Polytechnique Federale de Lausanne and Laurent Keller of the University of Lausanne reported that simple, small-wheeled Khepera and Alice robots can evolve behaviors such as collision-free movement and homing techniques in only several hundred "generations."

The robots were controlled by a neural network that mutated randomly, with input information from the robots' sensors. In an imitation of natural selection, the robots with the best maneuvering abilities were allowed to foster a new generation. Furthermore, selected robots were "paired" by having their neural net connections mixed and passed to a new generation.

Within 100 generations, the robots were able to move through a maze without bumping into anything.

Another experiment involved the robots pushing round discs along a wall to get points. Robots were found to cooperate in pushing larger tokens together to earn points for their whole group. The researchers also observed altruistic behavior, in which robots would sacrifice getting points for themselves in favor of getting points for their whole group of related robots. This was measured by pushing small and large discs into a zone--the large discs required a cooperative effort to move.

The researchers described "spider" hunting techniques among the bots in which hunters would lie in wait for prey (which in this case, fortunately, consisted of other robots). The hunted, meanwhile, developed a strategy of "quickly (rotating) in place, which reduced the probability of being approached from the sides without sensors."

I'm glad these guys are experimenting with little robots on wheels, and not military Swords robots.

(Via TG Daily)