Robots Evolve Cooperative Behaviors, Learn To Hunt And Be Hunted

Writing by Evan Ackerman on Tuesday, 2 of February , 2010 at 3:15 am

Last time we heard about robot evolution, the bots were figuring out how to deceive each other. Now, researchers at EPFL in Switzerland have been using the same sort of genetic programming techniques to enable robots to teach themselves how to solve mazes, cooperate on tasks, and hunt each other (we’ll save that one for last).

The way genetic programming works is that the robots are only programmed on a very basic level, with simple information on their sensors and objectives. At first, the robots are clueless as to how to take the information from their sensors and apply it to completing their objectives, but after each test, random variations (mutations) are introduced into the code. Robots that demonstrate the most improvement have their code passed on to the next generation, and the process was repeated a bunch of times. In this experiment, after 100 generations, the robots taught themselves how to navigate a maze without running into a wall, and figured out that having their sensors pointed in the direction that they were going was the best way to be. It gets cooler:

“In another experiment they programmed groups of robots to push tokens along a wall to a marked area to win points. They selected the robots that gained the most points to pass their code on to the next generation. Over time altruistic behaviors were observed, in which robots sacrificed points if the entire group would benefit, and the robots cooperated to push larger tokens together to earn more points. As in nature, the robots followed the biological principle of kin selection, in which they only helped robots having the same code lineage.”

Code-based kin selection. Crazy, huh? It makes sense, though, if you think about it… Robots work together best if they’re using the same code, and robots who aren’t using that code won’t know how to cooperate with the robots that are. So, they’ll end up being less efficient, and won’t be as likely to make it into the next generation.

The predator-prey dynamic is perhaps the most interesting. One group of robots with several sensors were programmed to chase after another group of robots, who had fewer sensors but were faster. At first, the predator robots simply chased after the prey robots, who ran away. After 125 generations of evolution, the predator robots had figured out how to stalk the prey robots, hiding out and then sneaking up on the prey’s blind spots. The prey robots, on the other hand, learned where the predators liked to hide and made sure to keep their sensors facing them.

All this, in just a hundred or so generations in a lab. Kinda makes you wonder why we don’t just set a couple hundred Roombas loose in a dust covered wearhouse, let them fight it out and breed with each other, and after a couple years we’d get the world’s smartest, most efficient, and deadliest robot vacuum.

[ Evolution of Adaptive Behaviour in Robots by Means of Darwinian Selection ] VIA [ Physorg ]

The paper also refers to another robot evolution project called Golem@Home, which used distributed computing to design a moving robot from scratch. Different simulations were raced against each other, and the winners were actually created. We posted about it back in 2007, it’s pretty cool.

Comments (2)

Category: General

• Add this post to
• Del.icio.us -
• Digg

Name (required)

Mail (will not be published) (required)

Website

XHTML: You can use these tags: <a href="" title=""> <abbr title=""> <acronym title=""> <b> <blockquote cite=""> <cite> <code> <del datetime=""> <em> <i> <q cite=""> <strike> <strong>