There was a cool story about a roboticist at University of Vermont that used genetic algorithms to teach robots how to walk. Robots were created in a sophisticated computer simulation and were given a task of walking toward a light without falling over. Some of the robots started like tadpoles, some like snakes with legs, some like lizards, and some like mammals with legs directly beneath their body. The genetic algorithm allowed the designs to evolve and let natural selection take its course. The interesting finding was that the robots that started out with a more primitive form and evolved to the mammalian form performed better than the robots that started in the mammalian form. These robots were mimicking how humans learn to walk. They first learn forward movement and don't worry about balance since it is harder to fall over with a reptile-like frame. Once they master forward movement, they can learn balance and improve their speed and gait. I think this may be a promising technology for future robots. Instead of trying to "play god" and program extremely complicated maneuvers manually, they can build the framework and leverage the power of evolution to come up with solutions itself.
http://www.sciencedaily.com/releases/2011/01/110120131542.htm
No comments:
Post a Comment