Genetic Learning Algorithms Combined With Novel Binary Hill Climbing Used for Online Walking-Pattern Generation in Legged Robots

Abstract

According to Darwin every species on this planet have developed froma small group of simple molecules into all the modern species living among us today. The reason why some species survive and others don’t is what Darwin called Natural Selection, which means that every individual have to fight for its existence. Those who are best fit will survive. This has brought life to the well known saying: "Survival of the Fittest". The best fit will have the best chance to reproduce, to pass its well fitted, surviving qualities on to their offspring. And the offspring of two well-equipped parents will have a high probability of adaptation, and so the circle of life goes on... A set of evolutionary search methods have been extracted from the Darwinian theories of evolution. These have been evolving in computer environments for several decades and have been passing through different areas of computer science, from theoretical tuning problems, algorithm developing, clustering, chip design, and several real world applications have been the foci the last years. In this thesis Genetic Algorithms and Evolvable Hardware is used for evolving gaits in a walking biped robot controller. The focus is fast learning in a real-time environment. An incremental approach combining a genetic algorithm with hill climbing is proposed. This combination interacts in an efficient way to generate precise walking patterns in less than 15 generations. Our proposal is compared to various versions of Genetic Algorithms and stochastic search, and finally tested on a pneumatic biped walking robot.