Evolving Hexapod Gaits Using a Cyclic Genetic Algorithm

Abstract

Gait generation is an integral part of a legged robot control. Hexapod gaits require the coordination of the simultaneous movement of all six legs in a cycle of activations. The problem is compounded when working with actual robots due to the variability in their capabilities. A learning algorithm that can optimize with the differences between and within robots would greatly reduce engineering calculations and increase robot adaptability. To be effective for the simplest of robot controllers, these algorithms should produce sequences of activations that can be used directly on the robot. Evolutionary computation would be ideal for this learning algorithm although few of its forms have a structure that conforms naturally to the cyclic nature of gaits. In previous work, we developed Cyclic Genetic Algorithms (CGAs), which are a variation on the basis genetic algorithm. Tests on robot simulations showed that these algorithms could rapidly converge to evolve the optimal tripod gait. In this paper, we successfully apply CGAs to an actual robot that is not only more complicated then the original simulation, but also has the increased variability innate in actual robots.