Adaptive Hexapod Gait Control Using Anytime Learning with Fitness Biasing

Adaptive learning systems that generate control programs for robots with varying capabilities is of importance in the implementation of autonomous robots. Learning done continuously with the best possible control program running the robot (anytime learning) can achieve the adaptability desired when implemented using some form of evolutionary computation. The difficulty with this method is that autonomous robots often lack the computational power required to run evolutionary computation along with their control program. In addition, anytime learning usually requires input from internal sensors, which are not often available in small autonomous robots, to make adjustments for capability changes. In this paper, we propose an anytime learning system that employs off-line learning, using evolutionary computation, with the control program being downloaded to the on-line controller. The off-line learning does not require internal sensors but uses global observation (external overhead camera) to make the required adjustments to guide the evolutionary computation. The results of periodic tests, done on the actual robot, are used to bias the fitnesses calculated by the evolutionary computation, which uses a model of the robot. Experiments reported in this paper use a simulation of the actual robot (a more accurate model), while construction of the actual learning system is in progress.