Learning Method for Multi-Controller of Robot Behavior

Abstract

In this paper, we propose a hierarchical behavior controller and a learning algorithm for the behavior controller which consists of several subcontrollers to indicate the desired trajectories for robot actuators.This algorithm selects the subcontroller which is not appropriate and needs to be tuned, by evaluating each subcontroller using multiple regression analysis based on previously obtained evaluation value.This process can reduce the learning iterations by avoiding attempts to tune good subcontrollers.The proposed algorithm is applied to the problem of selecting and tuning subcontrollers at the middle layer in the hierarchical behavior controller in order to compensate imperfect initial controllers.The hierarchical behavior controller is applied to the problem of controlling a seven-link brachiation robot, that moves dynamically from branch to branch like a gibbon, a long-armed ape, swinging its body like a pendulum.