Robot behavior learning with adaptive categorization in logical perceptual space

In this study, we adopt an existing learning framework to conduct robot behavior learning in two aspects, namely, behavior encoding, which is the learning of the sensorimotor mappings of individual primitive behaviors, and behavior coordination learning, which is the learning of the mapping of contribution of primitive behaviors on resultant robot action and sensor data. Since the two mentioned learning tasks usually involve high dimensional, nonlinear and discontinuous mappings, we propose to simplify the learning tasks by forming category mappings to approximate the sensorimotor and sensori-weight mappings. As the nature of the mappings are similar, we also suggest to perform learning of the mappings with the same learning architecture, like the Behavior Learning/Operating Modular (BLOM) Architecture, proposed in this thesis.; There are difficulties in online robot behavior learning, namely, (1) exponential memory increases with time, (2) difficulty for operators to specify learning tasks accuracy and control learning attention in priori, and (3) high dimensionality of perceptual space. Two strategies are proposed in this thesis to remedy the difficulties in behavior learning. In order to solve difficulties (1) and (2), the first strategy is to incorporate adaptive categorization in ART-type networks for perceptual and action patterns categorization. An other strategy is suggested to remedy difficulty (3) by employing logical perceptual space in robot behavior learning, instead of the high dimensional physical perceptual space. The logical perceptual space retains the shared information acquired from different sensor sources of the robot that are crucial to the behavioral control of robot action.; A Tele-Assisted Teaching System (TATS) is developed for data acquisition and behavior teaching for robot behavior learning. Our experiments demonstrate that the robot can exhibit the learned exploring and wall-following behaviors satisfactorily in novel environment under the guidance of the trained BLOM architectures with physical or logical perceptual spaces. (Abstract shortened by UMI.)...