Finding curvilinear path features in a layered learning paradigm for humanoid robot using monocular vision

Abstract

In this paper, a method to find curvilinear path features is proposed. These features are defined as centers and radiuses of circles that best fit to the curvature parts of the curvilinear path. In our previous research, we proposed a hierarchical layered paradigm for humanoid robot to learn how to walk in the curvilinear path. This model consists of four layers and each one has a specific purpose and is responsible to provide some feedbacks for the lower layer. In this study, we focus on the first layer which is high level decision unit responsible to provide some feedbacks and parameters for the lower layer using robot sensory inputs. The ultimate goal is that robot learn to walk in the curvilinear path and to reach this goal, the first step is to find robot position in the environment. In this work, Monte Carlo localization method is used for robot localization. Then we used artificial potential field to generate a path between robot and a goal. Finally, we proposed an algorithm that search the circles that best fit to the curvature parts of the path. Finding these features would help the learning process for lower layers in the learning model. We used robot camera as the only sensor to identify landmarks and obstacles for robot localization, path planning and finding curvilinear path features. © World Scientific Publishing Company.