Manifold Learning Approach Toward Constructing State Representation for Robot Motion Generation

摘要

This paper presents a bottom-up approach to building internal representation of an autonomous robot. The robot creates its state space for planning and generating actions adaptively based on collected information of image features without pre-programmed physical model of the world. For this purpose, image-feature-based state space construction method is proposed using manifold learning approach. The visual feature is extracted from sample images by SIFT (scale invariant feature transform). SOM (Self Organizing Map) is introduced to find appropriate labels of image features throughout images with different configurations of robot. The vector of visual feature points mapped to low dimensional space express relation between the robot and its environment with LLE (locally linear embedding). The proposed method was evaluated by experiment with a humanoid robot collision classification and motion generation in an obstacle avoidance task.