Bottom-up construction of state representation for humanoid robot based on behavior of image feature

This paper presents a framework for building state representation of autonomous robot without relying on designer´s knowledge on robot itself and its environment. Based on the proposed framework, a robot with an arm can acquire knowledge of its environment by interacting with objects. The knowledge is acquired by moving its arm with a small displacement around an object and observing behaviors of feature points in image. The conditional probability densities of location of points are estimated to analyze dependency between feature points. State is discriminated by dependency networks of the feature points through clustering of networks. Clusters of networks allow the robot to discriminate whether its arm will be occluded by an object, and whether it will move together with an object. By the proposed method, the robot can build state-representation regarding interaction with its environment without any specific assumptions. The proposed method was verified by experiment of network construction and estimation of state of the environment. Networks were well constructed, classified and state representation was built.