Visual navigation is a task of navigating a mobile robot by using a vision system fixed to the operating environment. It is necessary for a visual navigation system to detect and track the target robot, and recently some effective methods for this purpose have been proposed. However, in previous frameworks, it is assumed that the attributes (color, shape, texture, etc.) of the robot are a priori known. In this paper, we propose a method for a visual navigation system automatically learning a statistical model of the robot attributes. For the purpose, we employ an active recognition approach where the visual navigation system and the robot cooperate together. This method can evaluate how the learning was successful, therefore it becomes possible to optimize the system for individual operating environments.