This paper gives a team description of Osaka University “Trackies” for RoboCup-98, and related research issues. We focus on behavior learning of our goalie robot which has an omnidirectional vision system. A Q-learning method is applied by defining substates from visual information of the ball and the goal. To reduce the learning time, we propose an attention control method for an omnidirectional vision by means of an active zoom mechanism. We perform computer simulation and real robot experiments to show the validity of the proposed method.