In this paper, the kick of a ball to the virtual target point (VTP) for humanoid robots (HRs) by two visual navigation strategies is compared. The VTP is not necessary to be known all the time but it must be estimated on line. At the beginning, a neural-network-based vision system (NNBVS) is designed by the connection of four visual windows relative to four pitch angles of HR so that the transform between the image plane and the world coordinates is established for the visual navigation. The first strategy is assumed that the VTP is realized all the time. Based on the derived geometric relation, the corresponding design is established to execute the assigned task. As an HR is far away from the gate (e.g., larger than 1.8m), two columns of the gate are not recognized by the proposed NNBVS. In addition, the occlusion of an HR by other HRs often occurs for the soccer game. Then the VTP is not detected all the time. Therefore, another strategy is developed to deal with these circumstances. The condition for the second strategy is that the VTP is detectable as the distance between HR and gate is smaller than the recognizable distance. Finally, four representative experiments for the two navigation strategies confirm that each strategy possesses its own advantages and disadvantages.Index Terms-Humanoid robot, machine vision for localization, modeling using multilayer neural network, posture revision, strategy for visual navigation.