A control algorithm that estimates human grasping intention is developed for driving the Power-Assist Glove that assists grasping force. In order to operate the Power-Assist Glove with appropriate drive mode, we focused on finger-joint angles in the process of grasping. There is shown to be a correlation between the three principal grasping modes and the initial movement patterns of the finger-joint angles. To distinguish the patterns and predict each mode, a pattern classification method is applied to the algorithm. The control system achieved an 80% success rate in distinguishing the grasping modes of people. The Power-Assist Glove features a simple drive mechanism using soft actuators and decreases the muscle activity that corresponds to 1.5 kg loads.