The present paper develops a training-assist robotic system that can adapt a reference hand motion for a virtual tennis task to individual motor abilities. The system first measures maximum hand force and velocity in reaching arm movements prior to the training, and designs a suitable reference trajectory for teaching motion smoothness and timing in the virtual tennis task based on the measured results. A quantitative index for evaluating task performance and motor functions are then defined with consideration of task dynamics. Finally, the effectiveness of the developed training system is validated through a set of preliminary training experiments with health subjects.