Aiming to implement a more user-friendly powered prosthetic hand, this paper studies the real-time classification of intended hand motions using tactile sensors attached to the skin surfaces of a forearm. The proposed system used only two tactile sensors made of PVDF (polyvinylidene fluoride). Machine learning was applied to the classification of hand motion intentions using the tactile feature patterns. In this paper, we further studied the real-time motion classification methods in an online environment. We found that the average classification accuracy for the 6 types of motion in 8 experimental participants was 83.3 %. Participants reported no perceptible delay, which was also confirmed through video analysis. In conclusion, we showed real-time motion classification is possible by using two PVDF tactile sensors with simple training lasting for only several minutes.