Modeling the stroke process in table tennis robot using neural network

To hit incoming balls back to a desired position, it is a key factor for table tennis robot to get racket parameters accurately. For modeling the stroke process, a novel model is built based on multiple neural networks. The input data for neural networks are the ball velocity differences during the stroke, and racket parameters are the output data. To reduce the influences from the invalid data, a neural network based on each empirical data is established. The training data are clustered based on the empirical data. The way of choosing a neural network to compute the racket parameters depends on the comparison between the new coming data and the empirical data. Moreover, a novel way based on a binocular vision system to verify the stroke model is proposed. Experimental results have showed that the stroke model created via the proposed method is applicable and the verification method is effective.