Learning Batting Policy for a Robot Table Tennis Player Based on Support Vector Regression

A method based on support vector regression (SVR) is proposed to learn the batting policy to return the ball to a desired location for a 7-DoF (degree of freedom) anthropomorphic table tennis robot. Firstly, table tennis playing process is formalized as the batting evaluation function, which maps the state of the incoming ball and the parameters of the batting trajectory to the reward. Then, an exploration method based on the confidence region of the physical model is proposed to collect training data efficiently, and the batting evaluation function is obtained by generalizing the training data using ε -support vector regression (ε-SVR). Finally, the optimal batting trajectory is computed during decision process by maximizing the batting evaluation function using multi-start Quasi-Newton method. The proposed method is applied to a 7-DoF table tennis robot, and the results verifies its effectiveness.