Ball Trajectory Tracking and Prediction for a Ping-Pong Robot

Robot and vision systems designed for table tennis games have been a challenging and popular research topic. In this paper, we provide three indispensable procedures of a vision system for a ping-pong robot. To begin with, object recognition and tracking are preliminary in the system through image processing. In track the ball position in a three-dimensional workspace, we use cameras and adopt the binocular triangulation algorithm. In order to hit the ball precisely, the prediction of the ball position is necessary. Thus, we present a scheme for the ball trajectory prediction through Back Propagation Neural Network (BPN) without measuring the ball states of current velocity, acceleration, and self-rotational velocity. Because a ping-pong trajectory has two flight parabolas, we use two neural networks to learn the complete ping-pong flight trajectory. Namely, the first and second networks are trained for two parabolas, respectively. By doing this, the robot can only use the first ten points of the first parabola and then predict the striking point through the trained network. In our experiment, we provide an extensive validation of the proposed approach showing about 88% rate of success with flight trajectory prediction. Through refining the learning scheme, the rate of success could be up to 97%.