Ball Trajectory Tracking and Prediction for a Ping-Pong Robot

• Main Authors: HUANG, YI-CHEN, 黃奕晨
• Other Authors: LIN, HSIEN-I
• Format: Others
• Language: zh-TW
• Published: 2019
• Online Access: http://ndltd.ncl.edu.tw/handle/d84ckx

碩士 === 國立臺北科技大學 === 自動化科技研究所 === 107 === The robotic technology is developing greatly. Especially, sport robots gradually become a popular and interesting research topic, and one of them is the ping-pong robot. In the ping-pong robot, object trajectory tracking and prediction are important techniques in this research. To start with, the robot needs to recognize and locate the ball in the 3-D space. We detail the procedure how to carry out in this thesis. Ball trajectory prediction is another important issue. There are already several research literature about predicting methods, and that are roughly divided into two leading methods：analytic flying model and machine learning. The former builds the model considering gravity, air resistance, Magnus effect, and elastic collision, but the estimation of the ball state has to rely on high-performance vision devices. The latter is the method adopted in this thesis. The normal trajectory of ball is composed of two parabolas：Starting from serving to the table on the same side, after bouncing from the table on the other side to the robot side. Therefore, the proposed method implements two back propagation neural networks (BPNs) that learn the two parabolas respectively, and the robot system use the trained models to predict the ball trajectory. When the robot detects the few positions of the initial trajectory, the striking position can be determined by the BPN models. The advantage of the proposed method is that the requirement for costly vision device can be avoidable. The simulation results show that there was up to 97% success rate, and the average position error was about 26.313 mm. In the experiment, the robotic arm moved toward the striking location with 36.6 mm in average error even though the vision-based 3-D position surely had biases.