Ball Tracking and Trajectory Prediction for Table-Tennis Robots.

Area of Science:

• Robotics and Artificial Intelligence
• Sports Technology
• Machine Learning Applications

Background:

• Table tennis robots require accurate ball tracking and trajectory prediction.
• Existing methods include physical models and machine learning.
• Physical models often need high-frequency imaging for accurate force estimation.

Purpose of the Study:

• To develop a machine learning approach for predicting table tennis ball trajectories.
• To investigate the effectiveness of a dual artificial neural network architecture.
• To compare the proposed method against single-network models and traditional physical models.

Main Methods:

• Established two artificial neural networks to learn two distinct parabolic trajectories of the ball.
• Trained the dual-network model using real-world and physically generated trajectory data.
• Evaluated prediction accuracy using mean error and success rate in simulation and physical experiments.

Main Results:

• The dual-network method achieved a lower mean error (39.6 mm) compared to a single-network model (42.9 mm) in simulations.
• In physical experiments, the proposed method showed a mean error of 36.6 mm with a standard deviation of 18.8 mm.
• The machine learning model demonstrated a 97% success rate, surpassing the physical model's 70% success rate, with fewer parameters and shorter training times.

Conclusions:

• The proposed dual artificial neural network method significantly improves table tennis ball trajectory prediction accuracy.
• This machine learning approach is robust, maintaining performance without time-stamp data.
• The method offers practical advantages including reduced network complexity and faster training, making it suitable for real-world robotic applications.