<p>Table tennis is a highly complex sport that demands precise and efficient techniques to achieve optimal performance. Learning these techniques from human athletes involves capturing intricate patterns and strategies that are often difficult to replicate in robotic systems. In this paper, we propose a novel imitation learning method based on Adversarial Motion Priors, named AMP-Former, to improve the efficiency and accuracy of learning table tennis techniques from human athletes. AMP-Former is built upon a Transformer architecture and is designed to capture the nuanced relationships between different shots, generate suitable kinetic rules between joints, and produce effective strategies at the game level. AMP-Former excels at capturing the temporal and spatial relationships between different table tennis shots, enabling the model to understand the flow of the game and anticipate the next move. It is also adept at producing suitable kinetic rules between joints, ensuring that the learned techniques are biomechanically sound and efficient. Additionally, AMP-Former generates effective strategies at the game level, allowing the robot to make intelligent decisions based on the current state of the match. We conducted extensive experiments on two public benchmarks: OpenTTGames and Google Table Tennis Dataset. The results demonstrate that AMP-Former outperforms other state-of-the-art methods in terms of both efficiency and accuracy. Our method achieves superior performance in learning table tennis techniques, making it a promising approach for robotic table tennis training and beyond.</p>