Deep Reinforcement Learning-based Training for Virtual Table Tennis Agent

The advent of Virtual Reality (VR) has revolutionized various fields, notably by merging digital and physical realities, and has significantly influenced several fields like sports. This paper introduces a novel VR-based table tennis simulation game which employs Deep Reinforcement Learning (DRL) algorithms to enhance performance and skill development of the agents. The game features a VR table tennis scene with human models, where their movements are controlled by a neural network trained by using DRL. For the ball itself, we employ a physics engine to calculate its trajectory due to collision and gravity. The core contribution of this paper is the application of DRL algorithms to train the agent in order to simulate the behavior of a table tennis player. Comparison experiments with two algorithms validate the efficacy of the proposed method. Additionally, the game includes five human models, with varied characteristics including speed, force, and active reach, which are trained separately to explore optimal strategies for playing against diverse player profiles. Two training programs in skill mastery and skill generalization are set up to explore the optimal training strategy, which offers valuable lessons in real world for both amateurs and professional athletes. The source codes, trained networks, and video are available at https://github.com/extraordi-naryq/Master-Table-Tennis-game.