Training a Virtual Tabletennis Player Based on Reinforcement Learning

It is essential for virtual reality(VR)applications to be of high reality and immersion while the intelligence and rationality of behaviors taken by virtual agents in the virtual scene can significantly improve the authenticity and immersion of VR applications.We employ reinforcement learning to train the hitting ball strategy of rackets and design a set of rewarding functions under the guidance of table tennis rules in order to generate a rational racket trajectory with starting position and initial velocity of the ball given.We further bind the racket to the hand of a virtual player and then solve the hitting action of the player by combining inverse kinematics and reinforcement learning.This makes the virtual player be able to hit the ball with a sequence of reasonable postures.Careful ablation analysis was conducted to show the necessity and effectiveness of our rewarding policies and testing experiments demonstrate that our approach can successfully hit the ball with more than 93%accuracy,which is comparable to that of the imitation learning based method and higher than those of other methods.However,compared to imitation learning,reinforcement learning is less expensive due to using random generated training data.