3D Avatar And Its Reinforcement Learning Training

It is essential for virtual reality(VR)applications to be of high reality and immersion.Users usually demand that object quality,scene appearance and physically-based simulation should be as close to the real scene as possible,and also hope that the behavior and decision of interactive objects are of highly authentic and reasonable.This paper combines 3D human body modeling and reinforcement learning techniques together to design a virtual table tennis interactive system running on an HTC VIVE platform in order to provide users with highly immersive table tennis playing experience.For a virtual table tennis interactive system,immersion and authenticity are mainly embodied by two aspects: first,the virtual scene should look very real,namely,the appearance and physically-based simulation and global illumination of the table-tennis set as well as the virtual avatar should be very close to the real ones;next,virtual avatars are able to autonomously interact with users,i.e.they can hit the ball with a reasonable batting posture and batting strategy.This paper focuses on personalized modeling of virtual players and their intelligent behavior decision-making to promote the immersion of virtual reality applications.In geometric modeling,we implemented a 3D human body scanning data fitting system which consists of three stages,namely,rigid registration,feature-point-guided deformation and non-rigid dense point registration.At the second stage,we implemented the embedding graph deformation and the SMPL-driven deformation.The latter approach makes us be able to take advantage of the SMPL parameterization model to represent personalized 3d human bodies and therefore lays foundation for making use of inverse kinematics to generate hitting ball actions for personalized virtual players.Finally,we created geometries and physical attributes of other objects such as the table tennis courts,balls,and rackets based on the Unity 3D.As for independent decision-making of the virtual player,we employ reinforcement learning to train the hitting strategy of the rackets,and then produce human body postures by combining inverse kinematics and reinforcement learning to solve the hitting action of the virtual player when the racket is hit.This makes the virtual player be able to hit the ball with a sequence of reasonable postures.Our virtual ping-pong interaction system has a good immersion owing to the real appearance of our virtual scene and a degree of intelligence of the virtual player.We have conducted detailed experimental verification and analysis on our platform.The results show that the human body modeling method proposed in this paper can faithfully restore the geometric details of the scanning model,while the virtual ping-pong player designed in this paper can accurately complete the batting task.