Research On Optimization Of Character Animation Locomotion Controller Based On Neural Network

Designing locomotion controller for virtual characters is a challenging task.This paper uses deep neural network,as the locomotion controller to control the motion of virtual characters in OpenAI Gym toolkit.The existing method of using neural network as virtual character locomotion controller only focus on the character forward(robustness)and ignores the detailed body coordination,and this often creates awkward movements and reduces the naturalness and style of the synthetic motion.The environment of OpenAI Gym also lacks the motion simulation of virtual characters in obstacle terrain.We propose Space-time Constrained Optimization for Deep Locomotion Controller,to optimize the motion of virtual characters by integrating space-time constraints,as part of the reward function,during the learning process.We propose Motion Data based Neural Network as Locomotion Controller.We utilize the prior information of motion data to guide the learning of deep neural network as the locomotion controller of virtual characters.We propose to start from a simple reward function,and gradually higher requirements are imposed on gait patterns and the learning process is accomplished by a progressive reward function.We also propose the Multi-Critic Model to evaluate individual reward component,thus allowing dynamic adjustment to the reward function.We propose Terrain Information based Virtual Character Locomotion Controller in OpenAI Gym Environment.We construct complex terrain and place virtual character in complex terrain environment.Through sampling terrain information,virtual character learn to move forward in complex environment.The experiment results confirm that the introduction of the space-time constraints avoids the problem of generating awkward gait,and generate naturalness gait patterns.The results show that the integration of motion data not only ensures the consistency between the synthetic and original motions but also accelerates the learning of network.We demonstrate the application of our method to a variety of virtual characters(cheetah,hopper,2D walker,and 3D humanoid)performing various tasks(walking,running,jumping,and traversing at different velocities and across uneven terrains).The learning process based on complex terrain information enables 3D Humanoid to move forward in complex virtual environments.