Human Motion Generation Using Variational Recurrent Neural Network

∗ Human motion control, edit, and synthesis are important tasks to create 3D computer graphics video games or movies, because some characters act like humans in most of them. The purpose of this study is to construct a system which can generate various natural character motions. In this study, we consider that the process of human motion generation is complicated and non-linear, and it can be modeled by deep neural network. Since the motion generation process (deep neural network parameters) cannot be observed di-rectly, it needs to be estimated by learning from observable human motion data recorded by motion capture system. On the other hand, the process of inference which is opposite to the generation is also expressed by deep neural network. And inference and generation are performed for human motion data, and the parameters of the both deep neural networks are optimized based on the criteria that the original motion should be obtained through inference and generation processes. In this study, we constructed a human motion generative model using recurrent neural network and variational autoencoders, and confirmed that various human motions can be generated from a low-dimensional latent space.