Learning generative models for stylistic whole-body motions and its online adaptation

在这个研究项目中，我们开发了一种新的算法，用于学习全身人体运动的生成模型，可以显式地捕获运动序列的风格，并从非平稳的未标记的序列观察中估计模型的状态和风格变量。所开发的算法的适用性和有效性，通过几个应用程序，如实时预测的运动捕捉数据，模仿学习类人机器人的运动技能，学习肌电机器人接口。

In this research project, we developed novel algorithms for learning generative models of a whole-body human motions, that can explicitly capture the style of the motion sequence, and for estimating both state and style variables of the model from non-stationary unlabeled sequential observations. The applicability and effectiveness of the developed algorithms were demonstrated through several applications such as real-time prediction of motion capture data, imitation learning of motor skills by humanoid robots, and learning myoelectric-robot interfaces.