Human-motion synthesis refers to any computer-based computation used in generating motions performed by humans. It is an important field in computer graphics, with applications that include the computer and console game industry, and the development of situation simulators for training hazardous-duty personnel. It is also an essential part of imitation learning for humanoid robotics, since it is a principled tool that allows a human to teach a robot how to move as a human.
Traditional methods for motion generation of human figures are either based on purely data-driven models or purely mechanistic models. A problem with the former approach is that complex operations over large amounts of data are usually needed to generate complete new motions, whereas the problem with the later approach is that the differential equations obtained for the physical model are expensive to solve numerically. Our objective is to develop a methodology for human-motion synthesis based on physically-inspired machine learning models. Our preliminary results obtained by combining non-parametric regression with linear differential equations outperform accuracy results obtained by purely data-driven models, while keeping a competitive computational complexity. This proposal will build on and expand our preliminary findings.
The theoretical approach to the research problem is based on Latent Force Models. A Latent Force Model is a strongly mechanistic non-parametric model that combines Gaussian processes with 3 differential equations in a machine learning approach. Specifically, we will formulate our framework around the generation of motor primitives, the basic building blocks of any movement, by coupling nonparametric Gaussian process regression with ordinary differential equations. We will then combine these motor primitives in hierarchical and sequential ways to synthesis more complex movements.
To validate our framework, we will generate motions by extrapolating from well-known motion capture databases. Also, we will test the ability of the proposed models to identify motor primitives and to generate movements in a robotics system intended for humanoid robotics. Our methodology will allow us to synthesis persuasive human motions at a reasonable computational complexity, ultimately providing an alternative for motion generation of human figures to be used in computer animation and imitation learning in humanoid robotics.
