Keyframe animation of virtual humans via motion data learning

Abstract

SUMMARYThis paper proposes a method for creating human animations by utilizing motion capture data in traditional keyframe-based motion editing. This automatically generates complicated sequential movements of the whole body from multiple constraints of end-effectors. Different types of motion learning are introduced to optimize postures at each keyframe and motion curves for interpolating them, separately. Postures at every keyframe are determined by using hierarchical reinforcement learning so as to have similar features to motion samples with a data-centric objective function. Plausible postures are efficiently sought among the huge number of possible states because a skeletal structure is hierarchically decomposed and postures are efficiently quantized to narrow down the configuration space. Optimized postures are then interpolated using motion curves that are learned with acceleration templates from referential motion segments. This new technique of reusing motion data is well suited to design motions by manipulating end-effectors at each keyframe.