Directable Motion Texture Synthesis

We present an algorithm for creating animation by keyframing, stylistic sequencing, random sampling, or by a combination of these methods. We first “learn” a motion model from existing animation data (such as motion capture), and optionally label specific frames as corresponding to specific “styles.” The system can then generate a new motion in the style of the training data, but matching the pose constraints and/or style constraints specified by an animator. Realistic motion in animation is difficult to synthesize—it often requires either specific physical calculations based on the character being animated, or it must be hand-generated or captured. Animations are used, however, in a number of applications that would benefit from a more general approach to creating realistic motion, e.g. in video games that require constraints over terrain, keyframes, and general style, but have no restrictions on specific paths, locations, or characters. We propose a method for motion synthesis that, given a character’s hierarchical skeleton and bone lengths, will generate realistic motion similar, but not limited to, the content of sample motions. The new motion may also be constrained by keyframed location/orientation values and stylistic content.