Hierarchical Model-Based Imitation Learning for In-between Motion Synthesis

Motion in-between problem is the key to solve the issue of long-term motion synthesis, which are widely applied in animation production, virtual reality, video games and film industry. Motion in-between can be defined as a process generating transitions between previous motion and future motion. Previous researchers focus on motion sequence generation, which omits the interaction with environment and can cause foot-sliding problem. In this work we present a novel, generic system based on imitation learning for motion in-between problem. This hierarchical system contains two parts: first part is motion primitives, which utilize reference motion to learn general knowledge of human motion; second part is policy controller, which uses multiple primitives to generate natural human motion. The primitives are controlled by the combine weights generated by policy controller. Experiment shows that our method can effectively improve motion quality and reduce the amount of reference motion data.