Synthesizing close combat using sequential Monte Carlo

Abstract

Synthesizing competitive interactions between two avatars in a physics-based simulation remains challenging. Most previous works rely on reusing motion capture data. They also need an offline preprocessing step to either build motion graphs or perform motion analysis. On the other hand, an online motion synthesis algorithm [Hämäläinen et al. 2014] can produce physically plausible motions including balance recovery and dodge projectiles without prior data. They use a kd-tree sequential Monte Carlo sampler to optimize the joint angle trajectories. We extend their approach and propose a new objective function to create two-character animations in a close-range combat. The principles of attack and defense are designed according to fundamental theory of Chinese martial arts. Instead of following a series of fixed Kung Fu forms, our method gives 3D avatars the freedom to explore diverse movements and through pruning can finally evolve an optimal way for fighting.