DReCon: Data-Driven Responsive Control of Physics-Based Characters

摘要

Interactive control of self-balancing, physically simulated humanoids is along standing problem in the ield of real-time character animation. Whilephysical simulation guarantees realistic interactions in the virtual world,simulated characters can appear unnatural if they perform unusual movementsin order to maintain balance. Therefore, obtaining a high level ofresponsiveness to user control, runtime performance, and diversity has oftenbeen overlooked in exchange for motion quality. Recent work in the ield ofdeep reinforcement learning has shown that training physically simulatedcharacters to follow motion capture clips can yield high quality trackingresults. We propose a two-step approach for building responsive simulatedcharacter controllers from unstructured motion capture data. First, meaningfulfeatures from the data such as movement direction, heading direction,speed, and locomotion style, are interactively speciied and drive a kinematiccharacter controller implemented using motion matching. Second, reinforcementlearning is used to train a simulated character controller that is generalenough to track the entire distribution of motion that can be generatedby the kinematic controller. Our design emphasizes responsiveness to userinput, visual quality, and low runtime cost for application in video-games.