Motion Matching with Proximal Optimization Policy
Author
Najim, Yousif Ali Hassan
Term
4. term
Education
Publication year
2019
Abstract
This thesis investigates whether a motion matching system augmented with reinforcement learning, specifically Proximal Policy Optimization (PPO), can deliver scalable, high-quality real-time character animation in Unity. Building on a review of related work, the system was implemented in Unity to preprocess animation clips into poses and trajectories and to select the next pose using current and future cost terms; a PPO policy was trained to guide control decisions. The study evaluated motion-matching accuracy and runtime performance across varying numbers of agents. Results indicate that the motion matching component achieves decent accuracy, while the reinforcement learning setup provides clear benefits primarily when many agents are present. Further training and testing are needed to improve robustness and more fully characterize performance.
Dette speciale undersøger, om et motion-matching-system udvidet med forstærkningslæring, specifikt Proximal Policy Optimization (PPO), kan levere skalerbar, høj kvalitets realtidsanimation i Unity. Med udgangspunkt i en gennemgang af baggrundsarbejde blev systemet implementeret i Unity, hvor animationsklip forbehandles til poser og trajektorier, og valg af næste pose sker via omkostningsled for både nuværende og fremtidige tilstande; en PPO-politik blev trænet til at styre beslutningerne. Undersøgelsen evaluerede nøjagtigheden af motion matching samt ydeevnen på tværs af varierende antal agenter. Resultaterne viser, at motion matching opnår fornuftig nøjagtighed, mens forstærkningslæringen især giver fordele, når der er mange agenter. Der er behov for yderligere træning og test for at øge robustheden og bedre karakterisere systemets performance.
[This apstract has been generated with the help of AI directly from the project full text]