Algorithm Research of Generative AI Model in Virtual Character Behavior Simulation

Abstract

The traditional simulation method of virtual character behavior has some shortcomings in fidelity and interactivity. In order to overcome these limitations, this paper adopts Conditional Generative Adversarial Networks (CGAN) as the basic model, and makes necessary improvements and optimizations. In the research process, virtual character behavior generation based on conditional variables is realized by constructing CGAN model including generator and discriminator. The experimental results show that the improved CGAN model can generate highly realistic virtual character behaviors and perform well in different evaluation dimensions. The generated virtual characters can smoothly complete various actions, such as walking, running, jumping, etc., and they are naturally coherent when changing actions. In addition, the model can also generate role behaviors that meet specific situations according to conditional variables. In quantitative evaluation, the improved CGAN model is significantly superior to the baseline model in terms of fluency, consistency and similarity to real behavior. The research in this paper not only provides a new method to improve the fidelity and interactivity of virtual character behavior simulation, but also provides a new idea for future research in virtual reality and human-computer interaction.