Advanced articulated motion prediction

Motion synthesis using machine learning has seen rapid advancements in recent years. Unlike traditional animation methods, utilizing deep learning to generate human movement offers the unique advantage of producing slight variations between motions, similar to the natural variability observed in real examples. While several motion synthesis methods have achieved remarkable success in generating highly varied and probabilistic animations, controlling the synthesized animation in real-time while retaining stochastic elements remains a serious challenge. The main purpose of this work is to develop a Conditional Generative Adversarial Network to generate real-time controlled motion that balances realism and stochastic variability. To achieve this, three novel Generative Adversarial models were developed. The models differ in the architecture of their generators that utilize: a Mixture-of-Experts method, a Latent-Modulated Noise Injection technique, and a Transformer-based architecture respectively. We consider the latter to be the main contribution of this work, and we evaluate our method by comparing it to the other models on both stylized locomotion data and complex, aperiodic dance sequences, assessing its ability to generate diverse, realistic motions, being able to mix between different styles while responding to motion control. Our findings highlight the trade-offs between motion quality, variety and motion generalization in real-time synthesis by comparing by exploring the advantages and disadvantages of each architecture, contributing to the ongoing development of more flexible and varied animation techniques.