월드모델을 통한 Latent Diffusion Model 예측 고도화

초록

Human motion prediction is essential for applications ranging from autonomous systems to immersive virtual environments. However, most existing models focus on everyday motion, limiting their effectiveness in dynamic, expressive contexts like sports, VR, and assistive robotics. To address this, we target the LaFAN dataset, which captures high-energy movements with significant intra-class variation. We propose a novel framework combining three ideas: (1) a latent diffusion model for improved temporal consistency, (2) a reinforcement learning approach for stable long-term prediction, and (3) a robust noise optimization strategy to counter diffusion sampling noise. Experiments show our model produces realistic, coherent motion sequences, highlighting its potential for complex and expressive motion prediction.

키워드