• 강사 : 김태균 교수 (KAIST 전산학부)
• 일시 : 2026. 4. 20. (월) 16:00~17:30

Recent breakthroughs in large language models (LLMs) and generative AI have demonstrated remarkable capabilities in text, image, and video synthesis. However, despite their scale and fluency, these models remain fundamentally limited in their understanding of 3D space, physical interaction, and embodied reasoning. This talk explores the next frontier beyond LLMs: Spatial Physical Generative AI — systems that not only generate content but understand and reason about the physical world. We begin by examining the evolution from Transformers and large-scale pretraining toward vision-language and world models that attempt grounded intelligence. While current generative video and multimodal models achieve impressive visual realism, they often lack true spatial consistency, physical plausibility, and compositional generalization. Addressing these limitations requires integrating 3D scene representations, physics-based simulation, and generative diffusion frameworks. The talk presents recent advances in 3D spatial AI, motion diffusion, physics-guided video generation, and category-level 6D pose estimation. We introduce methods that combine regression and diffusion modeling with score scaling sampling to capture multi-hypothesis pose distributions efficiently. Furthermore, we highlight MPMAvatar, a hybrid mesh–3D Gaussian Splatting framework that enables physically accurate cloth simulation and photorealistic avatar rendering, demonstrating realistic deformation, collision handling, and zero-shot scene interaction. By unifying generative models with spatial representation and physical simulation, we move toward AI systems capable of embodied reasoning, real-world interaction, and scalable physical intelligence. Spatial Physical Gen AI, also called world models, represents a critical step toward grounded artificial general intelligence (AGI), bridging perception, generation, and physical understanding.