Unified MLLM framework with Sparse 3D VQ-VAE (70% token reduction) for part-level mesh decomposition and kinematic chain prediction, enabling physics-based robotic simulation from monolithic assets.