Spatiotemporally Integrated World Model for Physically Consistent Embodied Planning and Control

World models are a promising foundation for embodied intelligence, but existing approaches still struggle with long-horizon planning and control under partial observability, especially when physical consistency and execution robustness are required. We propose Spatiotemporally Integrated World Model (STI-WM), a unified framework for physically consistent embodied planning and control. STI-WM learns a compact latent world state that integrates spatial structure and temporal context, predicts future latent trajectories under candidate high-level skills, scores imagined futures by task progress, exploration utility, and physical consistency, and executes actions through closed-loop replanning. The model is trained with a compact objective that combines predictive modeling, physics-aware regularization, and low-level control supervision. Experiments on EMMOE and CALVIN benchmarks show that STI-WM consistently improves long-horizon embodied performance. It achieves 31.67 success rate and 53.64 task progress on EMMOE, and 4.51 average task-chain success length on CALVIN, outperforming strong prior baselines on both benchmarks. Real-world experiments and ablations further demonstrate the importance of physics-aware modeling, skill-space planning, and closed-loop replanning for robust embodied decision-making and execution.