Physical AI: The Convergence of Foundation Models, Sim-to-Real Transfer, and Embodied Intelligence

The frontier of artificial intelligence research has shifted dramatically from digital domains toward Physical AI - systems capable of intelligently interacting with the physical world through embodied agents. This paper presents a comprehensive technical survey of Physical AI, analyzing the mathematical foundations, key methodologies, and emerging industrial implementations. We provide rigorous mathematical treatment of the POMDP formulation underlying robot learning, compare the Diffusion Policy approach with the Flow Matching technique, and demonstrate how the latter achieves 10× faster inference (100+ Hz) through deterministic ODE solvers. We examine landmark systems including Physical Intelligence’s π0 (robot-agnostic generalist policy), and GenSim (text-to-simulation pipeline), alongside hardware acceleration through NVIDIA’s Jetson Thor and Cosmos World Model. Additionally, we address critical challenges including Real-to-Sim-to-Real transfer via language-guided parameter optimization, and data scarcity mitigation. This paper identifies open research directions for the next decade of embodied AI deployment.