Physical AI denotes artificial systems that, under real‑world constraints, continuously perceive, model, and act upon the physical environment, forming a closed perception‑prediction‑action loop. It differs from conventional digital AI, embodied AI, cyber‑physical systems, and digital twins in goals, interaction modalities, and implementation strategies.\ \ At the foundational level, Physical AI spans physics, control theory, cognition, machine learning, and reasoning, establishing a unified theoretical stack. The perception tier stresses multimodal sensing and physical data acquisition; the modeling tier focuses on world models and simulation; the decision tier covers motor control, policy learning, and multi‑agent coordination; the execution tier demands hardware‑software co‑design for real‑time and reliable operation.\ \ Core techniques comprise:\
- Multimodal perception and physical data fusion;\
- Simulation‑based world model construction;\
- Embodied foundation models with self‑supervised learning;\
- Motor‑control systems integrated with reinforcement learning;\
- Multi‑agent coordination and distributed decision making;\
- Hardware‑software co‑design for efficient real‑time execution.\ \ Representative applications include robotics, intelligent manufacturing, infrastructure monitoring, scientific experimentation platforms, and autonomous service systems. Deploying these applications requires addressing sim‑to‑real transfer, system reliability, real‑time execution, evaluation metrics, safety governance, and long‑term deployment challenges.\ \ By linking perception, prediction, decision, action, and adaptation in a closed loop, Physical AI is progressing from narrow automation toward more general physical artificial general intelligence (Physical AGI). Future research directions involve enhancing cross‑modal understanding, building interpretable physical world models, scaling multi‑agent collaboration, and establishing unified safety and evaluation frameworks.\ \ Review