Abstract: To address the challenges of dispersed maintenance objects, rapidly changing equipment states, and the disconnection between fault identification and field execution in new power systems, this paper defines the engineering boundary separating embodied intelligence from conventional power-sector artificial intelligence and automated inspection robots. A layered closed-loop architecture is proposed, comprising task input, task-level planning, perception and state cognition, motion planning and skill execution, physical agents, and independent safety supervision. Key enabling technologies—including multimodal perception, task planning and skill execution, precise manipulation, and safety control—are examined across three application scenarios: substation anomaly reinspection, distribution and substation equipment operation, and coordinated maintenance at renewable-energy stations. The analysis indicates that near-term deployment should prioritize abnormal reinspection and low-risk repetitive operations within structured environments, while any action that may alter equipment operating states must remain subject to both human authorization and execution-result verification. Finally, development pathways are proposed in terms of data infrastructure, cross-scenario generalization, real-time control, standardized interfaces, and graded verification.