Abstract: Embodied intelligence is recognized as a critical pathway for achieving efficient integration and collaboration among humans, robots, and the physical world, with its core lying in the deep fusion of humans, robots, and the physical environment, to enhance the capabilities of agents in perception, cognition, and collaboration toward the physical world. The key components and technological pathways of embodied intelligence are systematically investigated through 5 focused dimensions: A task-oriented multimodal active perception framework is first proposed, combining embodied interaction and active navigation to establish a vision-language-behavior collaborative environmental sensing system; Dynamic task decomposition and structured planning are then enabled via world models and task symbolization techniques, ensuring the generalizable decision-making abilities of agents; A virtual-to-real migration technology chain is subsequently constructed to efficiently transfer large-scale model training results to physical hardware, bridging deployment gaps between simulation and reality; Complex task transfer and generalization capabilities of embodied agents are further enhanced using vision-language-action models and a mixture-of-experts (MoE) framework; Finally, a domestically controlled ecosystem is developed based on the China Computing Power Network to advance the localization and large-scale implementation of core technologies. The technological framework and research progress of embodied intelligence are comprehensively reviewed, offering a clear technical roadmap and practical framework for its development, thereby laying a critical foundation for realizing general artificial intelligence.