Multimodal embodied large models: unified perception-cognition-action architecture and applications

With the rapid development of artificial intelligence technology, embodied intelligence has emerged as one of the key pathways towards achieving artificial general intelligence. This paper provides an in-depth exploration of the design principles, technical challenges, and future directions of the unified perception-cognition-action architecture in multimodal embodied large models. We analyze the limitations of traditional modular architectures and propose a new architectural paradigm based on deep integration of perception, cognition, and action. This architecture achieves organic unification of perceptual information, cognitive reasoning, and executive control through end-to-end learning, providing a theoretical foundation for autonomous agent behavior in complex physical environments. This paper systematically examines the core technical elements of the perception-cognition-action architecture, including multimodal representation learning, cross-modal attention mechanisms, world model construction, and hierarchical decision-making frameworks. We also explore the application prospects of this architecture in intelligent manufacturing, service robotics, autonomous driving, and other domains. Finally, we analyze current technical challenges and provide perspectives on future development trends of embodied large models, offering new insights for advancing embodied intelligence research.