Autonomous intelligent unmanned systems operating in real−world open environments—characterized by dynamic complexity, multi−agent coupling, incomplete information, and strong social constraints—face critical challenges such as insufficient compliance modeling, limited social risk perception, complex collaborative conflicts, and delayed abnormal response. To address these issues, this paper proposes an Embodied Social Perception Intelligence Framework, which integrates embodied perception (including proprioceptive, internal, exteroceptive, interactive, and intention perception) with social radar, and introduces Agentic AI as a top−level decision−making and control mechanism to achieve multi−level and autonomous cognitive decision−making. The framework adopts a five−layer architecture—perception, reasoning, execution, feedback, and meta−control—establishing a dynamic closed loop from multimodal perception to compliant behavior generation. By fusing physical and social environmental information, the proposed framework significantly enhances the task adaptability, collective coordination efficiency, and compliance reliability of autonomous intelligent unmanned systems in complex and uncertain scenarios such as urban governance, emergency rescue, and social security. This work provides a new technical pathway toward trustworthy, explainable, and sustainable autonomous intelligent systems.

Embodied intelligence represents a new stage in the evolution of artificial intelligence, marking a transition from "perception−cognition" to an integrated paradigm of "perception−cognition−action." The Vision−Language−Action (VLA) model provides a critical technological pathway for enabling autonomous agent operation in the real world by unifying visual perception, language understanding, and action generation. This paper systematically reviews the development trajectory and representative achievements of VLA technologies, and summarizes their architectural paradigm, which includes multi−modal perception, semantic fusion mechanisms, reinforcement and imitation learning, world models, and hierarchical action output. By considering application scenarios such as autonomous driving, human–computer interaction, and industrial equipment, we further analyze the core challenges faced by VLA development, including the scarcity of data resources, limited generalization and transferability, insufficient interpretability, and increasing computational demands, and we outline the future development trends.

With the convergence and innovation of emerging information technologies, Digital Twin (DT) technology has become a key enabler for digital transformation and the evolution of intelligent systems. It has been widely applied and received significant attention in fields such as industrial manufacturing, smart cities, and intelligent transportation. However, existing research on traditional DT technologies has predominantly focused on the modeling and analysis of physical entities ("objects"), with limited systematic integration of "human" and "environmental" factors. The lack of exploration into human−environment interactions makes it difficult for current digital twin frameworks to meet the advanced demands of complex intelligent systems for multi−level, comprehensive interaction capabilities. In view of this, this paper innovatively introduces the "object−human−environment" interactive vision and comprehensively and systematically analyzes the research frontiers and progress of digital twin technology from the three core dimensions of intelligent physical entity (object), intelligent individual (human), and virtual−real fusion environment (environment). Firstly, the paper analyzes the traditional digital twin technology system with "object" as the core and focuses on its theoretical origin, framework, and application. Secondly, it discusses the definition, development context, national policies, and core technologies of digital people driven by AI. Finally, expand the vision to the dimension of "environment" and explore the application practice of "environment" in multiple scenes of the meta−universe, deeply discuss the deep integration and interaction mechanism of the three elements of "object", "human" and "environment", reveal how the three interact and promote each other, and provide support for the construction of the meta−universe. Furthermore, this study discusses the current research challenges and future development trends of digital twin technology from the perspective of "Object−Human−Environment" interaction and proposes three key research directions: (1) developing an intelligent, multi−layered data fusion framework; (2) exploring AIGC−enabled intelligent virtual−real mapping and native virtual evolution; and (3) constructing novel virtual economy architectures and intelligent governance systems. The research outcomes provide both theoretical foundations and practical insights for building next−generation digital twin systems characterized by multi−agent collaborative perception, multimodal intelligent interaction, and closed−loop integration of virtual and real environments.

Humanoid robots, benefiting from their human−like morphology and locomotion capability, are regarded as promising platforms for future service, rescue, and industrial applications; however, achieving stable and reliable walking in unstructured environments remains highly challenging. This paper provides a comprehensive review of recent advances in humanoid locomotion planning and control, with a focus on gait planning, trajectory generation, whole−body control, and learning−driven approaches. We summarize the core concepts and implementation frameworks of representative methods, compare their applicable scenarios, strengths, and limitations, and present a hierarchical categorization of existing research. Moreover, this work discusses key technical bottlenecks that hinder environmental adaptability and dynamic stability. Finally, we outline future research directions, including multimodal perception integration, co−optimization of learning and control, whole−body motion skill learning, and safety assurance, and offer suggestions toward standardization and large−scale deployment.

The intelligence of unmanned vessel systems is undergoing a profound transformation from remote control to embodied autonomous forms, with the core being the realization of advanced intelligent behavior through multimodal perception, environmental interaction, and closed−loop learning. This paper systematically reviews the key advances of embodied intelligence in unmanned vessels, highlighting that semantic control loops, digital twin validation, and evaluation systems are moving from methodological exploration to engineering integration, and have already begun to provide preliminary application support in port and inland waterway scenarios. However, current technologies still face bottlenecks in perception stability, rule interpretability, and deployment resources. Therefore, this paper recommends focusing on breakthroughs in strengthening autonomous closed−loop intelligence systems, establishing standard and trustworthy validation environments, and promoting lightweight and collaborative deployment to enhance system reliability, compliance, and scalability, providing support for the development of intelligent ship technology and the implementation of our marine strategy.