Sensor-Driven World Models for Embodied Intelligence: A Survey of Sensor–State–Decision Modeling Across Far-Field and Near-Field Regimes

World models are becoming a crucial foundation for embodied intelligence, which transforms sensor observations into internal state representations that support prediction, planning, and control. However, existing research is scattered across different fields (autonomous driving, robotic manipulation, generative interactive environments, and physical intelligence), and lacks a description of the role of sensors in defining the observable world, shaping state representations, and influencing decision-making. This paper proposes a unified Sensor–State–Decision (SSD) framework to describe sensor-driven world models. Specifically, the Sensor-Layer acquires raw physical observation data, the State-Layer transforms it into a structured world state representation, and the Decision-Layer outputs the final executable decision. Furthermore, it proposes that the far-field and near-field mechanisms describe the relative positions between complete SSD instances in the SSD hierarchy. That is, higher layers constrain lower layers through goal injection, and lower layers correct higher layers through result feedback. Based on this framework, we synthesize representative architectures, applications, datasets, and benchmarks, particularly long-tail scenarios in the Far-Field and Near-Field domains. Finally, we summarize the current challenges. The SSD perspective promises to provide a new structured analytical framework for designing robust, physically based, and deployable embodied world models in the real world.