<![CDATA[Physical AI and embodied AI systems operating in real-world environments require continuous sensor grounding from deployed infrastructure, yet the predictive world-model architectures that power these systems were developed under laboratory assumptions of clean, synchronous sensor data that rarely hold in practice. The world-model literature and the IoT sensor infrastructure literature have evolved in largely separate research communities with minimal cross-citation, leaving a critical gap in understanding how latent-dynamics models can be robustly grounded by noisy, asynchronous, and resource-constrained real-world sensor streams. This paper presents a systematic survey of 34 peer-reviewed publications organized into three streams—world-model architectures (n = 12), IoT sensor infrastructure (n = 11), and bridge literature at their intersection (n = 6), with five cross-listed papers—following PRISMA 2020 guidelines for systematic reviews. The survey identifies three novelty claim spaces where current world-model architectures remain structurally unprepared for real-world sensor deployment: robustness to missing, noisy, and irregular sensor data; clock-synchronization-aware temporal alignment between asynchronous sensor streams and world-model inference; and architectural compression strategies enabling world-model deployment on resource-constrained edge devices. This is the first systematic survey to establish a Stream A × Stream B taxonomy that maps the integration pathways between world-model architectures and IoT sensor infrastructure, providing a structured research agenda for physically grounded predictive AI.]]>
Copyrights © 2025
