To address the challenge of unclear “ground-environment-machine-structure” interaction mechanisms in shield tunneling under complex urban environments, this study, inspired by the concept of embodied intelligence, proposes a quasi-online closed-loop control architecture comprising perception, learning, decision-making, and execution layers. In the perception layer, a wireless sensor network is utilized to collect multi-source heterogeneous data in real time, which is combined with the K-means unsupervised clustering algorithm to achieve the autonomous identification of stratum characteristics and structural response patterns in the absence of prior labels. In the learning layer, a physics-inspired world model based on a “data-physics dual-driven” approach is constructed. By adopting the physics-informed machine learning method, a three-dimensional refined numerical simulation surrogate model is embedded into the neural network training process as physical constraints. Fused with insitu measured data, a world model for the shield tunneling process is obtained, characterized by high accuracy, fast response, strong generalization, and forward deduction capability. In the decision-making layer, the pseudo-gradient estimation method in model-free adaptive control is improved based on the forward-simulation feedback of the world model, dynamically optimizing the shield tunneling parameters according to environmental states. Engineering application results demonstrate that this framework exhibits excellent adaptability during the actual construction of new three-line tunnels undercrossing existing three-line tunnels. The system can continuously and accurately identify stratum variations and deformation characteristics of existing structures. Furthermore, through implicit representation, it deeply mines the high-dimensional mapping relationship between the shield machine’s construction parameters and the additional stress of the segmental structure. The adaptive control method, improved by the physics-informed model, breaks through the traditional offline prediction paradigm, achieving precise dynamic closedloop control over the shield machine’s posture and the responses of existing structures.