Abstract:
Aiming at the problems of incomplete data, ambiguous instructions and dynamic state changes in unstructured railway dispatching scenarios, this paper proposed an embodied dispatching and decision-making paradigm for large language model (LLM) multi-agent clusters. This work constructed a three-layer architecture consisting of a physical perception layer, a logical cognition layer and a collaborative cognition layer. Embodied encoding represented asset states and constraint conditions, and role-based agents conducted semantic negotiation. The rights-responsibility-boundary rule, dual-brain verification mechanism and cognitive sandbox constrained the generation and execution of schemes. This work performed scene deduction and qualitative comparison with marshalling yard shunting operations under extreme weather conditions as an example, and realized a decision-making loop covering connection intention restoration, multi-agent negotiation and safety verification. The paradigm offered potential advantages in incomplete information processing, decision explainability and human intervention. Its dispatching efficiency and reliability remain to be empirically verified, and it provides an architectural reference for cognition-driven shunting operations and emergency response.