A Large Language Model-Driven Heterogeneous Air-Ground Search Swarm

In this paper, we propose a large language model(LLM)-driven task allocation and coordination framework in heterogeneous air-ground collaborative system, specifically applied to open-world emergency search scenarios. However, the increase in the number of agents and their heterogeneity increase the probability of errors in LLM reasoning. Secondly, the inherent latency of LLM inference compromises task allocation efficiency. To address these challenges, we present a centralized planning with asynchronous reasoning-execution (CPARE) framework. In this process, we adopt an element-based relevance matching mechanism to generate task and a time-efficient method based on weighted graph matching to complete task allocation.We validated the algorithm’s efficiency in a simulated environment and further built a physical system that integrated LLM with the air-ground platform, marking the first successful deployment in real-world search tasks.