From Queries to Plans: Strategic Swarm Coordination using LLMs

Swarm robotics involves coordinating multiple robots to accomplish complex tasks without a central controller. Programming such swarms for high-level missions is challenging, often requiring case-specific, expert-written logic that is time-consuming and limited in scope. Large Language Models (LLMs) have shown strong reasoning capabilities, and their application to multi-agent coordination remains a promising but underexplored direction. In scenarios such as disaster response or search and rescue, planning is critical, requiring integration of diverse assets and constraints. This work leverages LLMs to enable a single operator to coordinate swarms through natural language. Given a scenario description and asset attributes, the LLM generates a high-level, time-stepped mission plan. Falcon3-10B-Instruct was chosen as the smallest candidate and in-house solution of the Technology Innovation Institute, though it initially exhibited the weakest performance. To address this, a synthetic dataset was created by augmenting hand crafted samples and used for fine-tuning. The resulting model achieved significant improvements, consistently outperforming not only its base version but also the other candidates in inference.