TARA-LLM: A Knowledge-Driven Model for Task Assignment of Multiplayer Reach-Avoid Games

Large language models (LLMs) have shown great potential in decision-making and task planning due to broad training data and enhanced reasoning capacities. This article applies LLMs for multiplayer reach-avoid (M-RA) differential games, an adversarial scenario where a team of pursuers aims to intercept a team of evaders before the latter reaches a predefined goal region. The goal is to propose a knowledge-driven LLM called TARA-LLM for the task assignment of M-RA games, which consists of three modules to maximize the number of evaders captured. The model begins with the planning module to produce an initial matching by knowledge-driven prompts that incorporate analytical winning conditions, coalition reduction, and constraints. The revision module then addresses hallucinations of LLMs by correcting improper matchings and yielding a final matching. Finally, the implementation module generates a task assignment, including targeted evaders and explicit pursuit strategies. The TARA-LLM outperforms common prompt engineering and achieves better performance than the analytical baseline in certain scenarios in terms of time efficiency and generalization as the number of players increases. Numerical simulations and experiments over autonomous aerial vehicles validate the feasibility of the TARA-LLM.