SwarmBrain: Embodied agent for real-time strategy game StarCraft II via large language models

Abstract: LLMs have recently garnered significant accomplishments in various exploratory tasks, even surpassing the performance of traditional reinforcement learning-based methods that have historically dominated the agent-based field. The purpose of this paper is to investigate the efficacy of LLMs in executing real-time strategy war tasks within the StarCraft II gaming environment. In this paper, we introduce SwarmBrain, an embodied agent leveraging LLM for real-time strategy implementation in the StarCraft II game environment. The SwarmBrain comprises two key components: 1) a Overmind Intelligence Matrix, powered by state-of-the-art LLMs, is designed to orchestrate macro-level strategies from a high-level perspective. This matrix emulates the overarching consciousness of the Zerg intelligence brain, synthesizing strategic foresight with the aim of allocating resources, directing expansion, and coordinating multi-pronged assaults. 2) a Swarm ReflexNet, which is agile counterpart to the calculated deliberation of the Overmind Intelligence Matrix. Due to the inherent latency in LLM reasoning, the Swarm ReflexNet employs a condition-response state machine framework, enabling expedited tactical responses for fundamental Zerg unit maneuvers. In the experimental setup, SwarmBrain is in control of the Zerg race in confrontation with an Computer-controlled Terran adversary. Experimental results show the capacity of SwarmBrain to conduct economic augmentation, territorial expansion, and tactical formulation, and it shows the SwarmBrain is capable of achieving victory against Computer players set at different difficulty levels.

• The paper introduces SwarmBrain, an LLM-powered agent combining the Overmind Intelligence Matrix for strategic planning and the Swarm ReflexNet for real-time tactical responses.
• It demonstrates robust performance with a 100% win rate on lower difficulty settings and 76% on Hard, effectively managing economic and tactical gameplay.
• The study identifies challenges such as visual information gaps and LLM response latency, outlining future directions for enhanced strategic AI in real-time strategy games.

Introduction

LLMs have been forging new paths in AI, demonstrating transformative progress in tasks particularly suited to exploration. Notably, their success has overtaken that of reinforcement learning (RL)-based agents in specific arenas. This paper explores the functionality of LLMs, specifically within the real-time strategy (RTS) game environment of StarCraft II. The research introduces 'SwarmBrain,' an agent utilizing LLM for real-time strategic deployments, most notably tested within the Zerg faction. Two integral components of SwarmBrain are detailed: the Overmind Intelligence Matrix, which handles macro-level strategy, and the Swarm ReflexNet, designed for more reactive and immediate tactical decisions.

Overmind Intelligence Matrix and Swarm ReflexNet

The Overmind Intelligence Matrix simulates the strategic overmind of the Zerg faction, synthesizing a comprehensive view of the battlefield to direct resources, expansions, and tactical offensives. This macro-strategic module ensures a high-level orchestration of the game strategy. Its counterpart, the Swarm ReflexNet, compensates for the latency issues inherent in LLM reasoning by employing a swift, condition-response state machine framework that allows Zerg units to execute basic maneuvers autonomously. The SwarmBrain agent, tasked with leading the Zerg against computer-controlled Terrans, displayed prowess in managing economic growth, expansion, and tactical engagements across different difficulty levels, with a remarkable 100% win rate against lower difficulty settings and a substantial 76% win rate at the Hard level.

Experimental Results

The experiments conducted offer quantitative insights into SwarmBrain's performance. In the control setup, spanning 30 matches on the "Automaton LE" map, SwarmBrain managed an unyielding victory rate against Computer opponents set to Very Easy through Medium Hard levels. The agent upheld a notable win percentage even against the Hard level. This performance showcases the robust capabilities of SwarmBrain in leveraging economic, territorial, and tactical strengths to secure victories.

Discussion and Conclusion

Despite the achievements of SwarmBrain in handling the complexities of StarCraft II, challenges remain. Notably, the visual information deficiency within the agent and the latency of LLM responses are areas requiring improvement. The paper acknowledges these challenges and positions them as interesting future research directions. The contributions of SwarmBrain within the domain of LLMs and AI have significant implications for future developments in agent-based strategic gameplay.