Adaptive Command: Real-Time Policy Adjustment via Language Models in StarCraft II

TL;DR

Adaptive Command proposes a language-conditioned policy for real-time StarCraft II play by integrating an LLM-based strategic advisor with a behavior-tree executor and a natural language interface. It introduces a Chain of Summarization (CoS) to compress game state into $\hat{s}_{t}$ and uses two phases: Initial Policy Selection with $π_0 = f_{LLM}(s_0, c_0, θ)$ and Real-time Policy Adjustment updating $π_{t} = f_{LLM}(\hat{s}_{t+1}, a_t, c_{t+1}, θ)$, with $a_t = g_{BT}(π_t, s_t)$. User studies across novice, intermediate, and expert players show that Adaptive Command increases win rates for novices (and improves intermediate performance) while providing high instruction-following and helpfulness scores for non-expert users, though expert users reveal mixed results and higher expectations. The work demonstrates the practicality of real-time, language-guided human-AI collaboration in dynamic decision-making tasks, with implications for accessibility and broader domains beyond RTS games.

Abstract

We present Adaptive Command, a novel framework integrating large language models (LLMs) with behavior trees for real-time strategic decision-making in StarCraft II. Our system focuses on enhancing human-AI collaboration in complex, dynamic environments through natural language interactions. The framework comprises: (1) an LLM-based strategic advisor, (2) a behavior tree for action execution, and (3) a natural language interface with speech capabilities. User studies demonstrate significant improvements in player decision-making and strategic adaptability, particularly benefiting novice players and those with disabilities. This work contributes to the field of real-time human-AI collaborative decision-making, offering insights applicable beyond RTS games to various complex decision-making scenarios.

Figures (4)

• Figure 1: Adaptive Command Framework: Two-stage process integrating LLMs with behavior trees for strategic decision-making in StarCraft II.
• Figure 2: Adaptive Command in action: Real-time strategy adjustment through human-AI collaboration
• Figure 3: Average Win Rate by Skill Level and Group
• Figure 4: Individual participant scores for Instruction Following and Helpfulness. Each point represents a participant, with numbers indicating participant IDs. Point size and color denote skill level (Novice, Intermediate, Expert). Instruction Following measures the system's ability to accurately implement player instructions, while Helpfulness reflects the perceived usefulness of the Adaptive Command system.