TacticCraft: Natural Language-Driven Tactical Adaptation for StarCraft II

2025-07-21

AI Summary

Current StarCraft II AI agents struggle to execute high-level natural language tactical directives. To address this, we propose an adapter-based tactical modulation framework: building upon the frozen pre-trained DI-Star policy network, we introduce lightweight adapter modules into the action head and design a learnable tactical tensor encoding strategic preferences—including aggression level, expansion pattern, and technology tree progression. Semantic instruction guidance is enforced via KL-divergence regularization, enabling fine-grained, instruction-driven policy adaptation. Experiments demonstrate that our method achieves efficient, low-overhead behavioral modulation across a multidimensional tactical space, with less than 1% parameter increase. It significantly improves policy controllability while preserving the original competitive performance. This work establishes a scalable and practically viable paradigm for instruction-driven real-time strategy AI.

Technology Category

• Multiagent Systems: Adversarial Agents
• Cognitive Modeling & Cognitive Systems: Adaptive Behavior
• Game Theory and Economic Paradigms: Adversarial Learning

Application Category

• Semantics and Knowledge: Data modeling to support human-machine intelligence, including LLMs agents, intelligent system behavior, explanations, and user-friendly interactions
• Search and Retrieval-Augmented AI: Agentic search
• User Modeling, Personalization and Recommendation: Attacks and countermeasures in recommendation systems

Abstract

We present an adapter-based approach for tactical conditioning of StarCraft II AI agents. Current agents, while powerful, lack the ability to adapt their strategies based on high-level tactical directives. Our method freezes a pre-trained policy network (DI-Star) and attaches lightweight adapter modules to each action head, conditioned on a tactical tensor that encodes strategic preferences. By training these adapters with KL divergence constraints, we ensure the policy maintains core competencies while exhibiting tactical variations. Experimental results show our approach successfully modulates agent behavior across tactical dimensions including aggression, expansion patterns, and technology preferences, while maintaining competitive performance. Our method enables flexible tactical control with minimal computational overhead, offering practical strategy customization for complex real-time strategy games.

Problem

• Enabling StarCraft II AI to adapt strategies via natural language directives
• Modulating agent behavior across aggression, expansion, and technology preferences
• Maintaining core competencies while allowing tactical variations efficiently

Innovation

• Adapter-based tactical conditioning for AI agents
• Lightweight modules with KL divergence constraints
• Flexible control with minimal computational overhead