A Multimodal RTS Game Benchmark for Evaluating Vision-Language Models

Key Points

• The goal is to systematically evaluate the reasoning ability of vision-language models in real-time strategy game environments.
• Developed a multimodal observation space based on StarCraft II.
• Introduced queue-based action development and FSM-based group control.
• Conducted experiments comparing the proposed benchmark with existing text-based approaches.
• The new benchmark significantly improves the stability and performance of tactical decision-making compared to traditional methods.
• Effectively compares and analyzes the tactical reasoning capabilities of various vision-language models.