JEV-Star: Low-Cost StarCraft II Control with Language-Model Planning (github.com)

JEV-Star, a new AI framework, has been announced, enabling real-time macro control in StarCraft II through the integration of language-model planning. This system combines macro management with micro-control capabilities, utilizing two separate modules—one for high-level strategy (macro) and another for unit-level decision-making (micro)—each with distinct environments and action spaces. With five configurations to choose from, JEV-Star showcases its effectiveness, achieving 9 wins and 1 loss in configured matches against a Zerg opponent, highlighting its potential by utilizing the Astra planning module alongside the JEV action selection model. This development is significant for the AI/ML community as it exemplifies a cost-effective approach to complex game control, merging language processing with strategic gameplay in a well-structured repository. The seamless interaction between the macro module using the LLM Play SC2 and the micro module utilizing PySC2 demonstrates promising advancements in AI-driven game tactics. By providing a validated setup and clear implementation guidelines, JEV-Star not only contributes to the research landscape but also opens avenues for further exploration in AI applications within dynamic environments.