Two-Bridge: Exclusive Objectives and Extended Horizon StarCraft II Benchmark

Researchers have introduced Two-Bridge, a new intermediate benchmark for StarCraft II that bridges the gap between oversimplified mini-games and computationally expensive full-game scenarios. The benchmark isolates tactical skills like navigation and micro-combat while removing economy mechanics, enabling more efficient reinforcement learning research on real-time strategy environments.

Two-Bridge addresses a critical bottleneck in reinforcement learning research for real-time strategy games. The StarCraft II environment presents researchers with a false choice: mini-games are so simple that basic agents quickly plateau, while full-game scenarios demand massive computational resources that limit experimentation with modern RL algorithms. This new benchmark occupies the neglected middle ground, enabling researchers to study meaningful strategic complexity without requiring enterprise-scale computing infrastructure.

The research community has long recognized that curriculum learning—gradually increasing task difficulty—is essential for developing capable AI agents. However, the absence of intermediate benchmarks has forced researchers to either work with trivial tasks or commit to expensive full-game training pipelines. Two-Bridge's design philosophy of disabling resource collection, base-building, and fog-of-war mechanics focuses computational demand on tactical decision-making rather than state-space explosion, allowing researchers to evaluate RL algorithms under realistic budget constraints.

This development democratizes RTS game AI research. Academic teams and smaller organizations can now conduct serious experiments previously requiring institutional resources. The open-source release as a Gym-compatible wrapper with PySC2 integration lowers barriers to adoption. For the AI research community, Two-Bridge establishes a new standard benchmark that could accelerate algorithm development in complex domains.

Future impact depends on community adoption. If Two-Bridge becomes a reference benchmark comparable to Atari or MuJoCo, it could redirect significant research effort toward RTS game AI. The benchmark's effectiveness at measuring progress on meaningful tactical skills—rather than merely gaming simpler environments—will determine whether it becomes essential infrastructure or another niche research tool.

• →Two-Bridge fills a critical gap between trivial mini-games and computationally prohibitive full-game StarCraft II scenarios
• →By isolating navigation and micro-combat while removing economy mechanics, the benchmark enables efficient RL experimentation on realistic tasks
• →Open-source release as a Gym-compatible wrapper democratizes RTS game AI research for resource-constrained teams
• →The benchmark could accelerate reinforcement learning algorithm development if widely adopted as a standard reference
• →Agents learn coherent tactical behaviors without full-game infrastructure costs, validating the intermediate complexity approach