SAGE: Decoupling Spatial Logic from Metric Scale for Zero-Shot Multi-Robot Exploration

Effective multi-robot exploration in unknown environments faces a critical structural dilemma: high-level coordination typically relies on global spatial priors or fixed-scale representations, yet real-world missions operate in unbounded, irregular domains where such priors are unavailable. Here, we present SAGE, a framework that resolves this dilemma by treating exploration as a topological invariant, independent of metric boundaries. SAGE acts as a dimensional reduction operator, abstracting high-dimensional sensor streams into a sparse, dynamic graph that decouples decision logic from physical scale. To stabilize learning on these continuously growing structures, we introduce a capacity-controlled multi-agent reinforcement learning regime. By synchronizing the agent's observable horizon with its effective action space via a dynamic masking protocol, this mechanism constrains the complexity of the growing environment, ensuring the policy masters local topological correlations. We demonstrate that policies trained exclusively on minimalist, structured 60 × 60 grid worlds transfer zero-shot to environments of vastly different physics and scales—from large-scale unstructured maps to real-world deployment in dense, irregular forests. These results verify that abstracting spatial logic from metric geometry enables robust, boundless exploration without environment-specific tuning.