Decoupled Complementary Fields on 3D Gaussian Maps for Embodied Navigation and Reasoning

Embodied navigation and reasoning require an agent to actively acquire, organize, and verify task-relevant evidence in previously unseen environments. Recent methods typically rely on semantic maps, scene graphs, vision-language priors, or structured memory, but often underuse persistent fine-grained 3D evidence and entangle query-related ambiguity with environment-side reliability. This coupling can make evidence acquisition inefficient and final decisions vulnerable to weak or transient observations. We propose Decoupled Complementary Fields on 3D Gaussian Maps, an embodied navigation framework that maintains persistent volumetric evidence, separates query-related ambiguity from environment-side readiness, and grounds decisions through progressive verification. First, we construct a query-conditioned object-centric Gaussian navigation state built on online 3D Gaussian Splatting, which preserves fine-grained 3D evidence across viewpoints while providing an executable spatial interface for region reasoning and navigation. We further formulate Decoupled Complementary Fields, consisting of a Query-Ambiguity Field, which captures where task-relevant evidence remains unresolved, and an Environment-Readiness Field, which estimates where the environment is structurally reliable for motion and inspection. Finally, we couple these fields in a hierarchical navigation-and-verification policy, which selects informative frontiers, instantiates executable local goals, and verifies task-relevant candidates using persistent volumetric Gaussian evidence before stopping or answer handoff. Experiments on A-EQA and GOAT-Bench show consistent gains in active evidence acquisition, multi-modal lifelong navigation, and reliable final decision making.