Multi-Granularity Language Feature Fields as Zero-Shot Beyond-Category Object Navigators

Navigation driven by descriptive object intention introduces the challenge of understanding high-dimensional semantics from constrained observations. Existing object-centric navigation methods remain limited by their reliance on per-frame scalar-collapsed query results and the subsequent reconciliation of historical biases, which renders them inadequate when handling multi-attribute targets. In contrast, we couple divergenceless historical visual evidence with decision-granularity-aware late-querying to realize a zero-shot Beyond-Category Object Navigation (BCON) system. For spatial memory, we project regional features for co-occurrence reasoning and local features for scene understanding into 3D space, maintaining multi-granularity feature fields through feature concatenation and fusion re-perception, respectively. For searching, benefiting from visual cues revealed by spatial accumulation, we synthesize regional features involved in imaginatively observing the frontiers into viewpoint feature vectors to evaluate search potential. For grounding, leveraging the decisive manifold re-anchoring enabled by feature fusion, we aggregate semantic highlands queried from entire-field voxels of local features to ground conspicuous targets. The field insight demonstrates 1.76× success rate (65/96 episodes) over the baselines (highest 37/96 episodes) on beyond-category tasks. The generalizability of foundation models and the embodiment-agnostic action design enable the system to drive real-world quadruped robots.