CSG-Nav: Confidence-Aware Semantic-Geometric Fusion for Robust Exploration and Accurate Zero-Shot Object-Goal Navigation

Zero-shot object-goal navigation requires autonomous agents to explore un seen environments and locate target objects without task-specific training, posing challenges in perceptual reliability under observation uncertainty and physically grounded goal termination. Although foundation-model-based ap proaches improve generalization, they remain vulnerable to unreliable long range perception, hallucination-induced directional bias, and semantically plausible but geometrically incorrect stopping decisions. To address these limitations, we propose CSG-Nav, a lightweight zero-shot navigation frame work centered on confidence-aware multimodal information fusion. CSG-Nav introduces three complementary fusion mechanisms. First, multi-source se mantic fusion integrates open-vocabulary detection results, detection con fidence, scene observations, and relational reasoning to cross-validate se mantic evidence and suppress hallucination-induced directional bias. Sec ond, spatial-confidence fusion weights multi-view semantic evidence using a Gaussian decay function, suppressing unreliable distant observations while accumulating spatially consistent exploration preferences. Third, semantic geometric dual verification fuses semantic target detection with depth-based 3D localization to support physically grounded stopping decisions. Experi ments on HM3Dv1, HM3Dv2, and MP3D demonstrate competitive zero-shot performance, achieving success rates of 58.9%, 71.6%, and 44.3%, respec tively, with a compact 4B model configuration. Fusion-level ablations and real-world deployments in indoor and outdoor environments further verify the effectiveness and practicality of CSG-Nav for robust navigation under uncer tain perception.