Zero-shot goal navigation requires an agent to locate targets in unseen environments based on object categories, reference images, or text descriptions, placing high demands on scene understanding and reasoning. Existing methods mainly rely on online observations, modality similarity, or heuristic graph matching, and therefore still struggle with complex target search due to limited use of external knowledge and weak multi-step reasoning. We propose PriorNav, a prior-knowledge-enhanced framework for zero-shot goal navigation. PriorNav learns a unified retrievable knowledge space from semantic, instance, and relational knowledge, maintains a knowledge-enhanced scene graph by fusing retrieved priors with online observations, and performs progressive decision-making through multi-step iterative reasoning across exploration, verification, and approach stages. Experiments on Object-Goal, Image-Instance Goal, and Text-Goal navigation show that PriorNav improves the success rate over the baseline by 3.5%, 13.3%, and 3.5%, respectively, while also outperforming the strongest training-free baselines on all three tasks. Ablation studies further verify the effectiveness of multi-level prior knowledge, scene-graph enhancement, and iterative reasoning. These results show that combining prior knowledge with explicit reasoning is a promising direction for improving zero-shot goal navigation.
Liu et al. (Tue,) studied this question.