Have You Seen It Clearly? Uncertainty-Guided Belief Refinement for Zero-Shot Object Navigation

This research proposes RefBeliefMap, an uncertainty-guided mapping framework for zero-shot object navigation, which distinguishes target-location belief from spatial and semantic evidence uncertainty. The method jointly models intra-observation uncertainty and inter-observation disagreement to refine evidence before navigation. RefBeliefNav further evaluates unexplored frontiers and historical waypoints based on target-finding potential, information gain, and navigation cost. Experiments on HM3D, MP3D, and a Unitree Go2 robot demonstrate the effectiveness of the approach in both simulation and real-world settings.