Title: PSG-Nav: Probabilistic Scene Graph Navigation via Multiverse Decision Making

Abstract: Open-vocabulary navigation demands that embodied agents effectively handle substantial perception uncertainty, which arises from semantic ambiguity and model inaccuracies. Currently, most existing methods rely on local optimal deterministic strategies, thereby overlooking complex decision-making processes that consider multiple composite possibilities—factors that are essential for achieving globally superior solutions. To address this, we introduce Probabilistic Scene Graph Navigation (PSG-Nav), a framework that constructs a 3D Probabilistic Scene Graph utilizing full semantic categorical distributions to capture perception uncertainty.

To leverage local distributions for composing and reasoning about optimal navigation landmarks, we propose Multiverse Decision. This approach samples the most probable world settings from the joint distribution and evaluates navigation landmarks by assessing their compatibility with these multiverses. Furthermore, to reduce false positives caused by epistemic uncertainty in open-vocabulary navigation, we present the Evidential Experience Calibrator. This component facilitates online lifelong adaptation by cross-validating detections against memories of past successes and failures.

Extensive experiments conducted on standard benchmarks, including MP3D, HM3D, and HSSD, show that PSG-Nav sets new state-of-the-art performance levels. Specifically, it achieves Success Rates of 66.1%, 44.8%, and 67.9% on these respective datasets. The code for this project is available at: https://psg-nav.github.io/

Source: arXiv Generated at: 2026-06-02 00:00:00 UTC