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

Researchers introduce PSG-Nav, a novel navigation system that uses probabilistic scene graphs to help AI agents navigate complex environments while accounting for perception uncertainty. The system achieves state-of-the-art results on three major benchmarks by employing multiverse decision-making and an evidential calibrator to reduce false positives in open-vocabulary navigation tasks.

PSG-Nav addresses a fundamental challenge in embodied AI: enabling agents to navigate real-world environments despite inherent uncertainty in perception systems. Traditional navigation approaches rely on deterministic decisions based on single-point estimates, which can lead to suboptimal pathfinding when facing semantic ambiguity or model errors. This research tackles the problem by representing uncertainty explicitly through probabilistic distributions rather than treating perception outputs as definitive.

The core innovation involves three technical components working in concert. The 3D probabilistic scene graph maintains full categorical distributions for semantic understanding rather than point estimates. The multiverse decision-making approach samples multiple plausible world configurations from the joint distribution, allowing the system to evaluate navigation choices across multiple scenarios simultaneously. The Evidential Experience Calibrator learns from past successes and failures to calibrate detector confidence online, reducing false positives that plague open-vocabulary systems.

For the embodied AI industry, this represents meaningful progress toward more robust autonomous agents. The benchmark results—achieving 66.1% success on MP3D, 44.8% on HM3D, and 67.9% on HSSD—demonstrate substantial improvements over prior approaches. These gains matter for applications spanning robotic navigation, household robots, and search-and-rescue systems where navigation reliability directly impacts safety and effectiveness.

The approach's emphasis on uncertainty quantification and lifelong learning through experience calibration aligns with broader trends in AI toward more transparent, adaptable systems. Future work will likely extend these probabilistic reasoning methods to other embodied AI tasks requiring robust decision-making under uncertainty.

• →PSG-Nav uses probabilistic scene graphs to explicitly model perception uncertainty rather than relying on deterministic navigation approaches.
• →Multiverse decision-making evaluates navigation landmarks across multiple plausible world configurations to find globally optimal paths.
• →An evidential experience calibrator learns from past navigation successes and failures to improve detector confidence online.
• →State-of-the-art results on MP3D, HM3D, and HSSD benchmarks demonstrate significant improvements in embodied AI navigation performance.
• →The system addresses the open-vocabulary navigation challenge where semantic ambiguity and model errors typically degrade agent performance.