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

2 min read

PSG-Nav does open-vocabulary object-goal navigation over a 3D probabilistic scene graph. Each node keeps a full categorical distribution over labels instead of a single best guess. A “Multiverse Decision” step samples likely world configurations from the joint distribution and scores candidate landmarks by how well they fit those samples. An “Evidential Experience Calibrator” checks new detections against memories of earlier successes and failures to reject false positives, which the authors present as online lifelong adaptation. The README says it is built on SG-Nav and that the current release targets the HSSD ObjectNav benchmark in Habitat, with local Qwen2.5 LLM and VLM servers.

Findings

  • Success rates of 66.1% on MP3D, 44.8% on HM3D and 67.9% on HSSD, which the authors report as new best results (abstract). I have read only the abstract and README.

Limits

  • Simulation only (Habitat), one robot.
  • The README says the code is still being cleaned up for release.

Relevance

A team fusing robots that disagree about what they saw would need label distributions like these, so that one overconfident detector cannot win. Open question: does the calibrator’s memory of past failures hold up when the failures come from another robot? See robust-estimation, open-vocabulary-mapping and multi-robot perception.