Reproduce “AECNav: Active Evidence Consolidation for Efficient Zero-Shot Open-Vocabulary Object Navigation” on OpenResearch

Turn your favorite coding agents into research agents that review literature, develop hypotheses, and run experiments.

Universal app for Apple silicon and Intel. Requires macOS 11 or later.

About this paper

Guanlin Liu, Shaobin Ling, Renyuan Liu, Zeying Gong, Junjie Hu

AECNav addresses zero-shot object-goal navigation by gathering evidence over views before confirming a target. It reuses image features to gate segmentation, accumulates cluster-level log-odds from goal, confuser, and missed-detection evidence, and selects routes using semantic relevance, unknown space, and travel cost. On HM3D-OVON, it reports a 57.3% success rate, 9.0 points above MSGNav; this is benchmark evidence, not proof of real-world reliability.

Key insights

  • Unifying visual encoding with an evidence-gated mechanism significantly reduces perceptual latency and provides a coherent semantic understanding across navigation tasks.
  • A log-odds belief model that integrates positive semantic evidence, negative evidence from confusers, and implicitly negative evidence from missing detections robustly resolves target ambiguity.
  • Optimal exploration in open-vocabulary navigation requires a balanced consideration of semantic relevance, the potential for information gain from unknown areas, and the physical cost of traversal.

Summary by alphaXiv.

Tip: change arxiv.org to autoarxiv.org in any arXiv abstract link to open the paper here.