Zero-shot navigation framework reports higher scores on four benchmarks

single source· 1 articles · confidence: medium · first seen 2026-09-13 20:00 UTC

What this means for you

Nothing to act on yet: the paper is a preprint with no stated code or weights release, and the numbers come with no evaluation date or harness configuration, so they cannot be reproduced or compared like-for-like. If you work on embodied navigation, the idea worth borrowing is validating proposed actions against stored spatial evidence rather than trusting the planner.

A group has published HarnessVLN, a navigation system that needs no task-specific training. It wraps a multimodal language model — one that handles images as well as text — in a harness: the scaffolding that decides which tool to call next and checks each proposed move against spatial evidence, geometric feasibility and subgoal consistency. Reported success rates are 60.8% on R2R, 53.9% on RxR, 76.0% on HM3D-v2 and 59.3% on HM3D-OVON, above earlier training-free results. No evaluation date or harness configuration is given, the paper is a preprint, and a humanoid demonstration is reported.

Key facts

  • ·HarnessVLN reports success rates of 60.8% on R2R, 53.9% on RxR, 76.0% on HM3D-v2 and 59.3% on HM3D-OVON. source
  • ·The framework is zero-shot and training-free, wrapping a multimodal large language model behind a unified tool interface rather than fine-tuning it. source
  • ·The Agent Harness validates planner proposals against spatial evidence, geometric feasibility and subgoal consistency before execution. source
  • ·The method uses hierarchical event memory for task progress and a persistent spatiotemporal graph storing spatial evidence and failure annotations. source
  • ·The paper reports a humanoid deployment in real-world environments covering both instruction-following and object-goal navigation. source
  • ·The arXiv preprint, 2609.15195, is dated 13 September 2026; no evaluation date or harness configuration is given for the reported scores. source

What the sources say

  • Hugging Face Daily Papers — Abstract only: sets out the tool interface, memory and spatial graph components, and the four reported success rates.

Sources

The original reporting. Follow these — they did the work.

← the wire