LangMap: A Hierarchical Benchmark for Open-Vocabulary Goal Navigation

Bo Miao1,  Weijia Liu2,  Jun Luo3,  Lachlan Shinnick1,  Jian Liu3,  Thomas Hamilton-Smith1,  Yuhe Yang4, 
Zijie Wu5,  Vanja Videnovic6,  Feras Dayoub1,  Anton van den Hengel1
1 AIML, University of Adelaide   |   2 East China Normal University   |   3 NERC-RVC, Hunan University
4 University of Western Australia   |   5 Singapore University of Technology and Design   |   6 Breaker Industries
arXiv Dataset Code Annotation Tool Interactive Contrastive Visualization
Interpolate start reference image.

Abstract

We introduce HieraNav, a multi-granularity, open-vocabulary goal navigation task where agents interpret natural language instructions to reach targets at four semantic levels: scene, room, region, and instance.

We present Language as a Map (LangMap), the first benchmark built on real-world 3D scans with comprehensive human-verified annotations and tasks spanning these levels. LangMap provides region labels, discriminative region descriptions, discriminative instance descriptions covering 414 object categories, and over 18K tasks. Each target features both concise and detailed descriptions, enabling evaluation across different instruction styles.

LangMap demonstrates greater diversity, higher annotation quality, and larger scale, surpassing GOAT-Bench by 23.8% in discriminative accuracy with four times fewer words.

Evaluations of zero-shot and supervised models on LangMap reveal that richer context and memory improve success, while long-tailed, small, context-dependent, and distant goals, as well as multi-goal completion, remain challenging.

Qualitative Comparison: LangMap vs. GOAT-Bench Annotations

Figure 1. Blue indicates semantic errors and green indicates ambiguous descriptions.

BibTeX

@article{miao2026langmap,
  title = {LangMap: A Hierarchical Benchmark for Open-Vocabulary Goal Navigation},
  author = {Bo Miao and Weijia Liu and Jun Luo and Lachlan Shinnick and Jian Liu and Thomas Hamilton-Smith and Yuhe Yang and Zijie Wu and Vanja Videnovic and Feras Dayoub and {Anton van den} Hengel},
  journal = {arXiv preprint arXiv:2602.02220},
  year = {2026}
}