LangMap: A Human-Verified Benchmark for Hierarchical Open-Vocabulary Goal Navigation

NeurIPS 2026

Bo Miao1, Weijia Liu2, Jun Luo1, Lachlan Shinnick1, Jian Liu3, Thomas Hamilton-Smith1,
Yuhe Yang4, Zijie Wu4, Vanja Videnovic5, Feras Dayoub1, Anton van den Hengel1,6
1Adelaide University 2ECNU 3SUTD 4UWA 5Breaker Industries 6Metacognition AI
Paper Data and Code Video BibTeX
LangMap and HieraNav benchmark framework

Abstract

We introduce HieraNav, an open-vocabulary goal navigation task with goals at four hierarchical semantic levels: scene, room, region, and instance. We present LangMap (Language as a Map), a benchmark built on HM3D with human-verified region labels and discriminative region and instance descriptions, covering 414 object categories and over 18K tasks. Each target has concise and detailed descriptions, and our descriptions improve text-to-view matching accuracy over GOAT-Bench’s by 23 points. We also propose PlaNaVid, an RGB-only baseline that combines Bounded Diverse Memory with high-level planning, achieving top-tier success rates without depth, 3D scene representations, or object masks.

Highlights

Scene-, room-, region-, and instance-level goals, plus mixed-level multi-goal navigation.

Human-verified semantic labels and discriminative language annotations.

All HM3D-Sem validation scenes, 414 object categories, and over 18K tasks.

Task Levels

Examples of object, room, region, and instance navigation goals

Object: find any object of the target category, such as "Find the armchair."

Room: find the target object category in a room type, such as "Find the armchair in the bedroom."

Region: find the target object category in a described room instance, such as "Find the armchair in the bedroom that has a geometric rug."

Instance: find a specific object instance from its description, such as "Find the black leather armchair."

PlaNaVid Visualization

Qualitative example of PlaNaVid on multi-goal navigation.

Annotation Quality

Qualitative comparison between LangMap and GOAT-Bench annotations
Qualitative comparison of LangMap and GOAT-Bench annotations. Blue marks semantic errors, and green marks ambiguous descriptions. View full resolution.

BibTeX

@inproceedings{miao2026langmap,
  title     = {LangMap: A Human-Verified Benchmark for Hierarchical Open-Vocabulary Goal Navigation},
  author    = {Miao, Bo and Liu, Weijia and Luo, Jun and Shinnick, Lachlan and Liu, Jian and
               Hamilton-Smith, Thomas and Yang, Yuhe and Wu, Zijie and Videnovic, Vanja and
               Dayoub, Feras and van den Hengel, Anton},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}