HieraNav is an open-vocabulary language-conditioned goal navigation task where agents reach targets across four semantic levels: scene, room, region, and instance.
Language as a Map (LangMap) is the first large-scale benchmark for this task with comprehensive human-verified contrastive annotations.
-
Built on real-world HM3D indoor scans.
-
Human-verified region labels with discriminative descriptions.
-
Discriminative instance descriptions covering 414 object categories.
-
Both concise and detailed instructions to evaluate different instruction styles.
The full benchmark will be released.