LangMap evaluates hierarchical open-vocabulary goal navigation: agents navigate to goals specified at four semantic levels (scene, room, region, and instance), either one at a time or chained in multi-goal episodes.

This package contains the LangMap annotations for the 36 HM3D validation scenes: annotation metadata, single-goal and multi-goal task definitions, and this documentation. They are released under CC BY-NC 4.0, which covers only our annotations, task definitions, and documentation and does not supersede the terms of HM3D and HM3D-Sem.

The original HM3D and HM3D-Sem scene assets are not redistributed. Users must obtain the underlying HM3D and HM3D-Sem assets from their official sources and comply with their original licenses and terms of use.

Official data sources:

HM3D: https://aihabitat.org/datasets/hm3d/
HM3D-Sem: https://aihabitat.org/datasets/hm3d-semantics/

Quick Start

Download the annotations:

hf download bo-miao/LangMap --repo-type dataset --include "annotations/*" --local-dir data/LangMap

Load a scene and build a region-level instruction:

import gzip, json

with gzip.open("data/LangMap/annotations/00800-TEEsavR23oF.json.gz", "rt") as f:
    scene = json.load(f)

task = scene["episodes_by_region_level"][0]
region = scene["region_annotation"][task["region_id"]]
print(f"Find the {task['object_category']} in the {region['region_category'].lower()} that has {region['concise_description']}")
# Find the wardrobe in the bedroom that has white-sheeted bed
print(task["target_object_ids"])  # ['wardrobe_11', 'wardrobe_9']

Evaluation code is available at https://github.com/bo-miao/LangMap.

Top-Level Structure

Each JSON file is a dictionary with the following top-level fields:

goals
episodes_by_object_level
episodes_by_room_level
episodes_by_region_level
episodes_by_instance_level
episode_by_sequence
region_annotation

`goals`

goals is a list of object entries. Each entry corresponds to one object instance in the scene.

Fields

Field	Description
`object_category`	Object category.
`object_id`	Unique object instance identifier within the scene.
`position`	3D object position in the Habitat scene coordinate system.
`view_points`	Candidate agent viewpoints from which the object is visible; they serve as navigation goals.
`annot_region_id`	Region identifier.
`annot_manual_category`	Human-verified category labels.
`annot_unique_concise_description`	Concise discriminative description (only for objects with instance annotations).
`annot_unique_detailed_description`	Detailed discriminative description with richer visual and spatial context (only for objects with instance annotations).

Example

{
  "object_category": "wardrobe",
  "object_id": "wardrobe_9",
  "position": [-2.37497, 3.99337, -6.1985],
  "annot_region_id": "region_1",
  "annot_manual_category": ["wardrobe"],
  "annot_unique_concise_description": "tall wardrobe near white-sheeted bed",
  "annot_unique_detailed_description": "tall brown wood wardrobe in bedroom with plain white sheets on bed near Mickey Mouse plush toy"
}

`view_points`

Each object entry contains a view_points list. Each viewpoint records an agent pose and the target visibility from that pose.

Example

{
  "agent_state": {
    "position": [-1.86859, 3.14393, -6.26123],
    "rotation": [0.0, 0.74931, -0.0, 0.66222]
  },
  "iou": 0.8129
}

`region_annotation`

region_annotation is a dictionary mapping each region ID to its human-labeled region label and discriminative descriptions.

Fields

Field	Description
`region_id`	Region identifier.
`region_category`	Human-labeled room or region category.
`is_description_unique`	Whether the region description is intended to uniquely identify the region within the scene (regions with 0 may have empty descriptions).
`concise_description`	Short discriminative region description.
`detailed_description`	Longer region description with additional visual and spatial context.

Example

{
  "region_id": "region_1",
  "region_category": "Bedroom",
  "is_description_unique": 1,
  "concise_description": "white-sheeted bed",
  "detailed_description": "bed with plain white sheets and light teal pillows, tall wooden headboard, Mickey Mouse plush on dark wood nightstand, matching wardrobe and beige armchair, TV on wood cabinet"
}

Single-Goal Episode Lists

The file provides four lists of single-goal tasks, corresponding to four semantic levels.

Field	Description
`episodes_by_object_level`	Scene-level navigation tasks.
`episodes_by_room_level`	Room-level navigation tasks.
`episodes_by_region_level`	Region-level navigation tasks.
`episodes_by_instance_level`	Instance-level navigation tasks.

Each single-goal task defines an initial agent pose and a target set. A task is identified as <scene_id>-<scene_name>_<navigation_type>_<episode_id>, e.g., 00800-TEEsavR23oF_object_0.

Common Fields

Field	Description
`episode_id`	Task index within this semantic level (equal to its position in the list).
`start_position`	Initial agent position [x, y, z].
`start_rotation`	Initial agent orientation quaternion [x, y, z, w].
`min_dist`	Distances stored at task generation: the geodesic distance from the start pose to the nearest target viewpoint (Stretch navmesh) and the straight-line distance to that viewpoint. They are for reference only; the evaluation recomputes distances.
`navigation_type`	Semantic level of the episode. One of object, room, region, or instance.
`target_object_ids`	List of valid target object instance IDs. This list defines the target set; do not reconstruct it by matching object_category.

Level-Specific Fields

Field	Description
`object_category`	Target object category for object-, room-, and region-level tasks.
`room_name`	Target room category for room-level tasks.
`region_id`	Target region ID for region-level tasks.
`instance_id`	Target object instance ID for instance-level tasks.

Object-Level Example

{
  "episode_id": 0,
  "start_position": [-6.46888, 0.01338, -4.43047],
  "start_rotation": [0, 0.98206, 0, 0.18855],
  "min_dist": [{"geodesic_distance": 8.69194, "euclidean_distance": 6.46287}],
  "object_category": "wardrobe",
  "navigation_type": "object",
  "target_object_ids": ["wardrobe_11", "wardrobe_135", "wardrobe_137", "wardrobe_511", "wardrobe_9"]
}

Room-Level Example

{
  "episode_id": 0,
  "start_position": [-6.26957, 0.01338, -0.41912],
  "start_rotation": [0, 0.9958, 0, 0.09157],
  "min_dist": [{"geodesic_distance": 8.97802, "euclidean_distance": 7.83757}],
  "object_category": "wardrobe",
  "navigation_type": "room",
  "room_name": "Bedroom",
  "target_object_ids": ["wardrobe_11", "wardrobe_511", "wardrobe_9"]
}

Region-Level Example

{
  "episode_id": 0,
  "start_position": [-3.91646, 3.11338, -0.85312],
  "start_rotation": [0, 0.43328, 0, 0.90126],
  "min_dist": [{"geodesic_distance": 7.83349, "euclidean_distance": 4.28811}],
  "object_category": "wardrobe",
  "navigation_type": "region",
  "region_id": "region_1",
  "target_object_ids": ["wardrobe_11", "wardrobe_9"]
}

Instance-Level Example

{
  "episode_id": 0,
  "start_position": [-4.03902, 3.11338, -0.06032],
  "start_rotation": [0, 0.9173, 0, -0.39819],
  "min_dist": [{"geodesic_distance": 8.38719, "euclidean_distance": 5.08738}],
  "navigation_type": "instance",
  "instance_id": "wardrobe_9",
  "target_object_ids": ["wardrobe_9"]
}

Instruction Construction

The JSON files store structured task metadata rather than fixed natural-language instructions. The evaluation code builds instructions with the templates below, and the results reported in the paper use exactly these strings; other phrasings can be composed from the same metadata.

Scene level:    Find the {object_category}.
Room level:     Find the {object_category} in the {room_name}
Region level:   Find the {object_category} in the {region_category} that has {region_description}
Instance level: Find the {instance_description}

region_category and the region description come from region_annotation[region_id]; the instance description comes from the goals entry whose object_id equals instance_id.
Concise evaluation uses concise_description and annot_unique_concise_description; detailed evaluation uses detailed_description and annot_unique_detailed_description.
Room names of the form Tie: A & B mark areas that span two rooms. When the room name, region category, or region description contains Tie:, the instruction ends with , 'Tie: A & B' means the area spans both rooms A and B.
Multi-goal subtasks use the same templates with concise descriptions.

`episode_by_sequence`

episode_by_sequence contains multi-goal navigation episodes. Each sequence episode starts from one initial pose and asks the agent to complete a list of goals in order. The first subtask starts from the sequence start pose; each later subtask starts from the pose where the previous subtask ended.

Fields

Field	Description
`episode_id`	Sequence episode index.
`start_position`	Initial agent position [x, y, z].
`start_rotation`	Initial agent orientation quaternion [x, y, z, w].
`min_dist`	Distance from the initial pose to the first valid target.
`task_sequence`	Ordered list of subgoals. Each subgoal is represented as [navigation_type, episode_index].
`navigation_type`	Always sequence for this list.

Interpreting `task_sequence`

Each item in task_sequence points to one of the single-goal episode lists. For example:

"task_sequence": [["instance", 0], ["region", 43], ["region", 10], ["room", 31], ["region", 21]]

This means:

1. Use episodes_by_instance_level[0]
2. Use episodes_by_region_level[43]
3. Use episodes_by_region_level[10]
4. Use episodes_by_room_level[31]
5. Use episodes_by_region_level[21]

Evaluation Protocol

Actions:     MOVE_FORWARD (0.25 m), TURN_LEFT / TURN_RIGHT (30 degrees), STOP
Step limit:  500 steps per task (per subtask in multi-goal episodes)
Navmesh:     radius 0.17 m, height 1.41 m, max climb 0.10 m, cell height 0.05 m
Success:     the final position is within 0.25 m (geodesic) of a viewpoint of any target object (please use 0.25m rather than 1.0m for fair comparison)
SPL:         success * L / max(L, P), where L is the geodesic distance from the (sub)task start to the
             nearest target viewpoint and P is the length of the executed path
SeqSR@k:     fraction of multi-goal episodes whose first k subtasks all succeed

Excluded Tasks

The 106 single-goal tasks below are excluded from evaluation because of scene quality issues. They remain in the files so that task IDs and the indices in task_sequence stay unchanged; no multi-goal episode refers to them.

EXCLUDED_TASK_IDS = [
    "00800-TEEsavR23oF_room_74",
    "00810-CrMo8WxCyVb_object_30",
    "00810-CrMo8WxCyVb_object_31",
    "00810-CrMo8WxCyVb_region_75",
    "00810-CrMo8WxCyVb_region_76",
    "00810-CrMo8WxCyVb_room_53",
    "00810-CrMo8WxCyVb_room_54",
    "00820-mL8ThkuaVTM_object_29",
    "00820-mL8ThkuaVTM_region_47",
    "00820-mL8ThkuaVTM_room_39",
    "00823-7MXmsvcQjpJ_room_100",
    "00823-7MXmsvcQjpJ_room_101",
    "00823-7MXmsvcQjpJ_room_74",
    "00823-7MXmsvcQjpJ_room_98",
    "00823-7MXmsvcQjpJ_room_99",
    "00829-QaLdnwvtxbs_room_14",
    "00829-QaLdnwvtxbs_room_16",
    "00829-QaLdnwvtxbs_room_21",
    "00829-QaLdnwvtxbs_room_22",
    "00829-QaLdnwvtxbs_room_23",
    "00829-QaLdnwvtxbs_room_24",
    "00829-QaLdnwvtxbs_room_25",
    "00829-QaLdnwvtxbs_room_5",
    "00829-QaLdnwvtxbs_room_7",
    "00829-QaLdnwvtxbs_room_9",
    "00832-qyAac8rV8Zk_region_18",
    "00832-qyAac8rV8Zk_room_18",
    "00839-zt1RVoi7PcG_object_25",
    "00839-zt1RVoi7PcG_region_70",
    "00839-zt1RVoi7PcG_room_54",
    "00862-LT9Jq6dN3Ea_object_46",
    "00862-LT9Jq6dN3Ea_region_135",
    "00862-LT9Jq6dN3Ea_region_37",
    "00862-LT9Jq6dN3Ea_room_21",
    "00862-LT9Jq6dN3Ea_room_93",
    "00871-VBzV5z6i1WS_object_21",
    "00871-VBzV5z6i1WS_object_28",
    "00871-VBzV5z6i1WS_object_29",
    "00871-VBzV5z6i1WS_object_38",
    "00871-VBzV5z6i1WS_object_8",
    "00871-VBzV5z6i1WS_region_32",
    "00871-VBzV5z6i1WS_region_58",
    "00871-VBzV5z6i1WS_region_67",
    "00871-VBzV5z6i1WS_region_68",
    "00871-VBzV5z6i1WS_region_77",
    "00871-VBzV5z6i1WS_room_21",
    "00871-VBzV5z6i1WS_room_39",
    "00871-VBzV5z6i1WS_room_46",
    "00871-VBzV5z6i1WS_room_47",
    "00871-VBzV5z6i1WS_room_56",
    "00873-bxsVRursffK_object_29",
    "00873-bxsVRursffK_region_52",
    "00873-bxsVRursffK_room_39",
    "00876-mv2HUxq3B53_object_16",
    "00876-mv2HUxq3B53_object_31",
    "00876-mv2HUxq3B53_object_40",
    "00876-mv2HUxq3B53_region_33",
    "00876-mv2HUxq3B53_region_51",
    "00876-mv2HUxq3B53_region_69",
    "00876-mv2HUxq3B53_region_81",
    "00876-mv2HUxq3B53_region_92",
    "00876-mv2HUxq3B53_region_93",
    "00876-mv2HUxq3B53_room_19",
    "00876-mv2HUxq3B53_room_31",
    "00876-mv2HUxq3B53_room_45",
    "00876-mv2HUxq3B53_room_50",
    "00876-mv2HUxq3B53_room_59",
    "00877-4ok3usBNeis_object_44",
    "00877-4ok3usBNeis_region_58",
    "00877-4ok3usBNeis_region_8",
    "00877-4ok3usBNeis_room_52",
    "00877-4ok3usBNeis_room_9",
    "00878-XB4GS9ShBRE_object_18",
    "00878-XB4GS9ShBRE_region_33",
    "00878-XB4GS9ShBRE_room_32",
    "00880-Nfvxx8J5NCo_object_13",
    "00880-Nfvxx8J5NCo_object_30",
    "00880-Nfvxx8J5NCo_object_31",
    "00880-Nfvxx8J5NCo_region_22",
    "00880-Nfvxx8J5NCo_region_45",
    "00880-Nfvxx8J5NCo_region_46",
    "00880-Nfvxx8J5NCo_room_23",
    "00880-Nfvxx8J5NCo_room_44",
    "00880-Nfvxx8J5NCo_room_45",
    "00890-6s7QHgap2fW_region_19",
    "00890-6s7QHgap2fW_room_16",
    "00891-cvZr5TUy5C5_object_12",
    "00891-cvZr5TUy5C5_object_43",
    "00891-cvZr5TUy5C5_object_44",
    "00891-cvZr5TUy5C5_region_109",
    "00891-cvZr5TUy5C5_region_110",
    "00891-cvZr5TUy5C5_region_111",
    "00891-cvZr5TUy5C5_region_112",
    "00891-cvZr5TUy5C5_region_39",
    "00891-cvZr5TUy5C5_region_40",
    "00891-cvZr5TUy5C5_region_44",
    "00891-cvZr5TUy5C5_room_34",
    "00891-cvZr5TUy5C5_room_35",
    "00891-cvZr5TUy5C5_room_36",
    "00891-cvZr5TUy5C5_room_92",
    "00891-cvZr5TUy5C5_room_93",
    "00891-cvZr5TUy5C5_room_94",
    "00891-cvZr5TUy5C5_room_95",
    "00891-cvZr5TUy5C5_room_96",
    "00891-cvZr5TUy5C5_room_97",
    "00894-HY1NcmCgn3n_instance_7",
]

Citation

@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}
}

Downloads last month: 84

Total file size:

4.67 MB

Paper for bo-miao/LangMap

LangMap: A Hierarchical Benchmark for Open-Vocabulary Goal Navigation

Paper • 2602.02220 • Published Feb 2 • 1

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

Quick Start

Top-Level Structure

goals

Fields

Example

view_points

Example

region_annotation

Fields

Example

Single-Goal Episode Lists

Common Fields

Level-Specific Fields

Object-Level Example

Room-Level Example

Region-Level Example

Instance-Level Example

Instruction Construction

episode_by_sequence

Fields

Interpreting task_sequence

Evaluation Protocol

Excluded Tasks

Citation

Paper for bo-miao/LangMap

`goals`

`view_points`

`region_annotation`

`episode_by_sequence`

Interpreting `task_sequence`