jometeorie/urbanground-tasks

UrbanGround Tasks

UrbanGround Tasks is the evaluation dataset for UrbanGround: From Local Perception to Spatial Agency in a Real-Scale City. It contains exactly 810 georegistered task instances across five capability levels for evaluating multimodal agents in the UrbanGround real-scale 3D replica of Hong Kong.

• Paper: arXiv:2608.27456
• Code and desktop applications: UrbanGround/UrbanGround
• Interactive project page: urbanground.github.io

Quick start

from datasets import load_dataset

benchmark = load_dataset(
    "jometeorie/urbanground-tasks",
    split="test",
)  # 810 paper benchmark instances

print(benchmark[0])

The test split is the complete evaluation benchmark. It is not intended as a training/test partition. Reference answers and evaluator-only fields are intentionally public for reproducible evaluation; leaderboard runs must control which fields are exposed to the evaluated agent.

Benchmark composition

Level	Capability	Subtask	Abbr.	Instances
1	Local Environment Understanding	Visual Recognition	VR	80
1	Local Environment Understanding	Orientation Understanding	OU	60
1	Local Environment Understanding	Active Exploration Questions	AEQ	80
2	Navigation under Explicit Instructions	Short-Range Goal Navigation	SGN	80
2	Navigation under Explicit Instructions	Long-Range Goal Navigation	LGN	80
2	Navigation under Explicit Instructions	Instructional Navigation	IN	50
2	Navigation under Explicit Instructions	Constrained Navigation	CN	30
3	Exploration under Implicit Instructions	Place-Type Search	PTS	60
3	Exploration under Implicit Instructions	Implicit Intent Inference	III	60
4	Multi-Task Planning	Time-Window Scheduling	TWS	60
4	Multi-Task Planning	Multi-Stop Route Planning	MSP	60
5	Dynamic Environment Interaction	Dynamic Road-Closure Replanning	DCR	30
5	Dynamic Environment Interaction	Navigation among Pedestrians	NP	80
		Total		810

Data organization

data/
└── benchmark.jsonl             # all 810 task instances
metadata/
└── statistics.json             # machine-readable counts
scripts/
└── build_dataset.py

The JSONL file contains one complete task per row, uses snake_case field names, and includes human-readable subtask metadata for the Dataset Viewer.

Important fields

Field	Description
id	Stable task instance identifier.
task_type_id	Integer task type used by the UrbanGround application.
task_type, task_abbreviation	Human-readable subtask name and paper abbreviation.
capability_level, capability_name	Level 1–5 and its capability group.
start_point, end_point, qa_start_point, etc.	WGS84 positions; height is in meters and yaw is in degrees.
qa_options, qa_answer_index, qa_answer_text	Multiple-choice candidates and the reference answer.
restricted_zones	Closed-road geometry for constrained and dynamic replanning.
schedule, multi_route_targets	Ordered or unordered multi-destination task definitions.
notes	Annotator-only notes; these must not be sent to the evaluated agent.

Coordinates and reference fields are evaluator state, not a model-visible observation. In particular, qa_answer_index, qa_answer_text, explore_answer, and the reference end_point of implicit-intent tasks can leak the answer if passed to an evaluated model.

Evaluation protocol and answer visibility

This is an openly released benchmark, so the Dataset Viewer and downloaded files include reference answers and evaluator geometry. The benchmark is not a hidden-label test set. Evaluation must construct the model-visible prompt from the task type and the live UrbanGround observation instead of passing an entire row to the agent.

A conservative default allowlist is id, task_type, task_abbreviation, capability_level, and description. Task-specific controller code may also consume the appropriate start state or instruction parameters, but must keep answer fields, reference goals, route targets, restricted-zone geometry, and annotator metadata on the evaluator side. See the UrbanGround evaluation code for the authoritative per-task protocol.

Using the data with UrbanGround

The dataset contains task definitions, not rendered observations or trajectories. Running an episode requires an UrbanGround desktop build or the browser application. Official application releases already include the task library required by the runtime.

The Hong Kong 3D Visualisation Map, 3D Pedestrian Network, Unity runtime, Cesium, and Microsoft Rocketbox assets are not redistributed in this dataset. Those components remain subject to their respective upstream terms.

Data considerations

• Task coordinates refer to public locations in Hong Kong and use WGS84 longitude/latitude.
• Some questions reproduce public-facing business names, signs, addresses, or telephone numbers visible in the city data.
• Descriptions involving people or households are task scenarios; the dataset does not contain records about evaluation participants.
• The dataset does not contain participant trajectories, account credentials, API keys, private user data, 3D Tiles, screenshots, or application binaries.
• Geographic content can become outdated as streets, businesses, and access conditions change.
• Successful use depends on the matching UrbanGround application and external map services.

License

The original task annotations, schema, and dataset packaging are released under the MIT License, matching the UrbanGround project. External map and runtime components are not included and are not covered by this license. Public geographic and storefront facts reproduced in individual tasks remain subject to any applicable source terms.

Rebuilding the normalized files

From the UrbanGround source checkout, run:

python HuggingFaceDataset/urbanground-tasks/scripts/build_dataset.py \
  --source task \
  --output HuggingFaceDataset/urbanground-tasks

The script validates all 13 subtask counts and writes the single 810-row benchmark split.

Citation

@article{ju2026urbanground,
  title   = {UrbanGround: From Local Perception to Spatial Agency in a Real-Scale City},
  author  = {Ju, Tianjie and Wu, Zheng and Sun, Yueqing and Cui, Yuhan and Li, Bobo and Wu, Shengqiong and Cheng, Pengzhou and Zhao, Haodong and Wu, Zongru and Ma, Xinbei and Zhang, Doris and Li, Kunling and Lee, Mong-Li and Hsu, Wynne and Fei, Hao and Gu, Qi and Liu, Gongshen and Zhang, Zhuosheng},
  journal = {arXiv preprint arXiv:2608.27456},
  year    = {2026}
}