SatNav-Episodes-v0.1

118,494 episodes · 59 scenes · 3 navigation task families

SatNav is a benchmark for long-horizon vision-and-language navigation over satellite maps. Agents follow natural-language instructions through continuous geographic space, using pose-conditioned top-down observations to decide where to move and when to stop.

This dataset provides VLN episode annotations: instructions, scene identifiers, starting poses, goals, waypoints, reference paths, and task metadata. Use them with SatSim, the satellite-map simulator in the companion SatNav repository, to generate observations, collect training trajectories, and evaluate navigation policies.

Scenes · Code · Documentation · Model Zoo

What is included

• One original SatNav episode JSON file per split, covering all three task families.
• Parquet versions of the combined files for the Hugging Face Dataset Viewer and datasets library.
• scenes_list.yaml, containing scene identifiers and geographic bounds.
• Data documentation, license notices, and SHA256 checksums.

Obtain the 59 companion GeoTIFF scenes by requesting prepared files or generating them with your own API credentials.

data/                           Parquet files for all three splits
episodes/train/                 Training episodes (all_episodes.json)
episodes/eval/val_seen/         Seen-scene evaluation episodes (all_episodes.json)
episodes/eval/val_unseen/       Unseen-scene evaluation episodes (all_episodes.json)
scenes_list.yaml                Scene identifiers and geographic bounds
README.md                       Dataset Card and split configuration
LICENSE / DATA_LICENSE.md / NOTICE
SHA256SUMS                      File integrity checksums

Verify downloaded files from the repository root with sha256sum -c SHA256SUMS.

Tasks

Task family	JSON value	What the task evaluates	Subtypes
Boundary	Boundary	Following a boundary, tracking route progress, and stopping at the instructed location	loop, arc, extended
Landmark	LandmarkSet	Grounding landmarks and following relative spatial instructions	one_turn, two_turn
Route	Road	Following roads or waterways and reasoning through successive junctions	road, waterway, hybrid

Trajectories are paired with instructions in multiple linguistic styles; one trajectory can have several episodes.

Splits

Split	Boundary	Landmark	Route	Total	Scenes
train	26,121	40,343	38,700	105,164	56
val_seen	1,506	1,509	1,559	4,574	56
val_unseen	2,863	2,872	3,021	8,756	3
Total	30,490	44,724	43,280	118,494	59 unique

val_seen evaluates navigation in the 56 training scenes; val_unseen covers three held-out scenes.

Each split includes all three task families. Filter by trajectory_type to select one family.

Load the data

Install the Hugging Face Datasets library:

pip install datasets

Load all three splits from Hugging Face:

from datasets import load_dataset

splits = load_dataset("Eku127/SatNav-Episodes-v0.1")

example = splits["val_seen"][0]
print(example["instruction_text"])
print(example["scene_id"], example["trajectory_type"])

For private access, sign in with hf auth login. Use revision to load a specific release tag or commit.

Use with SatNav

The episodes/ directory contains an all_episodes.json file for each split, ready to use with SatNav training and evaluation.

Set the dataset root to the downloaded repository directory:

export SATNAV_DATA_ROOT="/absolute/path/to/downloaded-dataset"

Option 1: Request and download prepared scenes

Open SatNav-Scenes-v0.1, sign in to Hugging Face, enter your affiliation and research purpose, accept the terms, and submit. Access is granted after your request passes the system checks.

Download the 59 GeoTIFFs and their checksum manifest using the same account:

pip install -U huggingface_hub
hf auth login
hf download Eku127/SatNav-Scenes-v0.1 --repo-type dataset \
  --include "scenes/*.tif" --include SHA256SUMS \
  --local-dir /path/to/satnav_datasets
export SATNAV_SCENES_DIR="/path/to/satnav_datasets/scenes"
(cd /path/to/satnav_datasets && sha256sum -c SHA256SUMS)

The scene files total 64.79 GB (60.34 GiB). The scene checksum manifest is stored in the parent satnav_datasets/ directory, and the Episodes checksum manifest stays in its own dataset directory.

Option 2: Generate scenes with your own API credentials

Register for Google Map Tiles API or Mapbox and configure your credentials following the map preparation guide. For Google's batch downloader, run from the SatNav repository root:

export GOOGLE_MAPS_API_KEY="your-api-key"
export SATNAV_SCENES_DIR="/path/to/satnav_datasets/scenes"
python -m applications.map_downloader google \
  --scene-config "$SATNAV_DATA_ROOT/scenes_list.yaml" \
  --output-dir "$SATNAV_SCENES_DIR"

Train or evaluate

With either option, point SATNAV_SCENES_DIR to the directory containing the .tif files. Continue with trajectory generation, training, or evaluation.

Data fields

The Parquet representation exposes common fields as columns and retains each full original episode in episode_json.

Column	Type	Description
episode_key	string	Unique row key within this release, formed from the split and zero-based source row index
split	string	train, val_seen, or val_unseen
source_row_index	int64	Row position in the original split's all_episodes.json
instruction_text	string	Navigation instruction
instruction_type	string	Original instruction style label
trajectory_type	string	Boundary, LandmarkSet, or Road
trajectory_subtype	string	Task-specific subtype
scene_id	string	Identifier matching an entry in scenes_list.yaml
episode_id	string	Original episode ID normalized to a string for a consistent column type
trajectory_id	string	Original trajectory ID normalized to a string
difficulty_level	nullable int64	Original instruction difficulty value when present; null otherwise
start_position	list of float64	Starting [longitude, latitude, altitude]
start_rotation	float64	Initial heading in degrees
goal_positions	list of coordinate lists	Goal positions extracted from the original goals field
waypoints	list of coordinate lists	Sparse navigation waypoints
reference_path	list of coordinate lists	Reference trajectory
aux_info_json	string	JSON-encoded task-specific auxiliary metadata
episode_json	string	Complete original episode, including nested instruction and goal objects and original ID types

Coordinates use WGS84 longitude and latitude in degrees, with altitude in meters. Use episode_key as the unique row identifier; source IDs can repeat.

Data sources

Episodes are generated from selected city-scale regions using geospatial structures derived in part from OpenStreetMap. Boundary, landmark, and route construction produce trajectories paired with instructions in multiple linguistic styles.

Research scope

SatNav supports research on long-horizon memory, landmark grounding, route following, and instruction execution. It evaluates top-down navigation in selected urban scenes using real-world geographic coordinates. Full 3D perception, dynamic obstacles, weather, and flight dynamics are outside the simulation scope.

License and attribution

• Episode data (JSON and Parquet) and scene metadata: ODbL 1.0.
• Documentation: CC BY 4.0 unless otherwise stated.
• Contains information derived from OpenStreetMap data © OpenStreetMap contributors. See OpenStreetMap attribution.
• Prepared satellite scenes: SatNav-Scenes-v0.1 access terms. API downloads follow the selected imagery provider agreement.

See LICENSE, DATA_LICENSE.md, and NOTICE for details.

Citation

If you use SatNav in your research, please cite the SatNav paper, accepted at NeurIPS 2026, Track on Evaluations and Datasets.

@inproceedings{jiang2026satnav,
  title = {{SatNav}: A Scalable Benchmark for Long-Horizon {UAV} Vision-Language Navigation from Satellite Imagery},
  author = {Jiang, Jiajun and Hua, Chunliang and Chen, Zichun and Wu, Yanxing and Yang, Zeyuan and Song, Jie and Hu, Xiao},
  booktitle = {Advances in Neural Information Processing Systems},
  year = {2026},
  note = {Track on Evaluations and Datasets; accepted, to appear},
  url = {https://arxiv.org/abs/2609.31507}
}