🌍 PASSAGE — Real-Terrain Benchmark & Generator

Trustworthiness-oriented real-terrain benchmark and generation pipeline for pathfinding and surrogate modeling.

📖 Documentation • 🤗 Dataset • 💻 GitHub • 📝 Cite

---

---

Get started

PASSAGE is an open-source pathfinding dataset and generation toolkit built on real-world elevation data to benchmark and improve terrain-aware routing algorithms.

• Project goals: provide reproducible, multi-resolution benchmarks for pathfinding research, ML surrogates, and safety-critical AI evaluation.
• Sponsoring SIG: Anonymous Group open-source and critical AI research contributors.
• Community contact: use GitHub Issues, GitHub Discussions, or email oss@anonymous.invalid.

For installation and first run commands, see 🚀 Quick Start.

Documentation

Documentation is available at anonymous-org.github.io/passage.

The documentation site is published with GitHub Pages and includes user guides, API references, and architecture notes.

Please also set the repository About metadata:

• Description: short project purpose statement for PASSAGE.
• Website: https://anonymous-org.github.io/passage/.

---

🚁 Why PASSAGE?

Every year, Helicopter Emergency Medical Services (HEMS) crews across Europe perform over 300,000 rescue missions — racing against time through mountain passes, across valleys, and over ridgelines to reach patients in cardiac arrest or severe trauma. For every minute of delay, survival rates drop by 7–10%. The path the helicopter takes matters — yet pathfinding algorithms are still benchmarked on flat synthetic grids.

PASSAGE was created to change that. Built from real-world satellite elevation data (JAXA ALOS AW3D30 at 30 m/pixel), it provides the research community with large-scale, multi-resolution pathfinding benchmarks that capture the true complexity of terrain navigation.

We open-source this toolkit to make terrain-routing methods easier to compare, reproduce, and audit under a shared benchmark surface.

---

✨ Highlights

🌐 Multi-Resolution 64×64 → 4096×4096 px	🏔️ Real Elevation JAXA ALOS AW3D30 DSM	🔷 Rich Obstacles Superellipse shapes	⚡ 3 A* Cost Variants elevation, slope_uphill, energy
🧮 Grid A* Oracle Exact labels + timings	🤗 HuggingFace-Ready Parquet + one-command upload	🔁 Fully Resumable No wasted compute	✅ Tested & CI/CD 80%+ coverage, pre-commit

4.4 M Default Samples

7 Resolution Levels

3 A* Configurations

1 Solver Backend

30 m/px Spatial Resolution

---

🗺️ Visual Summary

Normalized elevation with start/end markers, procedural superellipse obstacles, and two computed paths (free terrain vs. obstacle-aware).

---

🚀 Quick Start

CPU-only by default. No GPU is required to download, generate, or evaluate the dataset, and uv sync does not install any proprietary NVIDIA CUDA runtime libraries. The optional passage-eval[torch] extra is wired to the CPU-only PyTorch index — see 🧾 Disclosures for the GPU install path.

# Install uv (if needed)
curl -LsSf https://astral.sh/uv/install.sh | sh

# Clone & install
git clone https://github.com/anonymous-org/passage.git
cd passage && uv sync

# One-command local rebuild (download -> calibrate -> generate -> export -> manifest)
make rebuild RESOLUTION=256

# Or run each stage manually
passage download                       # 1️⃣  Fetch JAXA elevation tiles
passage calibrate                      # 2️⃣  Compute global elevation stats
passage generate --resolution 256      # 3️⃣  Generate pathfinding samples
passage export --no-upload             # 4️⃣  Package as Parquet shards
make manifest                          # 5️⃣  Write outputs/export/manifest.sha256

# Point generate/export at a specific calibration file when needed
passage generate --calibration-filepath /path/to/calibrate.json --resolution 256
passage export --calibration-filepath /path/to/calibrate.json --no-upload

💡 Alternative: install with pip

python -m venv .venv && source .venv/bin/activate
pip install -e ".[dev]"

---

🔁 Pipeline Overview

1. Download JAXA DSM tiles

→

2. Calibrate min / max elevation

→

3. Generate samples + paths

→

4. Export Parquet + HF Hub

Step	Command	What it does
1	passage download	Download JAXA ALOS AW3D30 5°×5° tile archives with resume
2	passage calibrate	Scan tiles for global min/max elevation
3	passage generate	Create multi-resolution samples with paths, obstacles, tensors
4	passage export	Write Parquet shards, execute notebooks, push to HF Hub

📖 Full walkthrough: see the User Guide or the local USAGE.md reference.

---

🧭 Reproducibility

• splits.seed controls deterministic split assignment.
• generation.seed controls deterministic per-sample tile selection, crop placement, marker placement, and obstacle synthesis.
• Each sample metadata file records provenance.generation_seed and provenance.sample_seed.
• Runtime timestamps remain informational metadata and are not part of the deterministic sampling contract.
• make manifest writes outputs/export/manifest.sha256 so a released export tree can be checked for completeness.

Repository-root release metadata is tracked in CITATION.cff, DATASHEET.md, CHANGELOG.md, and croissant.json.

---

📦 Loading the Dataset

from datasets import load_dataset
from passage.utils import decode_tensor_blob

# Load a specific resolution from Hugging Face
ds = load_dataset("Anonym-2045/passage", name="256x256")
sample = ds["train"][0]

# Decode the compressed tensor → (256, 256, 3) numpy array
tensor = decode_tensor_blob(sample["tensor"], (256, 256, 3))
elevation = tensor[:, :, 0]  # Normalized [0, 1]
markers   = tensor[:, :, 1]  # -1 background, 0 start, 1 goal
obstacles = tensor[:, :, 2]  # Binary mask (0/1)

# Access solver paths
path_cols = [k for k in sample if k.startswith("path_")]
print("Solver paths:", path_cols)

---

⚙️ Configuration

Core settings live in config/config.yaml:

resolutions: [64, 128, 256, 512, 1024, 2048, 4096]
num_samples:
  64:    1000000
  128:   1000000
  256:   1000000
  512:   1000000
  1024:   250000
  2048:   100000
  4096:    50000
splits:
  ratios: {train: 0.90, calibration: 0.02, validation: 0.04, test: 0.04}
  seed: 42
generation:
  seed: 42
  preview_images:
    enabled: true
    resolutions: [64, 128, 256, 512, 1024, 2048, 4096]
    num_samples_per_split: 64
obstacles:
  enabled: true
  ratio_max: 0.30
  shape:
    n: { distribution: log_uniform, min: 1.0, max: 80.0 }
solvers:
  - name: astar_elevation
    backend: grid
    solver: astar
    cost_model: elevation
    weight: 5878.0

CLI options override config values. See the Configuration Reference in the documentation.

🔐 Hugging Face credentials

# .env
HF_USER_NAME=your-hf-username
HF_REPO_NAME=passage-dataset

# .env.secret (never commit!)
HF_TOKEN=hf_XXXXXXXXXXXXXXXXXXXXXXXX

---

📓 Notebooks

Notebook	Description
demo.ipynb	Quick-start: load, decode, and visualize a sample
costmodel.ipynb	Compare the configured terrain-cost variants on the same terrain
obstacles.ipynb	Explore superellipse obstacle generation
solve.ipynb	Solver timing and quality comparison
_dataset.ipynb	Template — auto-executed per resolution during export

---

🔬 Research Applications

Domain	Use Case
Classical AI	A*, Dijkstra benchmarking across cost models
Reinforcement Learning	Grid-based terrain navigation agents
Graph Neural Networks	Learned heuristics on elevation graphs
Surrogate Models	Fast path prediction from 4.4M+ multi-resolution terrain-routing examples
Computer Vision	Path segmentation from elevation images
Certified ML	ARP 6983 / ED-324 — bounded ODD, deterministic data

---

📁 Output Structure

outputs/
├── download/           # Tile archives + download.csv + extracted GeoTIFFs
├── calibrate/          # calibrate.json + elevation maps
├── dataset/            # Per-resolution samples
│   └── 256/
│       └── train/
│           ├── tensors/N045/E010/0000000000.npy.zst
│           ├── metadata/N045/E010/0000000000.yaml
│           ├── images/astar_energy/N045/E010/0000000000.png
│           └── paths/astar_energy/free/N045/E010/0000000000.csv
└── export/             # Hugging Face-ready parquet shards + release checksums
  ├── 256x256/train/data-00000.parquet
  ├── 256x256_samples/train/data-00000.parquet
  ├── README.md
  └── manifest.sha256

---

🧪 Development & Testing

# Install dev dependencies
uv sync --group dev

# Run tests (80%+ coverage required)
uv run pytest --cov

# Lint & format
uv run ruff check src/ tests/
uv run ruff format src/ tests/

# Pre-commit hooks
uv run pre-commit install
uv run pre-commit run --all-files

---

📝 Citation

Citation metadata is available in CITATION.cff. If you prefer BibTeX, use:

@inproceedings{anonymous2026passage,
  title     = {{PASSAGE}: A Real-Terrain Multi-Resolution Benchmark for
               Trustworthy Constrained Path Planning},
  author    = {Anonymous, Author and Anonymous, Author and Anonymous, Author},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS),
               Datasets and Benchmarks Track},
  year      = {2026},
  url       = {https://github.com/anonymous-org/passage}
}

---

Contributing

If you are interested in contributing to PASSAGE, start by reading the Contributing guide.

Contributions are welcome; the guide covers development setup, coding standards, pull requests, and issue reporting.

In brief:

# Fork & clone, then:
uv sync --group dev
uv run pre-commit install
git checkout -b feature/your-feature-name
# ... make changes ...
uv run pytest --cov
uv run ruff check src/ tests/

---

💬 Support

Need help? Here's where to go:

Channel	Link
Documentation	anonymous-org.github.io/passage
Bug Reports	GitHub Issues
Discussions & Q&A	GitHub Discussions
Email	anonymous@anonymous.invalid

Before opening an issue, please search existing issues and discussions to avoid duplicates.

---

👤 Maintainers

Name	Affiliation	Contact
Anonymous Author	Anonymous Group	anonymous@anonymous.invalid
Anonymous Author	Anonymous Group	anonymous@anonymous.invalid

---

🙏 Acknowledgments

• JAXA for the ALOS AW3D30 Digital Surface Model
• Hugging Face for the Datasets library and Hub infrastructure
• Anonymous Group for supporting open-source research in safety-critical AI

---

🧾 Disclosures

AI-assistance disclosure

Portions of this repository's source code, documentation, and notebooks were drafted with the assistance of an AI assistant (Anthropic Claude). All AI-generated material was reviewed, edited, and validated by the human authors listed in CITATION.cff, who take full responsibility for the released content. No AI-system co-authorship is asserted.

Third-party Python dependencies

Python packages declared in pyproject.toml and resolved in uv.lock are released by their respective upstreams under their own licenses, which may differ from the MIT license that covers PASSAGE's own source code. Users are responsible for reviewing and complying with each upstream's terms — in particular, GPU acceleration via the optional passage-eval[torch] extra relies on PyTorch and, when used with CUDA wheels, on NVIDIA CUDA runtime libraries (e.g. nvidia-cublas, nvidia-cuda-runtime, nvidia-cudnn) distributed under the proprietary NVIDIA Software License Agreement, not under MIT. PASSAGE itself does not redistribute these libraries.

Optional GPU / CUDA support

PASSAGE runs end-to-end on CPU — no GPU is required to download, generate, or evaluate the dataset. The uv.lock shipped with this repository resolves PyTorch from the official CPU-only index (https://download.pytorch.org/whl/cpu) and contains no nvidia-* / cuda-* packages. To enable GPU acceleration (only required by the optional surrogate-baseline tooling under passage-eval[torch]), install the CUDA build of PyTorch from the appropriate NVIDIA index after uv sync, for example:

uv sync --extra eval                          # installs CPU PyTorch via the eval extra
uv pip install --reinstall \
  --index-url https://download.pytorch.org/whl/cu128 \
  torch                                        # opt in to the CUDA build

By doing so you accept the NVIDIA license terms cited above for the CUDA runtime libraries pulled in transitively.

---

Security

Security reporting and handling procedures are documented in SECURITY.md.

If you discover a vulnerability, please use private reporting channels described in that file.

---

License

PASSAGE ships two distinct licensing surfaces and you should recognise which one applies to what you are doing:

Surface	License	What it covers
Generator and evaluator code	MIT License	every Python file in this repository — the dataset generator (passage), the evaluation contract (passage-eval), and the documentation source.
Generated elevation data	JAXA AW3D30 terms of use	every elevation tile that PASSAGE ingests, every sample whose elevation channel is derived from those tiles, and any model whose training data carries that channel. The full text mirrored at release time is in DATASET_LICENSE.md.

Respect JAXA's licensing terms. PASSAGE is built on the JAXA ALOS World 3D 30 m DSM (AW3D30). Any user of PASSAGE — or of any artefact derived from it — must respect the JAXA terms of use:

1. Cite the JAXA AW3D30 source in publications using PASSAGE (recommended: Takaku et al., "Updates of 'AW3D30' ALOS Global DSM with Other Open Access Datasets," ISPRS Archives, 2020).
2. Reproduce the AW3D30 attribution in any redistribution of the elevation channel or of any artefact directly derived from it.
3. Read and abide by the JAXA terms-of-use page linked above for the authoritative wording.

We sincerely thank JAXA for openly sharing AW3D30 with the research community.

---

📖 Documentation • 🤗 Dataset • 🐛 Issues • 💬 Discussions