🌍 PASSAGE — Real-Terrain Benchmark & Generator

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

🤗 Dataset • 🧾 Croissant • 📋 Datasheet • 📝 Citation

---

🚁 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, energy, slope-uphill
🧮 Grid A* Oracle Exact labels + timings	🤗 HuggingFace-Ready Parquet + one-command upload	🔁 Fully Resumable No wasted compute	🛡️ Certification-Aware ARP 6983 / ED-324

14,130 Default Samples

7 Resolution Levels

3 A* Configurations

1 Solver Backend

30 m/px Spatial Resolution

---

🗺️ Data Overview

Representative Sample

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

Dataset World Coverage

The dataset leverages the worldwide elevation data publicly provided by JAXA. Each cell below represents a 5×5 grid of 1°×1° tiles (3600×3600 pixels at 30 m resolution). Black cells indicate regions with no available tile — these are treated as flat tiles at sea level during dataset generation.

Maximum Elevation Map

This visualization shows the normalized maximum elevation within each downloaded tile. Brighter regions indicate tiles containing higher terrain (mountain peaks, elevated plateaus); darker regions correspond to low-lying areas (coastal zones, river basins).

📈 Elevation Range

According to the calibration across all tiles:

• Global minimum elevation: −430.0 m
• Global maximum elevation: 8 767.0 m

Detailed statistics are available in calibrate.json.

Peak-to-Peak Elevation Map

The peak-to-peak (PTP) elevation map displays the elevation range (max − min) within each tile. High PTP values (bright regions) indicate significant topographic variation — suitable for challenging pathfinding scenarios. Low PTP values suggest relatively flat terrain.

Hold-Out Region

To assess generalization of trained models, a region containing South America is held out. This region was chosen because it concentrates a wide range of elevations representative of all other places in the world map.

---

🚀 Quick Start

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("thalesgroup/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)

Manual Tensor Decompression

import numpy as np
import zstandard as zstd

resolution = sample["metadata"]["resolution"]

# Decompress the tensor
d = zstd.ZstdDecompressor()
b = d.decompress(sample["tensor"])
tensor = np.frombuffer(b, dtype=np.float32).reshape((resolution, resolution, 3))

# Unravel channels
elevation = tensor[:, :, 0]
markers   = tensor[:, :, 1]
obstacles = tensor[:, :, 2]

# Denormalize elevation to meters
global_min = sample["metadata"]["calibration"]["min"]
global_max = sample["metadata"]["calibration"]["max"]
elevation_m = (elevation * (global_max - global_min)) + global_min

See notebooks/demo.ipynb for a complete loading & visualization example.

---

📦 Dataset Structure

Each sample consists of:

Field	Type	Description
tensor	bytes	zstd-compressed (H, W, 3) float32 tensor payload
path_<solver>_free	list[[i, j], ...]	Path without obstacles for each configured solver
path_<solver>_obstacles	list[[i, j], ...]	Path with obstacles for each configured solver
image_<solver>	image	Optional per-solver visualization image (sample configs only)
metadata	dict	Sample metadata (coordinates, elevations, stats)

Tensor Channels

1. Channel 0 (Elevation): Normalized elevation in $[0, 1]$
2. Channel 1 (Markers): $-1$ everywhere, $0$ at start, $1$ at goal
3. Channel 2 (Obstacles): Binary mask ($0$ = free, $1$ = obstacle)

Note: Tensors are stored as compressed .npy.zst blobs.

---

🔁 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 notebooks/demo.ipynb, DATASHEET.md, and croissant.json.

---

📓 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

Per-resolution analysis notebooks are included with each configuration:

• 64×64: In-depth analysis notebook
• 128×128: In-depth analysis notebook
• 256×256: In-depth analysis notebook
• 512×512: In-depth analysis notebook
• 1024×1024: In-depth analysis notebook
• 2048×2048: In-depth analysis notebook
• 4096×4096: In-depth analysis notebook

---

🎯 Sampling Procedure

Each sample is generated through a carefully designed stochastic process ensuring diversity across elevations, geographic locations, and terrain types.

Step 1 — Target Elevation Sampling

A target elevation $e_{\text{target}}$ is sampled uniformly from the global calibrated range:

$$e_{\text{target}}\sim \mathcal{U}(e_{\min },e_{\max })e\_\{\backslash text\{target\}\}\; \backslash sim\; \backslash mathcal\{U\}(e\_\{\backslash min\},\; e\_\{\backslash max\})$$

where $e_{\min}$ and $e_{\max}$ are computed during calibration across all available tiles.

Step 2 — Tile Selection

From the set of tiles $\mathcal{T}$ containing the target elevation, a tile $T$ is selected uniformly at random:

$$T\sim \mathcal{U}(\{t\in \mathcal{T}:e_{\min }^{}\le e_{\text{target}}\le e_{\max }^{}\})T\; \backslash sim\; \backslash mathcal\{U\}\backslash left(\backslash \{t\; \backslash in\; \backslash mathcal\{T\}\; :\; e\_\{\backslash min\}^\{(t)\}\; \backslash leq\; e\_\{\backslash text\{target\}\}\; \backslash leq\; e\_\{\backslash max\}^\{(t)\}\backslash \}\backslash right)$$

Step 3 — Pixel Localization

Within the selected tile, the pixel $(i^*, j^*)$ closest to $e_{\text{target}}$ is identified:

$$(i^{\ast },j^{\ast })=\arg \underset{(i,j)}{\min }\mid \text{elevation}(i,j)-e_{\text{target}}\mid (i^*,\; j^*)\; =\; \backslash arg\backslash min\_\{(i,j)\}\; \backslash left|\; \backslash text\{elevation\}(i,j)\; -\; e\_\{\backslash text\{target\}\}\; \backslash right|$$

Step 4 — Crop Extraction

A crop of size $R \times R$ (where $R$ is the resolution) is randomly positioned such that the target pixel $(i^*, j^*)$ lies within the crop. The crop position $(y_1, x_1)$ is sampled as:

$$y_1\sim \mathcal{U}[\max (0,i^{\ast }-R+1),\min (H-R,i^{\ast })]y\_1\; \backslash sim\; \backslash mathcal\{U\}\backslash left[\backslash max(0,\; i^*\; -\; R\; +\; 1),\; \backslash min(H\; -\; R,\; i^*)\backslash right]$$ $$x_1\sim \mathcal{U}[\max (0,j^{\ast }-R+1),\min (W-R,j^{\ast })]x\_1\; \backslash sim\; \backslash mathcal\{U\}\backslash left[\backslash max(0,\; j^*\; -\; R\; +\; 1),\; \backslash min(W\; -\; R,\; j^*)\backslash right]$$

Step 5 — Marker Placement

Two markers define the pathfinding problem:

• Marker 1 (start): The target pixel $(i^*, j^*)$ within crop coordinates
• Marker 2 (goal): A uniformly sampled discrete pixel from the $R^2-1$ remaining cells:

$${\text{Marker}}_2\sim {\mathcal{U}}_{\text{discrete}}(\{(i,j)\in [0,R{)}^2:(i,j)\mathrm{\ne }{\text{Marker}}_1\})\backslash text\{Marker\}\_2\; \backslash sim\; \backslash mathcal\{U\}\_\{\backslash text\{discrete\}\}\backslash left(\backslash \{(i,j)\; \backslash in\; [0,R)^2\; :\; (i,j)\; \backslash neq\; \backslash text\{Marker\}\_1\backslash \}\backslash right)$$

Step 6 — Elevation Normalization

The elevation crop is normalized to $[0, 1]$ using global calibration statistics:

$$\widehat{e}(i,j)=\frac{e(i,j)-e_{\min }}{e_{\max }-e_{\min }}\backslash hat\{e\}(i,j)\; =\; \backslash frac\{e(i,j)\; -\; e\_\{\backslash min\}\}\{e\_\{\backslash max\}\; -\; e\_\{\backslash min\}\}$$

---

🧩 Obstacle Generation

Obstacles are procedurally generated using superellipse shapes (Lamé curves), allowing a rich variety of geometries — from circular to rectangular and everything in between.

Superellipse Equation

A superellipse centered at $(c_x, c_y)$ with semi-axes $a$, $b$ and exponent $n$ is defined by:

$${\left(\frac{\mathrm{\mid }x-c_x\mathrm{\mid }}{a}\right)}^n+{\left(\frac{\mathrm{\mid }y-c_y\mathrm{\mid }}{b}\right)}^n\le 1\backslash left(\backslash frac\{|x\; -\; c\_x|\}\{a\}\backslash right)^n\; +\; \backslash left(\backslash frac\{|y\; -\; c\_y|\}\{b\}\backslash right)^n\; \backslash leq\; 1$$

The exponent $n$ controls the shape:

• $n = 2$: Standard ellipse (circle when $a = b$)
• $n < 2$: Diamond/star shapes ($n = 1$ gives a rhombus)
• $n > 2$: Rounded rectangles (approaches a rectangle as $n \to \infty$)

Obstacle Parameters

Parameter	Distribution	Description
Position $(c_x, c_y)$	$\mathcal{U}([0, W) \times [0, H))$	Center coordinates
Size $(w, h)$	$\mathcal{U}([s_{\min} \cdot R, s_{\max} \cdot R])$	Width and height based on size ratios
Exponent $n$	Configurable (default: log-uniform)	Shape parameter
Angle $\theta$	$\mathcal{U}([0, 2\pi))$	Rotation angle

where $s_{\min}$ and $s_{\max}$ are configurable size ratios relative to resolution $R$.

Target Coverage

A target obstacle ratio $r_{\text{target}}$ is sampled uniformly:

$$r_{\text{target}}\sim \mathcal{U}(0,r_{\max })r\_\{\backslash text\{target\}\}\; \backslash sim\; \backslash mathcal\{U\}(0,\; r\_\{\backslash max\})$$

Obstacles are iteratively added until coverage approaches $r_{\text{target}}$, while ensuring no obstacle overlaps with the start or goal markers.

---

⚖️ Exported Path Cost Models

The current paper benchmark exports three grid-backend A* configurations. Each edge weight uses physical step distance $d$ (30 m cardinal, $30\sqrt{2}$ diagonal), weight $\alpha$, and elevation values from the current crop:

• elevation

$$w=d\cdot (1+\alpha \cdot {\bar{e}}_{\text{dst}}),{\bar{e}}_{\text{dst}}=\frac{e_{\text{dst}}-e_{\min }}{e_{\max }-e_{\min }}w\; =\; d\; \backslash cdot\; \backslash left(1\; +\; \backslash alpha\; \backslash cdot\; \backslash bar\{e\}\_\{\backslash text\{dst\}\}\backslash right),\; \backslash quad\; \backslash bar\{e\}\_\{\backslash text\{dst\}\}\; =\; \backslash frac\{e\_\{\backslash text\{dst\}\}\; -\; e\_\{\backslash min\}\}\{e\_\{\backslash max\}\; -\; e\_\{\backslash min\}\}$$

• slope_uphill

$$w=d\cdot (1+\alpha \cdot \frac{\max (0,e_{\text{dst}}-e_{\text{src}})}{e_{\max }-e_{\min }})w\; =\; d\; \backslash cdot\; \backslash left(1\; +\; \backslash alpha\; \backslash cdot\; \backslash frac\{\backslash max(0,\; e\_\{\backslash text\{dst\}\}\; -\; e\_\{\backslash text\{src\}\})\}\{e\_\{\backslash max\}\; -\; e\_\{\backslash min\}\}\backslash right)$$

• energy

$$w=d+\alpha \cdot \max (0,e_{\text{dst}}-e_{\text{src}})w\; =\; d\; +\; \backslash alpha\; \backslash cdot\; \backslash max(0,\; e\_\{\backslash text\{dst\}\}\; -\; e\_\{\backslash text\{src\}\})$$

These equations match the runtime implementation used by the grid-backend A* exporter.

---

🔬 Research Applications

Domain	Use Case	Relevant Features
Classical AI	A* benchmarking	Multi-resolution labels, timings, terrain-cost variants
Reinforcement Learning	Terrain navigation agents	Grid observations, reward shaping with configured cost models
Graph Neural Networks	Learned heuristics	Graph structure from elevation grids, path labels
Surrogate Models	Fast approximation of optimal paths	Input/output pairs at scale (14 K+ samples)
Computer Vision	Path prediction from elevation images	Tensor format, multi-channel images
Certified ML	ARP 6983 / ED-324 compliance	Bounded ODD, deterministic generation, traceability

---

📝 Citation

Use the metadata in CITATION.cff. Anonymous submission mirrors intentionally omit author-identifying citation fields until de-anonymisation.

---

📄 License

The PASSAGE generator code is MIT-licensed, but the exported dataset artifacts inherit attribution and reuse obligations from JAXA ALOS AW3D30. See DATASET_LICENSE.md for the dataset terms and LICENSE for the code license.