Why PASSAGE¶
Constrained path planning on real terrain underlies routing for search-and-rescue, robotics and airborne advisory systems, where rescue-helicopter crews fly through mountain passes, valleys and ridgelines. Existing public benchmarks cover this only in part: synthetic grids give algorithmic control without geographic realism, and geographic data give real terrain without controllable obstacles or shared baselines (the NeurIPS 2026 paper, introduction).
PASSAGE is a dataset and benchmark for constrained path planning on real terrain: each scene is a square crop of the JAXA ALOS World 3D–30 m (AW3D30) digital surface model with a start cell, a goal cell and procedural obstacles. Each scene has six route variants: three cost objectives (elevation, slope_uphill, energy), each on free terrain and with the obstacles blocked. Each variant carries an exact minimum-cost reference route for the stated graph (eight-connected, 30 m cells); start and goal cells that the obstacles disconnect are rejected by the generator. The dataset, the generator and the passage-eval evaluator are publicly released (code under MIT, data under the JAXA terms), so that classical planners, learned surrogates, graph networks and reinforcement learning can be scored on the same tasks.
From terrain to benchmark¶
PASSAGE v1.0 at a glance¶
| Item | Value |
|---|---|
| Version | 1.0, Hub tag v1.0 at commit c09cc5a21a55a93e7c09d47368d5f5bf75cb8fae |
| Core scenes | 4,400,000 at seven resolutions, from 64×64 to 4096×4096 cells |
| Route variants | 26,400,000, all exact routes (six per scene) |
| Splits | train / calibration / validation / test = 90 / 2 / 4 / 4 % |
| Test geography | South America hold-out; other tiles split by 5°×5° blocks |
| Preview rows | 1,792 (64 per split and resolution), with rendered images |
| Size | about 1.27 TB of Parquet (5,997 shards); 1.31 TB with receipts and notebooks |
| Intended tasks | Classical planning, path-mask segmentation, cost-to-go regression, graph and reinforcement-learning planners |
| Licence | Code: MIT. Data: JAXA AW3D30 terms, see DATASET_LICENSE.md |
| Paper | PASSAGE: A Real-Terrain Benchmark for Trustworthy Constrained Path Planning, NeurIPS 2026, Track on Evaluations and Datasets |
Illustrative validation previews are optional and are not included in every source distribution.
What you get¶
Seven grid sizes¶
From 64 × 64 to 4096 × 4096 cells: about 1.9 km to 123 km across at the nominal 30 m.
Real elevation¶
Every scene is a crop of AW3D30, JAXA's global digital surface model at a nominal 1-arcsecond (about 30 m) sampling (Takaku et al., 2020).
Procedural obstacles¶
Super-ellipses of varied shape, size and rotation, up to 30 % of a crop, never on the start or the goal.
Exact references¶
A Numba-compiled float64 grid A* labels every variant with an exact optimum; the generator rejects start and goal cells that the obstacles disconnect. A published notebook solves a sample again.
Bounded domain¶
The scene parameters form an operational design domain, a term the paper borrows from the EASA AI Concept Paper, Issue 02; the paper claims no compliance with EASA or FAA guidance.
Tested¶
865 tests with an 80 % coverage gate, deterministic seeding, pre-commit hooks and continuous integration.
Leaderboard¶
No entries in this index
The checked-in leaderboard index has no entries. The paper's measurements are reported separately.
Quick start¶
Most users only need the released data. Install datasets, zstandard and numpy, pin the dataset revision and decode one scene:
import numpy as np
import zstandard
from datasets import load_dataset
REVISION = "c09cc5a21a55a93e7c09d47368d5f5bf75cb8fae" # verified PASSAGE v1.0 upload
rows = load_dataset("thalesgroup/passage", "256x256_samples",
split="validation", revision=REVISION)
row = rows[0]
n = row["metadata"]["resolution"]
x = np.frombuffer(zstandard.ZstdDecompressor().decompress(row["tensor"]),
dtype="<f4").reshape(n, n, 3) # elevation, markers, obstacles
route = np.asarray(row["path_astar_energy_obstacles"]) # ordered [row, col] cells
The Dataset tab of this site shows how to recover metres, read the route status and stream the large configurations.
Rebuild the dataset from the JAXA tiles
The AW3D30 inventory holds 23,993 tiles, and the shipped configuration requests 1,000,000 scenes at 256 × 256. Follow the generation guide for the source checkout, terrain download, calibration and required geographic hold-out mask. Start with a trial generation before committing storage and compute to a full release.
Research uses¶
| Domain | Use | What PASSAGE provides |
|---|---|---|
| Classical search | A*, Dijkstra and anytime planners against exact references | Three cost objectives, obstacles, seven grid sizes |
| Reinforcement learning | Terrain navigation agents | Grid observations; exact optimal costs for rewards and regret |
| Graph neural networks | Learned heuristics | Eight-connected graphs from elevation grids, route labels |
| Surrogate models | Fast approximations of optimal routes | 4.4 M scenes and 26.4 M route variants |
| Computer vision | Route prediction from elevation images | (N, N, 3) tensors and route masks |
Cite PASSAGE¶
PASSAGE is accepted as a poster at the 40th Conference on Neural Information Processing Systems (NeurIPS 2026), Track on Evaluations and Datasets. Please cite the paper, and name the dataset version you used (PASSAGE v1.0):
@inproceedings{delhomme2026passage,
title = {{PASSAGE}: A Real-Terrain Benchmark for Trustworthy Constrained Path Planning},
author = {Delhomme, Geoffrey and Trap, Fran{\c{c}}ois and Kaakai, Fateh},
booktitle = {Advances in Neural Information Processing Systems (NeurIPS), Track on Evaluations and Datasets},
year = {2026},
url = {https://github.com/ThalesGroup/passage}
}
Future reference
The NeurIPS 2026 proceedings are not published yet. This entry will be completed with the proceedings details (volume, pages and paper URL) once they are available; the title, authors, venue and year will not change. Machine-readable metadata is in CITATION.cff.
Learn more¶
Generation process¶
Sampling, cost models, obstacles and the reference solver, with the mathematics.
Acknowledgments¶
- JAXA for the ALOS World 3D–30m (AW3D30) digital surface model.
- Hugging Face for the Datasets library and Hub infrastructure.
- Thales Group for supporting this open research.
JAXA license and attribution¶
The PASSAGE code is released under the MIT License. The elevation data every PASSAGE scene is built on are provided by JAXA under the JAXA Terms of Use of Research Data; the dataset is neither MIT-licensed nor public domain. Anyone who uses or redistributes PASSAGE data, or publishes derivatives of them, must follow those terms:
JAXA AW3D30 — user obligations
- Credit JAXA and AW3D30, including in any redistribution or published derivative, with the line: "The original data used for this product have been supplied by JAXA's ALOS World 3D–30m (AW3D30)."
- Notify JAXA before any commercial use. JAXA also asks to be notified, where possible, of publications that use the data.
- Read the JAXA terms of use linked above for the authoritative wording.
When you publish results on PASSAGE, we recommend citing the AW3D30 source with the PASSAGE paper: J. Takaku, T. Tadono, M. Doutsu, F. Ohgushi and H. Kai, "Updates of 'AW3D30' ALOS Global Digital Surface Model with Other Open Access Datasets", ISPRS Archives XLIII-B4-2020, 183–189, 2020, doi:10.5194/isprs-archives-XLIII-B4-2020-183-2020.
A summary of these terms is in DATASET_LICENSE.md; the JAXA terms of use remain the authoritative text. The Dataset tab lists the third-party elevation sources that AW3D30 uses to fill voids.