m1sc/reach-down-tiles-v2

Reach-down training dataset v2

The exact 2,500-tile dataset used to train the released reach-down fixed-wing landing-value ViT: 500 synthetic terrains, with five aircraft states per terrain. Original compressed NumPy files are preserved, totaling 13.33 GB. The model release contains the complete model weights, training histories, reconstructed training arguments, evaluation CSVs, and method documentation.

Files and split

• tile_00000.npz through tile_02499.npz: original training tiles.
• split.csv: per-tile split and terrain, aircraft, wind, noise, and state metadata.
• SHA256SUMS: SHA-256 checksums of every tile and the original split manifest.
• release.json: generation and training provenance plus source snapshot.

The split is by terrain: seeds 0–399 supply 2,000 training tiles; seeds 400–499 supply 500 validation tiles. All five states of a terrain remain in the same split. Additional evaluations on disjoint synthetic terrains and real DEMs are separate from this training archive; their CSVs are in the model release.

Generation

The generation command recorded in the repository release documentation is:

python scripts/make_dataset.py --n-terrains 500 --states 5 --out data/tiles_v2

The seed defaults to 0. Each terrain is a 512 × 512 heightfield at 10 m per pixel. The generator combines spectral fractal relief, flattened valleys, rock fields, and graded landing pads, with independent avoid masks. Terrain relief and spectral statistics, airframe, landing limits, aircraft state, adversarial wind bounds (0, 3, or 6 m/s), and correlated elevation-observation noise are randomized. Inputs use noisy observed terrain; privileged labels use the clean terrain and the Hamilton–Jacobi reach-avoid oracle. Full details and code are in docs/datagen.md and the source repository.

Tile schema

Every tile contains eight 512 × 512 arrays and a scalar JSON meta record:

field	meaning
z_obs	observed elevation, metres
slope_obs	slope of the observed terrain, degrees
rough_obs	roughness of the observed terrain, metres
hazard	avoid mask
margin_analytic	analytic glide reach margin, metres
margin_hj	privileged Hamilton–Jacobi reach margin, metres
label_value	unbuffered joint reach-avoid value target
label_run	hazard-masked landability target
meta	terrain seed, pixel size, platform and jittered limits, wind, noise, terrain statistics, and aircraft state

The released model uses seven normalized input channels and resizes them to 364 × 364. Training recomposes the target from label_run and margin_hj with a 40 m safety buffer; the stored label_value is the original unbuffered target. See reachdown/data.py in the source repository. Training arguments were reconstructed after the original run, as documented in each args.json; the dataset files and split are the original artifacts.

Download and read

import json
import numpy as np
from huggingface_hub import snapshot_download

root = snapshot_download(
    repo_id="m1sc/reach-down-tiles-v2",
    repo_type="dataset",
    local_dir="data/tiles_v2",
)
with np.load(f"{root}/tile_00000.npz", allow_pickle=False) as tile:
    meta = json.loads(tile["meta"].item())
    value = tile["label_value"]

cd data/tiles_v2
sha256sum -c SHA256SUMS