---
license: mit
task_categories:
  - robotics
tags:
  - drone
  - rescue
  - trajectory-planning
  - alpine
  - geospatial
  - xyz-semantic
  - skyfull
  - lerobot-compatible
size_categories:
  - 10K<n<100K
---

# skyfull-zermatt-rescue-v1

Rescue drone trajectory dataset over **real Zermatt terrain** (Valais, Switzerland).
Generated with [Skyfull](https://github.com/Ethgar/skyfull) — a geo-scale simulation
platform for physical intelligence.

## What makes this different

No pixels. No camera. Pure XYZ coordinates.

The terrain is encoded as an implicit neural function (`TerrainNet`), not a texture or image.
Every trajectory is a sequence of `(lon, lat, alt_m)` waypoints computed analytically,
then used as supervised signal for a tiny policy MLP.

This is the same principle as [LeRobot](https://github.com/huggingface/lerobot) —
collect → train → deploy — but the "lab" is a real Alpine mountain range and the
"robot" is a rescue drone.

## Tile

| | |
|---|---|
| Region | Zermatt, Valais, Switzerland |
| Bounds | lat [45.95, 46.07]  lon [7.68, 7.82] |
| Terrain source | AWS Terrarium tiles, zoom 12 (~26 m/px) |
| Terrain model | TerrainNet — 4k-param MLP, 37 KB |
| City base | Zermatt (7.7480°, 46.0200°) |

## Dataset

| | |
|---|---|
| Episodes | 10,000 |
| Steps per episode | 40 |
| Observation | (lon, lat, alt_m) — 3 floats per step |
| Trajectory shape | (40, 3) float32 |
| Victim min elevation | 2200 m |
| Terrain clearance | 80 m minimum |
| Total size | ~5 MB |
| Terrain violations | 0 / 10,000 |

## Files

| File | Shape | Description |
|---|---|---|
| `episodes.npy` | `(10000, 40, 3)` float32 | Full trajectories: lon, lat, alt_m per step |
| `victims.npy` | `(10000, 3)` float32 | Victim locations: lon, lat, elev_m |
| `dataset_info.json` | — | Schema, tile metadata, generation params |

## Quick start

```python
import numpy as np
from huggingface_hub import hf_hub_download

episodes = np.load(hf_hub_download("Ethgar/skyfull-zermatt-rescue-v1", "episodes.npy", repo_type="dataset"))
victims  = np.load(hf_hub_download("Ethgar/skyfull-zermatt-rescue-v1", "victims.npy",  repo_type="dataset"))

print(episodes.shape)   # (10000, 40, 3)
print(victims.shape)    # (10000, 3)
```

## Companion model

Trained policy: [`Ethgar/skyfull-policy-zermatt-rescue-v1`](https://huggingface.co/Ethgar/skyfull-policy-zermatt-rescue-v1)

Input: `(victim_dlon, victim_dlat, victim_elev)` — 3 floats
Output: full trajectory `(40, 3)` — planned in < 2 ms on CPU

## Skyfull tile library

| Tile | Region | Repo |
|---|---|---|
| Aosta Valley | Valle d'Aosta, Italy | [Ethgar/skyfull-aosta-valley-rescue-v1](https://huggingface.co/datasets/Ethgar/skyfull-aosta-valley-rescue-v1) |
| Chamonix | Haute-Savoie, France | [Ethgar/skyfull-chamonix-rescue-v1](https://huggingface.co/datasets/Ethgar/skyfull-chamonix-rescue-v1) |
| Zermatt | Valais, Switzerland | [Ethgar/skyfull-zermatt-rescue-v1](https://huggingface.co/datasets/Ethgar/skyfull-zermatt-rescue-v1) |
| Dolomites | South Tyrol, Italy | [Ethgar/skyfull-dolomites-rescue-v1](https://huggingface.co/datasets/Ethgar/skyfull-dolomites-rescue-v1) |

## Cite

```
@misc{skyfull-zermatt-rescue-v1,
  author       = {Ethgar},
  title        = {Skyfull: Zermatt rescue drone dataset},
  year         = {2026},
  publisher    = {HuggingFace},
  howpublished = {\url{https://huggingface.co/datasets/Ethgar/skyfull-zermatt-rescue-v1}},
}
```