skyfull-policy-chamonix-rescue-v1

skyfull-policy-chamonix-rescue-v1

Rescue drone trajectory policy trained on real Chamonix terrain. Part of Skyfull — physical intelligence over geo-scale terrain.

Architecture

Tiny MLP policy. No vision encoder. No transformer. No diffusion. Just coordinate geometry learned end-to-end.

Also includes terrain_chamonix.pt — the TerrainNet implicit neural terrain (4k params, 37 KB) encoding the Chamonix DEM as a queryable function.

Training

Performance

Quick start

import torch, numpy as np
from huggingface_hub import hf_hub_download

ckpt  = torch.load(hf_hub_download("Ethgar/skyfull-policy-chamonix-rescue-v1", "policy_chamonix.pt"), weights_only=False)
norm  = ckpt['norm']

import torch.nn as nn
class SkyfullPolicy(nn.Module):
    def __init__(self, hidden=256, depth=3, in_dim=3, out_dim=120):
        super().__init__()
        layers = [nn.Linear(in_dim, hidden), nn.SiLU()]
        for _ in range(depth - 1):
            layers += [nn.Linear(hidden, hidden), nn.SiLU()]
        layers.append(nn.Linear(hidden, out_dim))
        self.net = nn.Sequential(*layers)
    def forward(self, x): return self.net(x)

model = SkyfullPolicy(hidden=ckpt['hidden'], depth=ckpt['depth'])
model.load_state_dict(ckpt['state_dict'])
model.eval()

city_base = np.array(norm['city_base'], dtype=np.float32)  # [6.87, 45.924]
victim = np.array([6.9200, 45.9740, 2200.0], dtype=np.float32)

x_raw  = np.array([victim[0]-city_base[0], victim[1]-city_base[1], victim[2]], dtype=np.float32)
x_norm = (x_raw - norm['x_mean']) / norm['x_std']

with torch.no_grad():
    y_norm = model(torch.from_numpy(x_norm).unsqueeze(0)).numpy()[0]

traj = y_norm.reshape(40, 3) * norm['y_std'] + norm['y_mean']
print(traj.shape)   # (40, 3) — lon, lat, alt_m per waypoint

Tile coverage

Chamonix, Haute-Savoie, France. lat [45.86, 45.975] · lon [6.8, 6.99]

Specialist model for this tile only. Other regions: run pipeline.py <tile_id>.

Skyfull tile library