A county-scale autonomous drone fleet that detects wildfires before human spotters can.
The first 30 minutes decide whether a fire stays under an acre or burns 1,000 structures. Satellites and 911 calls miss that window in the wildland-urban interface; this project is designed to fill the gap.
A pre-flight Python monorepo spanning synthetic-data ML pipelines, a kinematic flight simulator, swarm consensus, and TAK/CoT interoperability. The target AOR is Gunnison Valley + Crested Butte, Colorado.
Status: simulation and software only. Zero flight hours. No trained production ML model.
# Install
pip install -e ".[dev]"
# Run full test suite (under 10 seconds)
python3 -m pytest -q
# Fly one drone with a synthetic plume
python3 -m sim.cli run sim/missions/gunnison_slate_river_1km2.yaml \
--scenario single_smoke_plume --speed-multiplier 5
# View the flight in browser
python3 -m sim.web.server # http://127.0.0.1:8088
# Valuation snapshot
python3 -m valuation.cli snapshotml/fire_detection/ Synthetic data, YOLOv8 training, inference gate
sim/ Kinematic flight simulator + swarm + perception
frontend/ Admin dashboard (Flask) — Cloud Run target
valuation/ Intrinsic-value engine + KPI dashboard
sapphire_integration/ Schema, TAK/CoT emitter, Foundry adapter
ground_station/ Raspberry Pi telemetry collectors
hardware/ Phased BOM with Blue-UAS-substitutable parts
missions/zones/ AOR zones with wilderness exclusions
- Kinematic flight simulator — deterministic seeding, JSONL logs, browser viewer
- Multi-drone swarm + k-of-N consensus — lossy-comms model, fused risk scoring
- GNSS-denied navigation primitive — VO + TRN + IMU fusion with spoof detection
- TAK / Cursor-on-Target emitter — 8 type-code mappings over TCP/UDP/TLS/multicast
- Continuous valuation engine — 4-method valuation band (comp, venture, DCF-lite, asset-floor)
- Blue-UAS-substitutable BOM — 3D-printable airframe, open parts list
- Python 3.11+ (stdlib-first in
sim/) - Flask 3.x, Leaflet, Chart.js for frontend
- Ultralytics YOLOv8 (lazy-loaded), OpenCV (lazy-loaded)
| Phase | Cost | Mission |
|---|---|---|
| Phase 0 | $0 | DJI Mavic Mini + simulator-only autonomy |
| Phase 0.5 | $215 | + RTL-SDR, sensors, LoRa mesh, edge YOLOv8 |
| Phase 1 | $2,613 | Holybro X500 + Jetson Orin Nano + RGB/LWIR fusion |
- Simulation-only — no real flight operations without FAA authorization
- Wilderness geofences — West Elk, Maroon Bells-Snowmass, Raggeds are no-fly per 36 CFR 261.16
- NDAA compliance path — DJI is a Phase 0 stopgap; Holybro/Cube/Jetson is the target stack
- No active fire penetration; detection and notification only
See AGENTS.md for monorepo architecture, safety boundaries, test requirements, and deployment procedures.