Hangar readme · solar case study
From polygon to QGC draft
in one hangar pass
This page is the operator README. Follow the solar-farm case study end to end — plain site polygon, or perception detections (YOLO / demo / VLM) — then export a draft mission. Everything here is experimental. You fly it; you own the risk.
1. The story
You run a mid-size solar site. You want two to four drones to lawnmower the array, skip a truck parked on a string, and mark a thermal hotspot for a closer look — then open the result in QGroundControl.
- Accept the liability gate on the homepage.
- Choose vertical preset
solarand a schedule mode (sortiedefault, orcolumnsto pack sorties). - Either draw/upload a site polygon or load perception detections.
- Re-confirm the disclaimer and hit customize / plan.
- Read
SAFETY_REPORT.md, then download QGC plans.
2. Path A — site polygon only
Best when you already have a boundary and do not need CV.
- Open Hangar planner.
- Set Vertical preset to
solar(≈5 m cells, mid altitude). - Load Demo solar farm, draw a polygon, or upload a
GeoJSON boundary
(repo sample:
examples/sites/solar_farm.geojson). Crop fields are on the fields page, grouped by state and crop (California DWR underexamples/sites/dwr/2023/, USDA sample underexamples/sites/csb/2024/). Check several blocks to plan a ranch; each field keeps its own flight lines. - Optional: add Point features with
"role": "depot"or"depot": truefor hangar placement. - Leave Schedule mode on
sortie, or pickcolumnsto pack sorties via LP pricing + set-partitioning (slower; useful when recharge interleaving looks wasteful). - Optional: check Pull OSM buildings as keep-outs to punch Overpass footprints out of the site (verify every polygon; a failed fetch still plans without them).
- Check I re-confirm the disclaimer → Customize mission.
- Wait for metrics (coverage, makespan, drones, partition family, schedule) and the path preview.
Fair-use grid cap on the public demo keeps jobs honest — not a paywall.
Local installs can raise FLEET_MAX_CELLS.
3. Path B — perception → mission
Best when YOLO / a thermal model / a VLM already labeled the site. Roles: cover avoid inspect depot
- Scroll to Perception Lab.
- Pick ontology
solar / infra. -
Load demo detections for a one-click walkthrough,
or upload a detections FeatureCollection
(sample:
examples/perception/solar_detections.geojson), or type a command likesurvey panels, avoid trucksand run NL. - Confirm cover / avoid / inspect overlays on the perception map.
- With the planner disclaimer still checked, hit Plan from perception.
- Read the perception report JSON — class counts are not a clearance.
Full adapter docs (YOLO, ONNX, llama.cpp / vLLM, domain plugins): docs/PERCEPTION.md
4. Export & preflight
- Download SAFETY_REPORT.md first — every checklist item is yours.
- Grab QGC plan — drone N files (one per scheduled aircraft).
- Optional: Missions GeoJSON, waypoints CSV, or the full job zip.
- In QGroundControl: inspect home / RTH, altitude mode, geofence, and every leg.
- Discrete sim deconfliction is not certified airborne separation — still separate yourself.
Case-study success looks like
- Coverage of reachable grid cells near 100% (or an explained skip list).
- Avoid zones empty of free cells (truck / crane punched out).
- Inspect points visible in the bundle for the hotspot follow-up.
- You still treat the export as a draft until GCS review is done.
What we measured (2026-10-03)
On this industrial polygon the solar battery is roomy (commute about 2%),
so auto and a pure lawnmower tie: same 458 s makespan,
flight within 0.2%. Pattern mixing pays on the bench matrix instead:
auto flies 15.8% less than boustro-only across 12 map/seed
pairs, and about 21–24% less on cluttered and 100×100 maps.
Numbers: benchmarks/lawnmower.md.
5. Same case study on the CLI
Mirror of Path B for scripts and CI:
git clone https://github.com/mmorri/pimpmydrone.git cd pimpmydrone make build cd python && uv sync # Perception → site → plan uv run python -m fleet perception ingest \ --input ../examples/perception/solar_detections.geojson \ --output /tmp/site.geojson --ontology solar uv run python -m fleet perception plan-from-perception \ --input ../examples/perception/solar_detections.geojson \ --output-dir ../outputs/case-solar --vertical solar # Or polygon-only (auto picks region vs strip) uv run python -m fleet from-geo \ --geojson ../examples/sites/solar_farm.geojson \ --output-dir ../outputs/site-solar --vertical solar \ --partition auto --schedule-mode columns --osm-buildings
Local UI: make serve → http://127.0.0.1:8000.
6. What to try next
- Crop fields — California orchards and a northern-plains sample, grouped by state and crop.
- Security ontology — person / vehicle inspect targets on a fence corridor.
- Thermal ontology — smoke / flame / hotspot priorities.
- Wire a real model: set
FLEET_YOLO_MODELorFLEET_VLM_URL(see perception docs). - Tip the shop if it saved you a sortie — /donate (not a subscription).