OpenAIO

NIDAR RescueSwarm

Three autonomous drones that search a flood zone, find survivors, geotag them,
and drop medical kits — with no network, no pilot, and one operator who only presses start.

NIDAR 2026–27 · Track 1 · Mission 1
_{MeitY · Drone Federation of India · SwaYaan}

1. Status

Phase 0/1. Sized end to end, requirements baselined against the rulebook, and the ground station runs as a system against three real ArduCopter SITL autopilots.

What exists: coverage planner, geotag geometry and its error model, ArduPilot parameter sets, safety-link protocol, ground station. All tested — 129 + 17 + 67 tests, 13 committed model outputs.

What does not: the aircraft, and the detector. There is no camera anywhere in the loop; survivors have only ever come from a simulator. Nothing has flown on hardware.

_{Three real autopilots flying the coverage planner's output in AUTO, in 6 m/s wind, over locally cached imagery. Every number came from an autopilot.
▶ mission-flight.mp4 · ground-station/README.md}

Next actions

Registration Aug 2026 · reviews Oct and Dec · finals Jan 2027. About 21 weeks, but monsoon and exams cut the flight-test window to ~8 weeks. See the development plan.

2. Design point

Outputs of a closed model in tools/sizing-model/, reconciled with the BOM to ~2 %.

⚠ The reserve is spent. The 300 g parachute (SYS-41) takes endurance to ≈2.0× against a self-imposed ≥2.0× policy. Any further mass growth breaks it. See bom_reconcile.py.

Under review: 40 m beats 60 m — 140 px on a person instead of 93, for 2.5 min of a 15 min allowance. Pending recall measurement in P7.

3. Proof

Three ArduCopter SITL instances, the real rescueswarm-drone{1,2,3}.parm, the real planned missions. Every number below came off MAVLink. Telemetry is committed in simulations/recordings/; the figures redraw from it.

3.1 The battery failsafe brings them home

BATT_FS_LOW_ACT = 2 had been reviewed and unit-tested for weeks, and no simulation ever loaded the file — every SITL script used stock defaults, where the value is 0 and a low battery does nothing at all. Now proven end to end: RTL fires at 10 809 mAh of a 10 800 mAh trip, on all three aircraft, within two seconds of each other.

3.2 They no longer collide

Three defects, none visible in review, all found by running the real artifact:

• Launches were simultaneous. Three aircraft leaving slots 1.22 m apart measured 1.3 m from each other at 2–3 m altitude. Now NAV_DELAY 0/15/30 s in the mission file.
• Two of three swept the wrong way round, finishing 516 m and 540 m from the pad on the lowest state of charge of the flight, because sweep direction keyed on drone index. Now all three finish inside 130 m, same path length.
• They landed on top of each other. See §4.

3.3 The search pattern

Each strip is swept twice, the second pass on the reverse heading, and every sweep finishes near the pad.

3.4 The pad

"Measured" means measured in simulation. That rules out whole classes of error and none of the physical ones. HANDOFF.md §2 keeps measured, modelled and assumed apart.

4. Decisions

Numbers in configuration trade.

5. Scoring, and what reading it properly changed

1000 points: 600 flight · 200 design review · 200 business strategy.

Three findings that moved the design more than any engineering analysis did:

• Geolocation gates 450 of 600 points, not 250. Kits are scored from the survivor, so a drop is never better than the tag it aimed at. RTK alone is worth 82 delivery points.
• We were designing to the worst zone. The old ≤3 m target scored 8 of 20 per drop; the old "90 % within 5 m" requirement sat outside every zone and would have scored nothing.
• Speed is worth 50 points and is already won. The bonus needs 15 min; we fly 7.7. Spend the surplus on recall — fly lower, take more looks.

Full breakdown in rulebook compliance.

6. Open

Four questions are drafted and unsent — how "correctly geotagged" is verified, whether a canopy descent scores −10 or −50, whether motor-out tolerance is required, and whether the pad can be surveyed beforehand. The first decides whether ~1–2 m of absolute base error costs anything against 250 points. Answers already received are recorded in the same file.

7. Repository

HANDOFF.md                  read this first if you are picking the project up
TRAPS.md                    read this before adding a config, a generated file or a test
docs/                       requirements, sizing, business, system overview
tools/sizing-model/         the model everything traces back to
hardware/bom/               Indian BOM + indigenisation scorecard
autonomy/coverage_planner/  boundary in, one AUTO mission per drone out
perception/geotagging/      pixel to lat/lon, and a Monte Carlo of its error
firmware/ardupilot-params/  the five failsafes, as parameters not code
simulations/                SITL harness, committed telemetry, figures
ground-station/             mission_backend + the GCS engineering record

The running ground station — Flask server, React client, SITL launch scripts — is in NIDAR-GSC. What lives here is ground-station/mission_backend/ (fleet model, KML parser, MAVLink ingest, and the SYS-20 module split that makes a retask route structurally unreachable), plus the requirement evidence: PLAN.md, the flight screenshots and mission-flight.mp4.

⚠ The two copies of mission_backend have drifted. mavlink_ingest.py is 186 lines here against 304 in NIDAR-GSC, and only the GSC copy calls SET_MESSAGE_INTERVAL — without which a passive listener receives nothing but heartbeats. The GSC copy is the one that flies; this one is what CI tests. See HANDOFF.md §4.5.

communication/ is still planned.

pip install -r tools/sizing-model/requirements.txt
python3 tools/sizing-model/rescueswarm_sizing_model.py   # the design point
python3 tools/sizing-model/bom_reconcile.py              # read before ordering

export NIDAR_SYS=$PWD
../NIDAR-GSC/scripts/test-battery-failsafe.sh            # low battery brings it home
python3 simulations/sitl/fly_and_record.py               # three aircraft fly the plan
python3 simulations/sitl/proof_figures.py                # redraw §3 from telemetry

NIDAR_SPEEDUP=3 slows the sim enough to see the launch stagger. Do not use NIDAR-GSC/scripts/run-sim.sh — it launches a fixed wing with no project parameters.

Change the model, re-run it, commit the new output in the same commit. Every number in this README traces to docs/sizing/, and CI fails if one drifts.