Why the build looks the way it does

This TerraHawk is our own aircraft, and every part of it is a choice that closes off some others. This page records the ones made so far, and why, so a later change can see what it’s undoing.

6S, decided 2026-09-25

The pusher decides this, not the lift motors. A 3115 pusher comes in two windings: 900KV for 4S and 640KV for 6S. The ArduPilot values, like Q_M_BAT_VOLT_MAX of 25.2, assume 6S. Mixing voltages means either an under-revving pusher or retuning those numbers. So the whole powertrain is bought for one voltage.

Going to 4S would save about $143: the 4S lift motors and battery cost that much less than the 6S ones, and the pusher costs the same. It’s still in the parts list in case the decision gets revisited. Both ESCs already cover 4S, so switching would mean swapping the motors and the battery and nothing else.

A wing board over a budget stack

The Corvon (or Aero Selfie) H743 is $33 cheaper than the Kakute H743-Wing, and ArduPilot genuinely supports it. What it gives up matters on a quadplane. It has no dedicated servo rail, so both elevon servos share 5 V with the receiver and GPS. And it has no onboard current sensor: amps come only from the 4-in-1 ESC pin, so the pusher’s draw never shows up in the logs or in current-based failsafes. The electronics notes have the full comparison. For a first build, seeing the whole power budget is worth $33.

The wing board itself changed on 2026-09-26, to the TBS Lucid H7 Wing. It keeps everything that argued against the budget stack (an 8 A servo rail and a 165 A current sensor) and adds a second IMU. The Kakute H743-Wing stays as the fallback while the Lucid is on backorder.

Then it changed again on 2026-09-27. The plan is now a flight controller of our own: an H743 with two CAN buses, split by what a fault would cost. Its ArduPilot board definition is already written and builds (Flight controller firmware). Until it exists the aircraft flies a Holybro Pixhawk 6C Mini with a PM02 V3 power module, in place of the backordered Lucid. The 6C Mini fits the same tray and has heated, isolated IMUs. Much of the aircraft now talks DroneCAN (the lift ESCs, airspeed and the charge node), so the output rail matters less than it did when the Lucid was chosen. The GPS, a Waveshare Q43 NEO-F9P, plugs straight into the flight controller’s GPS1 port. The electronics notes have the reasoning.

On 2026-10-04 the board for the first flights changed twice: to the HKUST NxtPX4v2, then, the same day, to the Lumenier LUX F765 NDAA. The LUX has one CAN port, off for the first flights, and outputs for the cargo door, but only one IMU, so the EKF has no second IMU to change to if it fails. With CAN off, the first flights fly the 4-in-1 lift ESC, an I2C MS4525DO airspeed sensor and no charge node. Our own H743 board is still the plan after them.

“19 × 19” is a circle, not a square

Both printed motor mounts have four M3 holes on a 13.44 mm square, which is 19.00 mm corner to corner. Motor vendors’ “19 × 19” or “19 mm” is the bolt-circle diameter. So every motor we considered fits, and a cheaper pusher with conflicting vendor data (one reseller lists the EMAX ECO II 3115 as 16 × 16) was left out until EMAX confirms it.

Print it here, not at a service

Only three print services quote lightweight PLA, and all three are in Europe. The cheapest was $865.95 shipped, black only. Printing on the Core One+ costs about $80 in filament. The bigger problem with a service is that nothing guarantees the lightweight-PLA profile’s single perimeter, 4% cubic infill and 0.60 flow. Standard PLA at a service’s default walls makes the skins two to three times heavier on a roughly 2 kg aircraft. The quotes are kept in print service quotes as a fallback.

Group the beds

The cautious default is to print one part at a time. When the first airframe alone was scheduled (2026-09-25), grouping parts with identical settings onto one bed took it from 4.6 days to 3.6 days for someone home all day, and from 7.6 to 4.3 on workdays. The build has grown since with our own parts; the print timeline has today’s days. Much of the saving is heat-up, homing and first-layer setup happening once per bed instead of once per part. The first design’s four elevon halves took 4.1 hours together against 6.1 apart. The cost is risk: one failed part can take its bed-mates with it. The one-part-per-bed files are kept for reprints, and the Winglet test bed stays alone as a go/no-go gate. See the print timeline.

Survey3W over Survey3N

Both are built to be triggered by an autopilot, and both return an exposure pulse, so every photo is tagged with the flight controller’s position. The difference is overlap. At 1.5 s per JPG and 16 m/s, the camera takes a frame every 24 m. At 120 m the wide Survey3W gets about 85% forward overlap at 5.5 cm/px. The narrow Survey3N resolves 2.3 cm/px but only gets about 65% overlap, so it would need slower or higher flight. The wide lens suits the speed the aircraft wants to fly.

The payload’s roughly 88 g sits about 100 mm behind the CG. That moves the battery 13 mm forward, and it’s why the camera powers from a 5 V pad instead of its own 26 g battery.

Fewer, larger parts on top, decided 2026-10-03

Three glued joints of the first design went on one day. The spine rear fuselage (2026-09-30) first carried its radome on a removable aft hatch. Ryan removed that hatch: the top, the spine and the radome are now one 7 mm shell on Fuselage Mid B, and the radome follows the Q43’s real profile, which brings the aircraft to about 233 mm tall on its gear (airframe geometry).

The first design split each elevon into an inner and an outer part, glued flush at X 480. Ryan chose a single-part elevon, 315.6 mm long, so the control surface has no glued seam.

The first design closed the front of the fuselage with two hatch halves and a quarter-turn lock. Ryan did not like the lock, and chose a single-part hatch held by a nose tongue at the front and four pairs of 8 × 3 mm N52 magnets. From catalogue pull figures, four pairs at a 0.3 mm gap should hold about 60 N. The worst suction on the hatch at 21 m/s is about 7 N (270 Pa of dynamic pressure, a pressure coefficient of −1, over 0.026 m²). Both numbers are estimates until a bench pull test. On Fuselage C the Hatch is held differently (Ryan, 2026-10-11): two pairs of 20 × 10 × 5 mm N52 block magnets at its front, fitted so that their fields cancel at the compass, and a snap catch at its aft end. The Canopy comes off on snap features, with no magnets (hardware).

The single-part elevon is still cut from the kit’s two elevon halves, so this site shows it in renders only, never as a file.

Our own wing section, with a fence

The wing keeps the first design’s planform, spar lines, hinge line and root. What changed is the shape of the wing in section. th-aeronautics drew new sections for stations along the span, with their own camber and twist, and checked them in XFOIL and then in CFD. At the slow end of the envelope, where a quadplane spends its transitions, the new sections lift between 6 and 14 % more before they stall, and the stall starts inboard, so the elevons at the outboard end keep working while the root lets go (th-aeronautics d8b7892e).

The cruise number is what made it worth printing. In the fine CFD pair, trimmed at the cruise lift coefficient of 0.40, the new wing has about 6.5 % less drag than the first design’s wing (a drag coefficient of 0.0476 against 0.0509), and it trims with about 2.3 to 3.4 degrees less up elevon (th-aeronautics 8e4da533 and the correction 6a86f927). Less drag in cruise is more survey area on the same pack.

The coarse runs on the stall fixes found one more thing. Without help, the outboard panel still separated near 14 degrees and the aircraft pitched up at the stall. A thin PETG fence on each wing at X 445, where the Wing Inner meets the Wing Outer, keeps the outboard panel attached through 14 degrees, costs nothing in cruise, and turns the stall nose-down (th-aeronautics e58f87e6). The confirming runs then showed the elevons keeping about 50 % more control power at the stall with the fence than without it (e21bcc84). Ryan put the fence on the build on 2026-10-17, and th-blender drew it into the wing as one part that the two panels trap at their joint (W73).

All of these are CFD estimates on a model, not flight data. The stall test card is where they meet the real aircraft.

A roll-off hangar, not a clamshell

The first hangar design had two roof halves hinged on the long walls. Its weak point is wind during the move. Each half is about 1.9 m² and passes through vertical on the way open, and a 15 m/s gust on that face is roughly 140 N·m about the hinge. The timing makes it worse: the roof has to open while the aircraft is inbound, committed and low on battery. A roof that rolls off on rails never presents a new face to the wind. The hangar page has the full comparison, including the lean-to that prop wash ruled out.