The barometer was honest all along

Written from flight logs 45 to 49.

Back at the same site with the override latch from the 13th, the altitude top-up reduced to a single attempt, and the takeoff height down to 1.5 m. The question for the day was simple: does the transmitter actually win now?

It does. Every time I flipped Land during a Hawk routine, the routine stopped on the spot and Hawk said “You have control, sir.” Five flights, two of them ended by the switch, both clean. The thing I had claimed on the 13th and could not back up is now in the log.

The slow sink

Then the last flight. “Take off.” It rose to 1.5 m, held, and started coming down. Not Land mode, not a command from me, not a stick input. It settled onto the grass with Hawk still believing it was two metres up.

I had moved the height reference to the barometer after the 13th, and I had assumed that was the end of GPS altitude in the loop. It was not. The GPS altitude was still fused, and with the noise values PX4 ships, the estimator weights a GPS altitude at half a metre of claimed error far above a barometer at three and a half. So the reference was the barometer and the number was still the GPS. Log 49 shows the estimated height tracking the GPS by one to three metres while the barometer, read from the same file, stayed flat and true. On log 47 the same thing climbed the aircraft to three metres real when I had asked for 1.5.

The fix is one bit in one parameter: stop fusing GPS altitude. Position from the satellites, height from the air. Set it in QGroundControl, rebooted, and the next flight is the proof.

Two things that were not bugs

Twice the aircraft took off in Position mode and started drifting on the sticks. Both times the mode switch was still in Land from the previous flight; moving it to Position mid-flight selects Position. The switch goes to Position before “take off”, every time, and that is now written down.

And once I released the kill switch a second too early, while the aircraft was still armed, and the motors restarted on the ground. Kill stays engaged until the controller says disarmed.

What the GPS is doing here

The sideways wander in the hover is not Hawk and not the controller. The raw GPS position wanders a metre or two over twenty seconds while claiming thirty-centimetre accuracy. That is the buildings. Thirty satellites do not help when half of them arrive by reflection. The fixes are a site with sky, or a downward optical-flow sensor with a rangefinder, which is how a consumer drone holds still for a quarter of the price. The second one is a part, not a redesign, and it is on the list.