Field Test

Everything here has been checked on a bench against a real aircraft except the flight. The onboard mission sequencer, the arrival dwell, the take-off altitude climb and the re-evaluated position check have each only ever run in a compiler. Treat the first flight as a test of the sequencer, not of the waypoints.

Three tools, in the order you’ll want them:

Tool	What it’s for
field_check.py	Walks the link, settings and payload checks and prints pass/fail
QGroundControl	Proves the aircraft is a standard MAVLink vehicle, not just one this ground station understands
The dashboard’s MAVLink tab	Runs both wires side by side and shows where they disagree — the only place this project’s bugs have ever been visible

Before leaving

export LB_MAVLINK_PORT=14551        # QGC keeps 14550; a shared port means each gets half
export LB_MAVLINK_PEER_PORT=14550   # where the aircraft listens
export LB_MAVLINK_SIGNING_KEY=...   # optional; with it you are the Safety Computer

Check the phone and the ground station are on the same build of the dialect. A mismatch is refused on the checksum, so fields go missing rather than wrong — which reads like a weak link and is not one.

1. Link only — safe with props on

cd GroundStation/Python
PYTHONPATH=. python test_scripts/field_check.py <PHONE_IP>

Sends nothing. Confirms telemetry arrives, that the core fields decode, that the config streams, and that the position isn’t (0, 0).

Read the warnings, don’t just count the failures. homeLocation, remainingFlightTime and distanceToHome absent is a state of the aircraft — no home point, no battery estimate — not a gap in the interface. The script never fails those.

2. QGroundControl sees a normal vehicle

Point QGC at UDP 14550 and confirm, without touching anything else:

• the vehicle appears and stays connected, no “waiting for vehicle”
• position, altitude, heading and battery on the HUD, all moving
• the flight mode reads a name, not a number
• QGC and field_check.py run at the same time, both live. If one freezes they’re sharing a listen port — a UDP datagram reaches exactly one socket.

This is the check that matters most for interoperability: it proves the aircraft speaks MAVLink, rather than speaking whatever this ground station happens to accept.

3. Settings, on the ground

PYTHONPATH=. python test_scripts/field_check.py <PHONE_IP> --phase ground

Writes the RTH altitude and reads it back through telemetry, not from the ack — the ack only proves the message arrived. Also confirms a non-numeric setting comes back refused rather than being smuggled through as a magic number.

Cross-check in QGC’s parameter editor: LB_RTH_ALT, LB_MAX_HEIGHT, LB_MAX_DIST.

4. Payload — the gimbal will move

PYTHONPATH=. python test_scripts/field_check.py <PHONE_IP> --phase payload --move

Gimbal absolute and relative, photo, thermal, temperature, rangefinder, AutoSensing.

The gimbal check is worth watching directly. The sign convention was settled by a 48-sample measured sweep, not by argument; if the read-back lands near +45 after commanding −45, the sign has inverted. A reading of 6553.5 is DJI’s unset marker (65535/10), not an angle.

Point the camera at a person for the AutoSensing step — zero detections with an empty frame is the correct answer and tells you nothing.

Then confirm the photos landed:

PYTHONPATH=. python test_scripts/ftp_exercise.py <PHONE_IP>

Lists the card over MAVLink FTP and cross-checks the sizes against HTTP /send/listMedia.

5. The signing gate

PYTHONPATH=. python test_scripts/signing_exercise.py <PHONE_IP> [KEY_HEX]

Proves an unsigned command is refused once a signed one has latched authority to the Safety Computer, and that only a signed release hands it back. Verified against pymavlink’s signer, but no real Safety Computer has ever sent one.

6. Flight

PYTHONPATH=. python test_scripts/field_check.py <PHONE_IP> --phase flight --fly

prints this list. Pilot’s thumb on the sticks throughout. Every command has a manual override, and item 5 confirms the override works before anything later depends on it.

#	Check	What proves it
1	Take-off	isFlying true, altitude climbs
2	Take-off altitude	Ask for a height in param7 and get that height. The climb runs after DJI’s fixed-height auto-take-off, so watch for an overshoot
3	gotoAltitude / gotoYaw	Each acks IN_PROGRESS, then again on arrival
4	Manual override	Take the sticks mid-command: motion stops, and the next command is refused rather than silently ignored. Land, RTH and abort must stay available — those are the recovery actions
5	Release override	Control returns
6	Single waypoint	Flies there, and intermediaryWaypointReached latches once, with the right seq
7	Arrival dwell	It latches where the aircraft actually is, not on one noisy fix at the edge of the box, and not repeatedly. Three ticks by default, per-waypoint from the plan
8	Hold heading	Swarm-Steward uses hold-heading, not nose-forward: the aircraft should keep its heading rather than yaw into the leg
9	Multi-leg plan	Two or three legs, low. Each completes in order, MISSION_CURRENT advances, nothing is skipped. The single largest untested path
10	Settle vs pass through	It settles where the plan says to settle, and flies through where param3 says to pass
11	Supersede	A new plan mid-leg makes the old one report SUPERSEDED, not FAILED
12	Abort	Stops where it is, acks CANCELLED
13	RTH	Returns and lands at the home point, at the altitude written in step 3
14	Swarm-Steward on LB_TRANSPORT=mavlink	An action completes end to end. Fall back to both if something is unmapped — that keeps MAVLink telemetry and HTTP for the rest

Re-test first: the two QGC guided commands

Both failed on the first field attempt and both are fixed, so these are the ones to fly before anything else depends on them.

QGC → Go to location. It flew away from the point. MAV_CMD_DO_REPOSITION’s ground-speed parameter is documented as “less than 0 (-1) for default” and QGC sends -1; that reached the waypoint loop as the speed ceiling, so the commanded speed was −1 m/s and the aircraft retreated at walking pace. It could never recover on its own — the along-track term stays positive while it retreats, so nothing reversed the sign. Confirm it now flies toward the point, at the profile cruise speed.

QGC → Change Altitude. Nothing happened. QGC expresses it as the same DO_REPOSITION with latitude and longitude NaN, meaning “hold position, change only altitude”. Those were flown as coordinates: over COMMAND_INT NaN scales to 0, which is a real position in the Gulf of Guinea. It now routes to the altitude controller. Confirm the aircraft climbs or descends without rotating — a zero-length leg has no bearing, and the nose-forward controller would have read atan2(0, 0) and turned to north first.

Also worth watching: a goto sends yaw = NaN, “don’t change yaw”. The arrival heading was hardcoded to zero, so the aircraft used to finish by rotating to north. It should now hold the heading it already had.

Keep this open while flying

The dashboard’s MAVLink tab, with cross-check against HTTP on.

Not one MAVLink defect in this work was found by reading code. Every one was a frame that decoded cleanly and meant the wrong thing — a camera heartbeat overwriting the flight mode, two gimbal fields transposed because MAVLink packs largest-first, a status message read twelve bytes late. A compiler can’t see any of those, and neither can a reviewer.

A disagreement that appears only in the air is the interesting one. Everything above has been checked at rest, so anything new belongs to motion: a value that’s fine stationary and wrong at speed, or one that saturates.

MAVLink commands are now recorded in the flight log alongside the HTTP ones, with their raw parameters, in /sdcard/Documents/lyrebird/FlightLogs/<date>/. They were not, which is why the first two field failures had to be diagnosed from the source rather than from evidence — a mission flown over MAVLink left no record of having been commanded at all. Both defects are a sentinel value flown as a real one, and both are plain in a line reading p1=-1.0 or a NaN latitude. Pull the log after flying:

adb pull /sdcard/Documents/lyrebird/FlightLogs/$(date +%F)/