Hi all, I’ve been building FlightMD, a free and open-source diagnostic tool for PX4/ArduPilot/MAVLink flight logs. Upload a .ulg/.bin/.tlog and it runs 8 deterministic rule-based checks, oscillation, vibration, EKF, battery, GPS, parameters, motors, plus ascent/recovery detection for rockets and HABs, and returns a health score plus specific parameter fixes. No cloud AI in the diagnosis path, no account needed, fully self-hostable.
Rather than trust my own synthetic test fixtures, I validated it against 50 real logs pulled from PX4’s own public database, quads, hexes, VTOLs, fixed-wing, even a coaxial helicopter and a rover. Full writeup and all 50 logs are on GitHub, including where my own cross-checks disagreed with the noisy metadata.
Posting here because I’d genuinely value feedback from people who know these logs better than I do. If you’ve got a PX4 or ArduPilot log lying around, I’d appreciate you running it through and telling me where the analysis is wrong or unhelpful. There’s also an opt-in checkbox to anonymously contribute a log + its report to a growing validation dataset, if you don’t mind.
Just check a couple, and smarttune cli is honestly the closest match: also deterministic/rule-based, also multi-format (ArduPilot+Betaflight+PX4), also gives validated parameter recommendations, only thing is CLI only, rather than a hosted web app, along with no PDF/GPX/KML export or 3D path visualization. then we have, ardupilot-log-diagnosis (BeastAyyG) is a genuinely solid hybrid rule especially with the ML tool, but what I understand is here, the results come from a trained classifier rather than fixed thresholds, so it’s not auditable from the code alone in the same way. And the ArduPilot’s own AI Log Analyzer goes the opposite direction from mine, an AI agent deciding what to parse, whereas mine has zero AI in the diagnosis path.
As for my work, FlightMD differs from all of these: a hosted web app with PX4+ArduPilot+MAVLink auto-detection in one tool, fleet-ops features (cross-flight trends, maintenance tracking, webhook alerts), PDF/GPX/KML export + 3D/map visualization, and a published validation run against 50 real public PX4 logs rather than just my own tests, and they were randmly picked, for variable samples. Should’ve researched more before working, and posting, genuine overlap, especially with smarttune-cli.
As for the vibe thing, or developing it, was leaning towards developing it, but given that list, will have to re-evaluate things. And since you asked directly: yes, I used AI coding assistance building this, alongside my own testing and validation work. Not going to pretend otherwise.