Is my flight log private when I upload it to LogHat?

Yes. The log is uploaded over a 1-hour SAS URL to a private Azure Blob container. Only your account can read it back. The PDF report has an optional share link you control — by default it's private; you choose whether to enable it per ticket.

What file formats can I upload?

ArduPilot dataflash logs (.bin and .log), MAVLink telemetry logs (.tlog), and PX4 ULog files (.ulg). All four normalise into the same internal schema, so the analysis depth doesn't depend on which format you started with. We don't currently parse Blackbox or BetaFlight logs.

How is LogHat AI different from MAVExplorer or UAV Log Viewer?

MAVExplorer is a power-user tool for plotting and querying log fields manually — fast, scriptable, but it doesn't know which fields matter for your specific crash. UAV Log Viewer is a browser-based viewer with 3D replay. LogHat adds the causal-hypothesis layer on top: it ranks the failure modes against your log and explains the chain. The three tools overlap; we use MAVExplorer ourselves for one-off field exploration.

Can LogHat tell me whether a crash was the pilot's fault?

We can identify what the autopilot did and what the sensors reported. We deliberately don't assign blame — pilot decisions are visible in the log but the context (visibility, training, mission objective, fatigue) isn't. The report describes the technical cause; the human-factors investigation belongs with whoever has that context.

Does the Vector AI have a knowledge cutoff?

The base model has a cutoff and we surface it in every regulatory or version-sensitive answer. The Qdrant knowledge base, which the RAG layer cites for ArduPilot and PX4 forum content, is updated continuously by background crawlers — so version-specific guidance (Copter 4.6, Plane 4.7 dev) reflects what's currently in the community archives. When asked about regulations or anything time-sensitive, the answer is stamped with the cutoff and a reminder to verify against the live source.

How does LogHat handle PX4 ULog versus ArduPilot bin?

Both pipelines extract the same set of normalised events: failsafes, mode changes, sensor health flags, vibration profile, GPS-fix quality, EKF status. The downstream forensic analysis treats them identically once normalised, so a PX4 ULog gets the same root-cause ranking as an ArduPilot bin. Where the underlying message schemas differ (PX4's estimator_status versus ArduPilot's XKF4), we map them into a common health representation rather than presenting raw fields.

How LogHat AI Analyses a Drone Crash Log Automatically

TL;DR: Drop a .bin, .tlog, .log, or .ulg into LogHat and the analysis pipeline returns a crash diagnosis in under a minute. The Vector AI copilot reads every ERR and MSG line, identifies failsafe chains, and surfaces the three or four signals most likely to be the root cause. It won't replace the screwdriver work, but it compresses the first hour of forensic analysis into 60 seconds you can act on.

What happens to your log in those 60 seconds

Upload. The file lands in Azure Blob via a 1-hour SAS write URL. The log never sits on a public endpoint.
Format detection. .bin and .log route through the ArduPilot DATAFLASH parser; .ulg through the PX4 ULog adapter; .tlog through the MAVLink telemetry adapter. All three normalise into the same internal schema so the downstream analysis is identical.
Four parallel parsers. A demuxer reads the file once and fans each MAVLink message to four concurrent cursors: deep flight analysis, sim-core extraction, AI digest, and telemetry graphs. The slowest stage is the bottleneck, not the I/O.
Vector AI digest. Your log is cross-referenced against a Qdrant knowledge base of ArduPilot and PX4 forum threads, official docs, and prior LogHat analyses. The pipeline surfaces the specific ERR Subsys and ECode values your log produced, the failsafe sequence, and the parameters most likely to be misconfigured.
3D replay + PDF. The complete report lands as a downloadable PDF plus an interactive 3D replay you can scrub through frame by frame.

What the copilot tells you that grep can't

A raw messages ERR dump in MAVExplorer lists every error code in the log. What it doesn't do is order them by causal probability. The Vector AI weights events — an isolated ERR Subsys=11 ECode=2 GPS glitch is information, but a Subsys=11 ECode=2 followed within 500 ms by Subsys=17 ECode=1 (EKF failsafe) followed by Subsys=12 ECode=1 (crash check) is a causal chain that points at the GPS as the originating cause. The chain matters more than any single error.

An example workflow

Suppose your hex went down ten minutes into a survey mission. You upload the .bin.

Within 45 seconds you see the "Root Cause Hypotheses" section in the PDF. Top hypothesis: GPS-induced position drift triggering EKF failsafe. Confidence: high.
The supporting evidence: GPS.HDop climbed from 0.9 to 2.4 over 8 seconds, GPS.NSats dropped from 18 to 7, XKF4.SP crossed 1.0 four times in the next 3 seconds, and an ERR Subsys=17 ECode=1 wrote at TimeUS=614223110.
You click "Show in 3D replay" and the timeline scrubs to that timestamp. You watch the drone's track drift north over an industrial roof — structures that suggest multipath.
Recommended actions appear: elevate the GPS antenna, enable GPS_GNSS_MODE = 65 for multi-constellation, and gate arming on HDop < 1.2. Each links to our deeper post on the topic.

Total time from upload to actionable plan: under a minute.

What it won't do for you

It won't replay your decisions. The pilot's stick history is in RCIN. The copilot reports the inputs but won't tell you that switching to STABILIZE during a position glitch was the wrong call — that judgement is yours.
It won't see what the log can't see. Direct sun on the barometer, a loose servo cable, a fuel tap that wasn't fully open, prop damage from a hard landing the day before — these only show up as downstream symptoms.
It won't replace tuning. If the log shows oscillations at 40 Hz, we can flag them and link to our VIBE and harmonic notch posts. The actual INS_HNTCH_* configuration is something you decide based on your airframe.
It won't certify your operation. The PDF report is a forensic narrative; it isn't a substitute for an airworthiness investigation, an insurance assessor's report, or a regulator-compliant audit.

The artefacts you can integrate

Every output we generate is also available as a JSON blob. vector_analysis.json contains the normalised event timeline; flight_metadata.json carries the flight envelope summary; report_digest.json is the text the Vector chat references. If you want to integrate LogHat into your own dashboard or maintenance-ticket system, those three files are the contract — the PDF and 3D replay are surfaces built on top.

How to make the AI sharper for your fleet

Every hypothesis carries a thumbs-up/down. Down-votes are the highest-value feedback we get — each correction sharpens the next analysis for the next engineer with a similar crash. Use them generously.