Munim Technologies

Research/Software

Many drones run with a GPS delay setting that is far off

Across 164 public PX4 flight logs, GPS velocity lagged the IMU by a median 99.8 ms, close to the 110 ms default. But 29.6% of logs were more than 50 ms off their own setting, and an independent sample replicated both figures.

Oct 30, 2026 · 7 min read

A drone's flight controller fuses GPS with its inertial sensors, and it needs to know how stale each GPS reading is. PX4, the open-source autopilot, assumes 110 ms unless the owner changes the setting EKF2_GPS_DELAY. We measured the real lag in public flight logs. On 164 logs from 115 vehicles, GPS velocity lagged the IMU by a median 99.8 ms, so the default is about right on average. It is wrong for many vehicles: 29.6% of logs were more than 50 ms off their own setting, and 98% of logs ran the default. An independent sample of 98 logs replicated both figures.

The mismatch is not cosmetic. On 256 further logs, EKF2's own GPS consistency check rose more sharply during acceleration on mismatched vehicles. A companion study of 208 public flights found that three quarters raise at least one health flag, and that PX4's battery-resistance estimate collapsed to zero on 10 of 75 measurable flights.

99.8 ms

median GPS lag on 164 logs (default setting 110 ms)

29.6%

of logs more than 50 ms off their own delay setting (47 of 159)

5.3 ms

median difference between two logs of the same vehicle

+103.7 ms

extra lag from u-blox dynamic model 6 vs 7, one vehicle, same day

The question

GPS receivers filter their solution and send it over a serial or CAN bus, so each reading arrives some time after the moment it describes. EKF2, PX4's estimator, subtracts a fixed delay to line GPS up with the IMU. If the setting is wrong, the estimator compares GPS with the wrong moment of the flight. We asked how large the real lag is across the fleet of public logs, how much it varies, and whether a mismatch shows up in the estimator's own behaviour.

What we did

For each pair of consecutive GPS samples in flight, our tool compares the change in GPS north and east velocity with the IMU's acceleration, rotated into the world frame and integrated over the same interval shifted back by a candidate lag. The lag that fits best, allowing a constant accelerometer bias, is the measurement. It is taken against the same timestamp from which EKF2 subtracts its delay. On synthetic flights the method recovered known delays within 3 ms.

The fleet study was pre-registered: the plan was committed before any log was downloaded. A sealed job drew 400 eligible logs from Flight Review's public list of 483,966, from October 2025 on, at least 240 s long and not simulated. 164 passed the gates for enough manoeuvring and a measurable fit. Five of six predictions held. A separate pre-registered study, in px4-log-health, drew 300 logs from other vehicles, one per vehicle, and set ranges for the lag figures before it ran.

What we found

The spread is wide: half the logs fell between 71 and 133 ms, and a tenth below 48 or above 179. But two logs of the same vehicle agreed to a median 5.3 ms across 49 vehicles, so the lag behaves like a property of the vehicle, not of the flight. That makes a per-vehicle value measurable. One prediction missed: we expected DroneCAN receivers to lag more than serial ones, and they lagged less, 90.4 against 108.7 ms.

GPS lag in the first sample and an independent replication

fleet-2026Replication (fleet-health-2026)
Logs passing / vehicles164 / 11598 / 98
Median lag99.8 ms113.8 ms
Interquartile range71.3-133.0 ms88.5-159.4 ms
More than 50 ms off own setting29.6% (47 of 159)31.2% (of 93)
Logs running the 110 ms default98.1%95.7%
Both samples pre-registered; the replication's ranges (median 80-120 ms, 18-42% off by more than 50 ms) were set before it ran, and both figures fell inside them. Sources: px4-gps-lag studies/fleet-2026/GRADE.md; px4-log-health studies/fleet-health-2026/GRADE.md (R1, R2).

Two independent checks agree with the measurement. On the 98 replication logs, EKF2's own GPS velocity innovations leaned the way the measured mismatch predicts in 25 of the 27 logs more than 50 ms off, and the two methods ranked logs alike (Spearman +0.86). On 256 further logs, EKF2's GPS velocity test ratio while accelerating rose to a median 2.7 times its steady-flight value on mismatched logs, against 1.7 times on the rest. We had also predicted no extra GPS rejections; that check failed, because the interval on the difference (20% against 16% of logs) reached +14 points, above our +10 limit.

Part of the mismatch has a cause. u-blox receivers tune their navigation filter to a dynamic platform model, and PX4 set model 6 (airborne, under 1 g) for multicopters until v1.17 moved them to model 7. In 244 public u-blox logs, model-6 logs lagged a median 45.6 ms more than model-7 logs with the same receiver class. On one vehicle flown the same day under both settings, the lag was 214.2 ms at model 6 and 110.5 ms at model 7.

Median GPS lag by u-blox receiver and dynamic model

ReceiverModel 6 (<1g)Model 7 (<2g, current default)Model 8 (<4g)
F9P180.7 ms (15 logs)137.3 ms (12)120.9 ms (6)
M9108.7 ms (25)62.0 ms (19)69.1 ms (34)
M8186.1 ms (17)114.1 ms (9)124.4 ms (7)
Model not detected144.4 ms (47)116.2 ms (19)115.8 ms (31)
Pre-registered; all checks held. Because older defaults gave model 6 to multicopters, vehicle type and release could confound this; an exploratory re-score among multicopters only kept the gap (+45 ms). Source: px4-gps-lag studies/dynmodel-2026/GRADE.md.

Flight health more broadly

The replication sample doubled as a health census of 208 flights from 208 vehicles, scored against Flight Review's own limits. 75% raised at least one flag. Six of ten health predictions held; our priors for vibration and receiver jamming state were too low. On logs from v1.13 or newer, where Flight Review's vibration bands apply, 51.5% sat in the orange or red band. A follow-up found that PX4's battery internal-resistance estimate read exactly zero for the whole second half of the flight on 10 of 75 measurable flights. Nine of those ten relied on the estimate, so they flew with state of charge not corrected for load.

Share of public PX4 flights showing each issue

Denominators differ: each share is over the logs that report that figure. The vibration split by release was counted after grading and is exploratory. Uploaded logs over-represent flights with problems. Source: px4-log-health: studies/fleet-health-2026/GRADE.md; studies/battery-r-2026/GRADE.md (R1)

What it means

The default delay is a reasonable average and a poor fit for about a third of vehicles. Because repeat flights agree to a few milliseconds, an owner can measure their own vehicle's lag from one manoeuvring flight and use it as a starting value, checked with EKF2 replay. Owners of older multicopters on u-blox model 6 should expect the largest gap. Our checkup tool reports the measured lag, the fleet comparison and the health flags for any log, and runs in a browser with nothing uploaded.

More research