← Drone Dynamics

DYNAMICS LAB · ENGINE01

Comparing flight trajectory models

All models are compared on the same nine held-out civil flights. The charts show how errors differ and how often observations fall within the estimated uncertainty regions.

Go to flight playback ↓

Compare the nine recorded flights

The chart compares nine flights recorded with the same drone. Select a flight to highlight it and see its GPS coordinates and model results.

Test flight 9 · Lap030 · 8:55

Aircraft used
Volantex Ranger 2400 · SpeedyBee F405 / INAV
Model with the lowest mean error
Kalman filter · 10.43 m
GPS position at the selected time
Latitude 39.299810° · Longitude -0.611382°
All models tested

The six models and their ensemble were compared with observed GPS positions at 1, 3 and 5 seconds. Results apply to the selected flight.

ModelMean error (ADE)Root mean square error (RMSE)
Kalman filter (KF)10.43 m17.56 m
Correlated random walk (CRW)10.68 m17.21 m
Model ensemble (ENSEMBLE)10.81 m16.81 m
Constant velocity (CV)14.31 m23.52 m
Constant acceleration (CA)20.30 m31.66 m
Random walk (RW)31.36 m38.98 m
Ornstein–Uhlenbeck (OU)184.66 m212.40 m

The tracks share a geographic reference. Playback is accelerated. WGS84 coordinates come from the original GPS samples shown in the chart.

DME-001

Experiment conclusions

On nine test flights, Kalman had the lowest mean displacement error at 9.12 m, slightly ahead of CRW at 9.27 m. The ensemble had the lowest RMSE, 15.12 m, while its mean error was 9.60 m. Model choice therefore depends on the error measure that matters.

Study limits. The flights share a platform and circuit. GNSS is a noisy reference and the ranking may differ on other flights.

Model-by-model results
RW
Random-walk reference: 34.80 m mean error, above constant velocity.
CV
Constant velocity: 12.74 m mean error, a useful simple baseline.
CA
Constant acceleration: 18.13 m mean error, above CV in this test.
KF
Kalman: lowest mean error, 9.12 m.
CRW
Correlated random walk: 9.27 m mean error, close to Kalman.
OU
Ornstein–Uhlenbeck: highest mean error, 180.27 m. High coverage did not offset poor accuracy.
Ensemble
Model combination: lowest RMSE, 15.12 m.
HMM
Groups motion patterns into descriptive states. The states have no verified intent labels.
EKF
Extended Kalman filter: available in the core, without a measured result in DME-001.
Particle Filter
Available in the core, without a measured result in DME-001.
Monte Carlo
In DME-003, 100 synthetic random-walk replicates produced a distribution of final outcomes.

DME-001 · Real experimental data

Model errors on the same test set.
Model errors on the same test set.
Actual coverage of uncertainty regions in the test set.
Actual coverage of uncertainty regions in the test set.
Recorded flight path in local coordinates.
Recorded flight path in local coordinates.

The uncertainty intervals were estimated by resampling whole flights. The flights share an aircraft and circuit, and the recording dates are unknown, which limits independence. The comparison tests whether more complex models offer an advantage.

Data and technical files

The technical details below contain the study tables and limitations. The original data and instructions are available to download.

Technical details and reproducibility

Study context

Problem addressed
Which motion estimates perform best on the same held-out historical flights?
Contribution relative to existing methods
The benchmark puts established methods and a validation-fitted ensemble under one protocol.
Potential commercial use
Select candidate estimators for offline evaluation on a civil fleet’s own telemetry.

Market context and value assumptions on the engine page

Computed 2026-11-06T10:25:09.902316+00:00 · reproducible scientific prototype

Measured and computed results

modelADE_mFDE_mRMSE_mMAE_coordinate_mcoverage95
RW34.79599357.40528241.42632820.4674110.96096
CV12.74391920.82719422.1855878.2067630.90653
CA18.12908935.42205529.65208211.4919210.883761
KF9.12169517.24209815.832155.7986830.904366
CRW9.27140817.24023515.4396845.926360.909744
OU180.270339237.539602208.814684116.086710.98021
ENSEMBLE9.5996117.40035515.115736.1198570.926349

Paired comparison and uncertainty

modelreferencemean_ADE_difference_mtwo_sided_exact_signflip_punitsholm_adjusted_p
RWCV22.0520750.00390690.023438
CACV5.3851710.00390690.023438
KFCV-3.6222230.00390690.023438
CRWCV-3.4725110.00390690.023438
OUCV167.526420.00390690.023438
ENSEMBLECV-3.1443080.00390690.023438

Robustness

conditionmodelADE_mmatched_baseline_ADE_morigins
context_noise2mRW34.98976134.795993439
context_noise2mCV13.18690412.743919439
context_noise2mCA23.08131318.129089439
context_noise2mKF23.9340849.121695439
context_noise2mCRW23.2605769.271408439
context_noise2mOU180.268998180.270339439
context_noise2mENSEMBLE22.6866699.59961439
context_missing30percentRW35.76946234.874378435
context_missing30percentCV12.86528712.78599435
context_missing30percentCA18.33421818.112987435
context_missing30percentKF9.2439759.115594435
context_missing30percentCRW9.5046229.270341435
context_missing30percentOU182.589058179.941936435
context_missing30percentENSEMBLE9.8308669.594352435

Scientific result SHA-256: 1f0238e055f9f9a1d601982e68e3e6406f6148a4e768ca17fd0cdfed30427280. Numerical correctness is evaluated with independent controls and source replay; scientific validity remains bounded by source quality and model assumptions.

Limits

Recorded GNSS is a noisy reference, not known true state. Physical sensor calibration is unavailable.
Flights share platform and circuit; recording-day groups are unavailable. File-disjoint splits prevent row leakage but do not prove external generalization or independent days.
Gaussian nominal regions and validation scaling are empirically evaluated, not guaranteed confidence coverage.
ADE averages sampled1/3/5second horizons, actual timestamp offsets recorded; no interpolation.
All parameter and ensemble fitting use development/validation only; test outcomes are not tuning inputs.