Simulated IMU / Estimator comparison
Roll & gyroscope bias.
Four methods, shared measurements, known simulation truth.
Loading the recorded simulations…
Roll angle
Gyroscope bias
Inspection window
Error and model uncertainty
Measurement corrections
Reading these diagnostics
State errors use every original endpoint, held until the next recorded prediction and any correction. Steps extend to the window edges; no estimator restarts at the window boundary. State uncertainty is recorded after prediction and any correction. The band shows ±2 model standard deviations about zero; it does not guarantee accuracy.
Innovations and their covariance describe the prior before correction. Points appear only at recorded arrivals in the window; no new value is created between corrections. A latest innovation from before the window can appear in the readout, explicitly labelled as not plotted.
Full-run mean NIS includes startup and does not change with this window. The angle KF has one measurement dimension; the vector EKF has two. Their ideal-model means differ, so their raw NIS values are not a shared lower-is-better score. Neither individual values nor the displayed mean establish statistical consistency.
Use the shared time slider, move across a plot, or step between corrections. Plot scales fit the window, including all its values; RMSE, mean NIS and repeated-trial scores continue to describe their full runs.
Full-run error
| Method | RMSE (°) |
|---|
About these scores
Across noise seeds
Repeated trials
| Method | Mean | Observed min–max | Displayed |
|---|
Trial protocol and limits
Each point summarizes all 3,001 states of a 30 s run, including initialization. Trials keep the scenario and filter settings fixed while changing noise draws; irregular timing also varies between seeds.
These scores always use sample-count weighting and the full run, independently of the playback cursor, diagnostic window and weighting selector above. The plot scale fits the inspected method and includes the displayed run.
The observed range is not a general robustness ranking or a hardware accuracy bound. Selecting a point inspects its full-run score; it does not load another trajectory.
Experiment settings and limitations
Playback reads recorded Python results. It does not rerun or retune the filters. The nominal/translation pair shares sensor-noise realizations. Controlled cases reuse noise draws by sample order; motion timing, bias, noise scale and acquisition delay can change the raw measurements. Scales cover all methods, including hidden curves; they may change when selecting another scenario.
Accelerometer tilt assumes gravity dominates the measurement. Translation can violate that assumption. This single-axis simulation does not establish accuracy on a physical drone, or embedded or flight performance.
Download displayed data and metricsRecorded comparisons / Vector EKF
Parameter studies.
One assumption changes. Measurements stay shared within each trial.
Loading the recorded studies…
All three settings
| Setting | Mean | Observed min–max | Mean difference |
|---|
Paired differences
Protocol, provenance and limits
All settings are retained. No default is changed by this view. Means and observed min–max ranges summarize these 20 trials; ranges are not confidence intervals or hardware accuracy bounds. Final evaluation sets have already been used and are not an untouched reserve for future tuning.
These studies use single-axis simulations. They do not establish bench, embedded or flight performance. The replay view keeps its original seed-42 trajectories and constant-bias settings; this view contains trial scores, not additional trajectories.
Download both studies, trial scores and provenance