[clmazzac]

Problem

IMU white noise was a fixed standard deviation per sample (angularVelocityStdDev, linearAccelerationStdDev), whatever the rate. A real sensor is specified by noise density. Its per-sample noise grows with sample rate, and noise averaged over a fixed time does not change.

In the sim, raising the IMU from 100 Hz to 400 Hz kept 0.002 rad/s per sample, so the noise left after averaging halved. The simulated IMU got better at higher rates. A filter tuned from a datasheet would get the wrong noise level at any rate other than the one the per-sample values were picked for.

Change

• New optional calibration settings angularVelocityNoiseDensity (rad/s/√Hz) and linearAccelerationNoiseDensity (m/s²/√Hz). Each is one number for all three axes.
• With a density, each sample uses σ = density / √Δt. Δt is the time since the previous capture, so σ follows the actual capture rate after fixed-step rounding. The message covariance uses the same σ.
• Without a density, the per-sample settings apply as before. Saved and default runs do not change.
• The Config page shows both densities under the IMU advanced fields. The per-sample fields are relabelled as the fallback.
• docs/run-manifests.md explains the model and gives conversions from datasheet units (°/s/√Hz, °/√h, µg/√Hz, m/s/√h).

Why the default run is unchanged

Setting densities on the default run would change its hash. That breaks the pinned golden in VIS-02 explicit-camera v11 identity vectors (tests/render-scene-provider.test.js). Making densities the default is a separate decision that needs that fixture regenerated. 0.0002 rad/s/√Hz and 0.002 m/s²/√Hz reproduce the current defaults exactly at 100 Hz.

Tests

Live sim, car parked, 2.5 ms step, about 20 s of sim time per run, IMU read back from the recorded SFLog. Units: rad/s for the gyro, m/s² for the accelerometer.

Run	Gyro per sample	Reported σ	Gyro 0.1 s avg	Accel 0.1 s avg
Density, 100 Hz	0.00195	0.00200	6.1e-4	6.6e-3
Density, 400 Hz	0.00402	0.00400	6.2e-4	5.8e-3
Per-sample, 100 Hz	0.00196	0.00200	6.1e-4	6.6e-3
Per-sample, 400 Hz	0.00201	0.00200	3.1e-4	2.9e-3

• With densities, the 0.1 s averages match the predicted 6.3e-4 and 6.3e-3 at both rates.
• At 100 Hz the density and per-sample runs give identical values.
• Lag-1 autocorrelation of the gyro is about 0, so samples stay white.
• npm test on a clean checkout: 1964 pass, 6 skipped, 1 fail. The failure is Python decodes retained RGB, depth, and CameraInfo SFLog messages, which also fails on main.
• No unit test is added in this PR.

Noticed while testing, not changed here

The recorded logs are missing IMU messages. Only about half of the 0.1 s windows in each run had every sample. The averages above use complete windows only.

🤖 Generated with Claude Code