Struct Initialization 

pub struct Initialization {
    pub min_duration: Seconds,
    pub max_gyro_rate: RadiansPerSecond,
    pub max_accel_deviation: MetersPerSecond2,
    pub sigma_position: Meters,
    pub sigma_velocity: MetersPerSecond,
    pub sigma_tilt: Radians,
    pub sigma_yaw: Radians,
    pub sigma_accel_bias: MetersPerSecond2,
    pub sigma_gyro_bias: RadiansPerSecond,
}

Quasi-static initialization, equations (5)–(8).

Fields

min_duration: Seconds

How long the vehicle must have been still.

A duration rather than a sample count, because the same count means very different things across IMU rates: 100 samples is 2 s at 50 Hz and 0.25 s at 400 Hz, and a quarter second is too short to average sensor noise down or to tell stillness from a slow drift. The logs in data/manifest.txt alone span 50 Hz to 250 Hz.

max_gyro_rate: RadiansPerSecond

Largest angular rate magnitude still considered stationary.

Compared against the peak over the window, not a filtered value, so at the same number this is the stricter test: one vibration spike is enough to fail it.

max_accel_deviation: MetersPerSecond2

Largest departure of the specific-force magnitude from gravity still considered stationary. Peak over the window, as with max_gyro_rate.

sigma_position: Meters

Initial position standard deviation.

sigma_velocity: MetersPerSecond

Initial velocity standard deviation.

sigma_tilt: Radians

Initial roll and pitch standard deviation. Gravity determines these well.

The whole tilt uncertainty of a leveled start, including the share an accelerometer bias explains: that share, σ_βa / γ, is carried as the tilt’s correlation with the bias rather than added to it, and where it exceeds this figure it is the prior, with the window’s own scatter across gravity added as the share the bias does not explain (equation (8)). At the defaults it does, 0.0204 rad against 0.02.

sigma_yaw: Radians

Initial yaw standard deviation. Much larger than sigma_tilt: yaw inherits the magnetometer’s calibration error.

sigma_accel_bias: MetersPerSecond2

Initial accelerometer bias standard deviation.

0.2 m/s², the switch-on bias both production estimators assume: PX4’s EKF2_ABIAS_INIT (src/modules/ekf2/EKF/common.h:340, c4e4ef98), and ArduPilot’s 0.2 · EK3_ACC_BIAS_LIM at its default of 1.0 (libraries/AP_NavEKF3/AP_NavEKF3_core.cpp:616, AP_NavEKF3_core.h:52, AP_NavEKF3.cpp:567, 368dc0c4). A static window cannot measure it, since (5) reads a horizontal bias as tilt, which is why the tilt carries its correlation with it.

At 0.1, harsh_imu’s 0.186 m/s² bias sat at 1.86σ and its attitude was overconfident on 50 seeds (evidence).

sigma_gyro_bias: RadiansPerSecond

Initial gyroscope bias standard deviation, before the window’s mean rate is weighed.

A window at rest weighs its own mean rate against this, (7), and a still window measures the bias far better (evidence); so this is the prior of a start in motion, where the bias starts at zero, and no start at rest ends wider than it but for the Earth’s rotation, which (8) adds. PX4 starts every bias at zero under 0.1 rad/s (EKF2_GBIAS_INIT, src/modules/ekf2/params_gyro_bias.yaml:5-14, c4e4ef98), and ArduPilot’s fallback is 2.5°/s (libraries/AP_InertialSensor/AP_InertialSensor.cpp:1536-1542, 368dc0c4).

0.01 rad/s against both, because the corpus’s biases are smaller than either assumes (EKF2’s per-log medians 1.2 × 10⁻³ rad/s RMS, none past 4.5 × 10⁻³), and a wider prior lets a coarse start explain its attitude error as bias: at PX4’s 0.1, 7ce66f0d, the hand launch, rejects 58157 measurements against 4877 (evidence).

Trait Implementations

impl Clone for Initialization

fn clone(&self) -> Self

Returns a duplicate of the value.

fn clone_from(&mut self, source: &Self)

Performs copy-assignment from source.

impl Copy for Initialization

impl Debug for Initialization

fn fmt(&self, f: &mut Formatter<'_>) -> Result

Formats the value using the given formatter.

impl Default for Initialization

fn default() -> Self

The stationarity tolerances are PX4 EKF2’s — 15°/s and 20% of gravity — chosen against the replay corpus: tighter ones called four of its first five logs moving while they sat on the ground, on idle vibration and prop wash (evidence).

impl PartialEq for Initialization

fn eq(&self, other: &Self) -> bool

Equality operator ==.

fn ne(&self, other: &Rhs) -> bool

Inequality operator !=.

impl StructuralPartialEq for Initialization