Struct ImuNoise 

pub struct ImuNoise {
    pub gyro_white: f32,
    pub accel_white: f32,
    pub gyro_bias_walk: f32,
    pub accel_bias_walk: f32,
}

IMU noise, as the continuous-time densities of equations (16)–(21).

Fields

gyro_white: f32

Gyroscope white noise, rad s⁻¹ / √Hz.

accel_white: f32

Accelerometer white noise, m s⁻² / √Hz.

gyro_bias_walk: f32

Gyroscope bias random walk, rad s⁻² / √Hz.

accel_bias_walk: f32

Accelerometer bias random walk, m s⁻³ / √Hz.

Trait Implementations

impl Clone for ImuNoise

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 ImuNoise

impl Debug for ImuNoise

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

Formats the value using the given formatter.

impl Default for ImuNoise

fn default() -> Self

White noise at ten times PX4 EKF2’s density, measured to be what this filter needs; bias walks at ArduPilot EK3’s density.

These are deliberately far above what an IMU datasheet or an Allan variance plot gives for the sensor alone, because the process noise of a real airframe absorbs what the model leaves out: vibration, scale-factor and cross-axis error, timing jitter, and the coning and sculling a first-order propagation does not capture. Datasheet-grade figures would make the covariance claim a precision the estimate does not have, and an overconfident covariance gates out measurements that were fine.

Both estimators add a bias walk as (σ Δt)² per prediction step, so their σ is per-step at their own rate and the density it implies is σ √Δt: PX4 at src/modules/ekf2/EKF/covariance.cpp:152-153,166-167 with a 10 ms step (module.yaml:22, EKF2_PREDICT_US), at c4e4ef98e9; ArduPilot at libraries/AP_NavEKF3/AP_NavEKF3_core.cpp:1065,1071 with a 12 ms step (AP_NavEKF3_core.cpp:36, EKF_TARGET_DT), at 368dc0c428, its bias states being b Δt. PX4’s 1.0e-3 rad s⁻² and 3.0e-3 m s⁻³ (params_gyro_bias.yaml:19, params_accel_bias.yaml:19) come to 1.0e-4 and 3.0e-4; ArduPilot’s copter 1.0e-3 and 2.0e-2 (AP_NavEKF3.cpp:30-31) to 1.1e-4 and 2.2e-3. These are ArduPilot’s, the larger pair.

The conversion is worth most where only tilt · g + b_a is observed: read unconverted, the walk drags a grounded vehicle’s tilt to a 5.2° peak against EKF2’s 1.08° (2c42096b), and converted it peaks at 2.5°.

The white noise is PX4’s per-step σ taken as a density, which is ten times the density PX4’s 1.5e-2 rad s⁻¹ and 0.35 m s⁻² stand for at its 10 ms step (module.yaml:118-131). The factor stands for two things this filter does not model: the floors both estimators put under a receiver’s reported accuracy, and vibration, which PX4 meets only by inflating accelerometer noise on clipping (covariance.cpp:125-133). Replay measured both spending it: at 0.3× accel_white, 2c42096b vibrating on the ground peaks at 4.6° of tilt against 2.5°, and a299e722’s raw R rejects 493 velocity solutions against 283. A change that models either is the moment to measure this again.

The floor the sensors themselves measure, StaticWindow::noise, sits 9 to 180 times below the gyroscope’s and 6 to 230 times below the accelerometer’s on the corpus, so the factor is measured from below too (evidence).

impl PartialEq for ImuNoise

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

Equality operator ==.

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

Inequality operator !=.

impl StructuralPartialEq for ImuNoise