When the mag is pulled away before takeoff, the EKF can end up learning
a bad yaw gyro bias. This bad gyro bias makes the estimated heading spin
and even if the bias continues to be learned, it can end up being stuck
in limit cycle due to the innovation changing its sign every time the
heading estimate overshoots the measurement.
To prevent this, the Z gyro bias variance is reset when the mag heading
fusion test ratio fails on ground.
* added python script with ekf derivation (WIP)
Signed-off-by: RomanBapst <bapstroman@gmail.com>
* worked on c code auto-generation
Signed-off-by: RomanBapst <bapstroman@gmail.com>
* save before variable name change
Signed-off-by: RomanBapst <bapstroman@gmail.com>
* changed symbol names
Signed-off-by: RomanBapst <bapstroman@gmail.com>
* added codegeneration class
Signed-off-by: RomanBapst <bapstroman@gmail.com>
* improve 3D mag fusion derivation
Signed-off-by: RomanBapst <bapstroman@gmail.com>
* EKF: Extend ekf sympy derivation to include all observation types
* EKF: Add custom ecl::powf function for integer powers
* EKF: Convert ekf covariance prediction to use sympy output
* EKF: Add test program to compare sympy and matlab covariance prediction
Also tests ecl::powf(x,exp) function
* EKF: simplify ecl::powf function
* Generate code to subfolder generated/
* Add printouts for showing code generation progress
* Move generated covariance code to generated folder
* Upgrade code generation to python3
* main.py: Remove unused create_symbols function
& making code more compact
* main.py: move main part into function
* Code generation: fix passing wrong rotation matrix to yaw_observation ()
* EKF: Amend generated code filename for consistency
* Move ecl::powf function test to unit tests
* EKF: Use updated ecl:powf functionality in test program
* Move ecl::powf to utils.hpp
* Update ecl::powf test
* Update output change indication
* test: update expected output for change indicator
* test: update expected output for change indicator again
Co-authored-by: RomanBapst <bapstroman@gmail.com>
Co-authored-by: kamilritz <kritz@ethz.ch>
* ekf: disable xy accel bias learning before takeoff
As those biases are usually poorly observable before takeoff because
they are almost perpendicular to the gravity vector, learning is often
driven by noise and numerical issues. This results in incorrect bias
learning before takeoff when the drone is static on ground for a long
period of time.
* ekf: update unit test and change indicator
This is a non-functional change required to select accel bias estimation
per axis selection. The intent is then to disable the learning before
takeoff of the components that are poorly observable.
* EKF: Improve covariance prediction stability
Eliminates collapse of vertical velocity state variance due to rounding errors that can occur under some operating conditions.
* EKF: Fix typo
* test: Fix initialisation test cases
Provide sufficient time for variances to stabilise and fix calculation of reference quaternion for alignment.
* test: Allow for accumulated rounding error in IMU sampling test
* test: Allow sufficient time for quaternion variances to reduce after initial alignment
* test: Increase allowance for tilt alignment delay and inertial nav errors
* test: Increase allowance for tilt alignment delay and inertial nav errors
* adpat reset velocity test
* test: update change indication file
* test: Adjust tests to handle alignment time and prediction errors
* README.md: Add documentation for change indicator test
* ekf2: centralize mag covariance reset. A complete and clean reset of the
mag states covariances is now performed through the "resetMagCov"
function only. This avoid having slight differences of implementations
across the code.
* Ekf: reset quat cov with initial uncertainty instead of zero
This operation is expensive when done to the whole covariance matrix and unnecessary after covariance prediction because we calculate the upper diagonal and copy across so it is already symmetric.
* ekf_control: Inhibit mag fusion when field magnitude is large
Move mag inhibition check in separate function
* ekf_control: pull out of functionalities out of controlMagFusion
- yaw abd mag bias observability checks
- mag 3D conditions
- load mag covariances
- set and clear mag control modes
* ekf_control: refactor mag heading/3D start/stop.
Move mag declination, mag 3d and mag heading fusion out of the main function
* ekf_control: extract mag yaw reset and mag declination fusion requirements
* ekf_control: use WMM in isStronMagneticField for mag fusion inhibition
- Correct units of WMM strength table
* ekf_control: extract mag_state_only functionality of AUTOFW (VTOL custom)
Also split inAirYawReset from onGroundYawReset
* ekf_control: extract mag automatic selection
- transform if-else into switch-case for parameter fusion type selection
* ekf_control: extract run3DMagAndDeclFusion, reorganize functions, fix
flag naming in Test script
* ekf_control: do not run mag fusion if tilt is not aligned.
Reset some variables on ground even if mag fusion is not running yet. It
could be that it runs later so we need to make sure that those variables
are properly set.
* ekf_control: move controlMagFusion and related functions to mag_control.cpp
* ekf control: check for validity of mag strength from WMM and falls back
to average earth mag field with larger gate if not valid
* ekf control: remove evyaw check for mag inhibition
* ekf control: change nested ternary operator into if-else if
* Ekf: create AlphaFilter template class for simple low-pass filtering
0.1/0.9 type low-pass filters are commonly used to smooth data, this
class is meant to abstract the computation of this filter
* ekf control: reset heading using mag_lpf data to avoid resetting on an outlier
fixes ecl issue #525
* ekf control: replace mag_states_only flag with mag_field_disturbed and
add parameter to enable or disable mag field strength check
* ekf control: remove AUTOFW mag fusion type as not needed This was implemented for VTOL but did not solve the problem and should not be used anymore
* ekf control: use start/stop mag functions everywhere instead of setting the flag
* ekf control: Run mag fusion depending on yaw_align instead of tilt_align
as there is no reason to fuse mag when the ekf isn't aligned
* AlphaFilter: add test for float and Vector3f
angle- and delta velocity bias variance
- the contribution of process noise per iteration for these states can be so
small that it gets lost if using standard floating point summation
Signed-off-by: Roman <bapstroman@gmail.com>
bresch