EKF Reset

[dellaert]

Implemented "reset" covariance transport in the EKF filter hierarchy also added a notebook showing how the NEES value is better when reset is being used. This is now the default setting for all update methods in the manifold EKF.

[Copilot]

Pull request overview

This pull request implements "reset" covariance transport in the EKF filter hierarchy to improve filter consistency. The reset step transports the covariance from one tangent space to another after state retraction, which theoretically improves the Normalized Estimation Error Squared (NEES) metric. The feature is demonstrated through a Python notebook showing improved NEES values when reset is enabled.

Changes:

• Added reset() method to ManifoldEKF class that retracts state and transports covariance using the retract Jacobian
• Modified update() methods to accept optional performReset parameter (default true)
• Added type trait HasRetractJacobian to detect manifolds supporting Jacobian computation
• Added comprehensive tests including Monte Carlo validation

Reviewed changes

Copilot reviewed 5 out of 7 changed files in this pull request and generated 8 comments.

Show a summary per file

File	Description
gtsam/navigation/ManifoldEKF.h	Core implementation of reset() method, HasRetractJacobian trait, and updated update() methods
gtsam/navigation/tests/testManifoldEKF.cpp	New tests for reset functionality including ResetUsesRetractJacobian and Monte Carlo validation
gtsam/geometry/tests/testPose2.cpp	Test for Pose2 retract Jacobian computation
gtsam/geometry/tests/testGal3.cpp	Test for Gal3 retract Jacobians computation
gtsam/navigation/navigation.i	Updated wrapper interface to expose performReset parameter
python/gtsam/examples/Gal3ImuNEESReset.ipynb	Example notebook demonstrating NEES improvement with reset enabled
.gitignore	Added gtsam_unstable/timing/data/ entry

[scottiyio]

Interesting, never thought about resetting covariance. I don't see any issues

[dellaert]

Thanks. it’s the key step in the that makes the left and right invariant filter equivalent, in that ANU/Klagenfurt paper.

[dellaert]

@giuliodelama FYI