Recalibrated EKF Framework

• The recalibrated EKF framework is a set of advanced techniques that enhance state estimation by addressing linearization errors, sensor biases, and uncertainty in nonlinear systems.
• It integrates learning-based recalibration, affine and transformation-based updates, and covariance compensation to restore statistical consistency and correct observability.
• Practical applications in robotics, navigation, and sensor fusion show marked improvements in RMSE and NEES, validating both theoretical guarantees and real-world performance.

A recalibrated Extended Kalman Filter (EKF) framework refers to any systematic enhancement of the EKF that explicitly addresses inaccuracies or inconsistencies arising from model nonlinearity, sensor miscalibration, observability mismatch, or erroneous uncertainty quantification. Modern recalibrated EKF paradigms include learning-based observation model recalibration, affine/coordinate-transformed EKFs, explicit covariance compensation, distributionally robust estimation, and online adaptive noise inference. The overarching goal is to restore first-principles properties—such as statistical consistency, correct observability, and accurate uncertainty quantification—to recursive stochastic filters in nonlinear, multimodal, or real-world sensor fusion settings.

1. Motivation for EKF Recalibration

Standard EKFs, derived as local linearizations of nonlinear process and measurement models, are widely used in robotics, navigation, and sensor fusion. Despite their success, vanilla EKFs exhibit several critical deficiencies:

• Model linearization mismatch: Covariance recursion is based on first-order Taylor approximations, which systematically underestimate uncertainties for strongly nonlinear models (Jiang et al., 2024).
• Observability inconsistency: The linearized EKF can spuriously gain information along directions that are unobservable in the underlying nonlinear system, leading to overconfident and inconsistent state estimates (Song et al., 2024, Hao et al., 7 Feb 2025).
• Fixed or poorly tuned process/measurement noise: Mismodeling or calibration errors in $Q$, $R$, or sensor observation models corrupt both mean and covariance propagation, resulting in filter divergence or degraded accuracy (Cohen et al., 2024).
• Sensor complexity and black-box characteristics: Modern sensors may not admit closed-form parametric models for $z = g(x)$; physical calibration is laborious or infeasible, motivating data-driven alternatives (Guo et al., 18 Jan 2026).
• Safety and certification needs: Application domains such as robotics and autonomous systems require verifiable bounds for estimation error and rigorous consistency checks (Jang et al., 3 Apr 2026).

These limitations have motivated a suite of recalibrated EKF methodologies, each targeting distinct sources of failure.

2. Koopman-Inspired Learned Measurement Recalibration

The KILO-EKF ("Koopman-Inspired Learned Observations EKF") exemplifies a data-driven, measurement model–level recalibration (Guo et al., 18 Jan 2026). Its core innovation is to replace poorly calibrated, nonlinear geometric measurement models with a learned, linear map in a nonlinear feature (lifted) space:

• Sensor data lifting: Define a nonlinear map $\phi: \mathbb{R}^m \rightarrow \mathbb{R}^N$ applied to raw measurement $z_k$, producing $y_k = \phi(z_k)$.
• Linear-Gaussian measurement model in lifted space:

$$y_k = H x_k + b + w_k,\quad w_k \sim \mathcal{N}(0, R)$$

where $(H, b, R)$ are learned from annotated data via closed-form least squares/MAP without iterative optimization.

• EKF update in lifted space: For each measurement, compute innovation, innovation covariance, and gain with respect to $\phi(z_k)$ using the learned $(H, b, R)$. The correction step is structurally equivalent to a linear EKF, with Jacobians obtained from the learned lifting, and preserves the recursive, efficient update properties.
• Scalable to complex/unmodeled sensors: Sensor-specific nonlinearities, range/squashed-range, or multipath effects are captured without explicit geometric modeling.

Empirical results with quadrotor localization show substantial reductions in RMSE and improved NEES compared to both miscalibrated and data-calibrated EKF baselines.

3. Affine, Transformation-Based, and Symmetry-Invariant Extended Kalman Filters

Modern theoretical analyses have established that inconsistency in EKFs for partially observable nonlinear systems arises when the EKF's linearized subspace of unobservable directions ("unobservable subspace") fails to match that of the true system, typically due to state-dependence lost in the EKF's linearization. Multiple frameworks restore consistency via explicit state-space recalibration:

• Affine EKF (Aff-EKF): Constructs an affine coordinate transformation, $R$0, such that the unobservable subspace is rendered state-independent. The transformed system is then linearized and estimated in these new coordinates (Song et al., 2024). The procedure is fully algebraic and sidesteps the need for fixed linearization points or Lie group structures.
• Transformation-based EKF (T-EKF): Identifies a linear, time-varying transformation $R$1 so that the transformed system's unobservable subspace is constant and matches that of the nonlinear process. Two forms are available: T-EKF 1 (fully transformed coordinates) and T-EKF 2 (retrofit to existing EKF, applying gain and covariance corrections post-update via $R$2) (Hao et al., 7 Feb 2025). This framework is applicable to multi-robot cooperative localization, target tracking, and visual-inertial odometry.
• Invariant EKF / Error-State on Manifolds: Utilizes symmetry-induced group errors ($R$3, e.g., right-invariant error on $R$4) and retraction/linearization compatible with group action to ensure preservation of unobservable directions and statistical consistency (Brossard et al., 2019). Assumption: compatibility between error map and transformation group.

These consistency-guaranteed EKF variants rigorously match or outperform invariant and fixed-point approaches (e.g., FEJ-EKF, RI-EKF) both in RMSE and in statistical consistency metrics such as NEES.

4. Covariance Compensation and Robust Uncertainty Quantification

A central recalibration axis is to explicitly compensate for uncertainty underestimation by augmenting the EKF covariance:

• Covariance-compensated EKF: Introduces a positive semi-definite correction term, $R$5, defined as the difference between the predicted covariance of a nonlinear KF (UKF, CKF, EKF2) and the EKF:

$R$6

The recalibrated update uses $R$7 (Jiang et al., 24 Mar 2026).

• Design guidelines: Three foundational requirements for the recalibrated EKF are established:
  1. Covariance compensation must be invariant under orthogonal transformations and PSD,
  2. Sufficient compensation $R$8 must exceed the lower bound set by EKF2 with radially symmetric priors,
  3. Tuning should bias toward underconfidence; the minimum MSE in post-update covariance occurs for $R$9.

Empirical Monte Carlo studies across several estimation problems validate improved actual vs. estimated RMSE and a marked reduction in overconfidence.

• Distributionally robust EKF (DR-EKF): Recasts the EKF estimation step as a minimax optimization over a Wasserstein ball of probable noise distributions, incorporating linearization residuals as explicit error sources (Jang et al., 3 Apr 2026). The stage-wise robust update solves a semidefinite program yielding a re-optimized covariance, robust gain, and deterministic error-bounds.

These schemes yield provable and empirically validated statistical consistency and safety bounds.

5. Adaptive and Learning-Based EKF Recalibration

Contemporary frameworks treat the remaining uncertainty due to changing system behaviour, hardware, or poorly modeled noise as a learning problem:

• Neural calibration maps: Learn a corrective mapping $z = g(x)$0 from the EKF's raw posterior covariance to a recalibrated covariance, based on ground-truth or ergodic sample estimates (Tsuei et al., 2021). Implementations include scalar, linear, and deep neural network architectures, and are trained via supervised (ground truth) or self-supervised (innovation statistics) means. The recalibration is injected at each EKF step as $z = g(x)$1.
• Online adaptive process noise via transformers: A-KIT (Adaptive Kalman-Informed Transformer) uses a set-transformer to infer a diagonal scaling for the process noise covariance $z = g(x)$2 from batches of sensor updates and innovation history, with Q updated per-timestep in real time (Cohen et al., 2024). The end-to-end Kalman-informed loss trains the entire pipeline.
• Koopman-inspired measurement learning: The KILO-EKF method (see above) can be viewed as a special case where the measurement model is recalibrated via (nonparametric) regression in a lifted feature space.

Quantitative results across navigation, vision, and robotic localization demonstrate significant improvements (>30–90%) in both state accuracy and covariance reliability.

6. Practical Implementations and Applications

Recalibrated EKF architectures have been deployed and validated in:

• Multi-sensor extrinsic calibration: Joint LiDAR–IMU–camera, multi-IMU-to-body calibration, and radar-inertial odometry with online temporal alignment. States include full extrinsic, bias, and offset parameters, with EKF updates tightly coupling calibration quantities with global states (Mishra et al., 2022, Hartzer et al., 2023, Kim et al., 2 Feb 2025).
• Visual-inertial odometry and SLAM: Real-world and synthetic datasets confirm full restoration of NEES to dimensionally consistent values, RMSE reductions, and cross-validation alignment with batch/optimization-based estimators (Tsuei et al., 2021, Song et al., 2024, Hao et al., 7 Feb 2025).
• AUV navigation, terrain-referenced tracking, target tracking with nonlinear measurements: Robust under heavy sensor noise/low process or measurement uncertainty, outperforming classical EKF, UKF, and CKF baselines (Jiang et al., 24 Mar 2026, Jang et al., 3 Apr 2026, Cohen et al., 2024).

Implementation complexity is generally negligible compared to the underlying EKF computations. For transformation-based recalibrations, T-EKF 2 and similar methods facilitate retrofitting to legacy EKF codebases via gain/covariance correction factors, requiring only Jacobian and algebraic operations per step.

7. Theoretical Guarantees and Statistical Consistency

The central contribution of recalibrated EKF frameworks is to guarantee, by construction:

• Statistical consistency of posterior uncertainty: Empirical NEES tracks the dimension, actual coverage matches predicted uncertainty, and overconfidence is eliminated even in the presence of strong nonlinearity or miscalibration (Song et al., 2024, Tsuei et al., 2021).
• Correct matching of system observability: Consistent filters preserve the unobservable subspace, ensuring that the EKF does not spuriously reduce uncertainty where the true system is unidentifiable (Hao et al., 7 Feb 2025, Brossard et al., 2019).
• Distributionally robust error bounds: DR-EKF approaches yield deterministic, computable upper bounds on the MSE between the filter state and the true state, including the effect of model mismatch and linearization error (Jang et al., 3 Apr 2026).
• No loss in real-time performance: All methods, including learning-based approaches, run in real time due to analytic formulas, fixed-dimensional updates, and negligible inference overhead.

These guarantees are established both theoretically—by geometric and algebraic analyses of the underlying dynamical systems—and validated via large-scale simulation and real-data experiments.

---

References:

• KILO-EKF: "Koopman-Inspired Learned Observations Extended Kalman Filter" (Guo et al., 18 Jan 2026)
• Affine EKF: "Affine EKF: Exploring and Utilizing Sufficient and Necessary Conditions for Observability Maintenance to Improve EKF Consistency" (Song et al., 2024)
• Transformation-based EKF: "A Transformation-based Consistent Estimation Framework: Analysis, Design and Applications" (Hao et al., 7 Feb 2025)
• Covariance Compensation: "Design Guidelines for Nonlinear Kalman Filters via Covariance Compensation" (Jiang et al., 24 Mar 2026)
• DR-EKF: "Residual-Aware Distributionally Robust EKF: Absorbing Linearization Mismatch via Wasserstein Ambiguity" (Jang et al., 3 Apr 2026)
• Learning-based Covariance: "Learned Uncertainty Calibration for Visual Inertial Localization" (Tsuei et al., 2021)
• Adaptive Kalman-Informed Transformer: "Adaptive Kalman-Informed Transformer" (Cohen et al., 2024)
• Joint LiDAR–IMU–Camera calibration: "Extrinsic Calibration of LiDAR, IMU and Camera" (Mishra et al., 2022)
• Multi-IMU calibration: "Online Multi-IMU Calibration Using Visual-Inertial Odometry" (Hartzer et al., 2023)
• Radar-inertial odometry temporal calibration: "EKF-Based Radar-Inertial Odometry with Online Temporal Calibration" (Kim et al., 2 Feb 2025)
• Invariant EKF on Lie groups: "Exploiting Symmetries to Design EKFs with Consistency Properties for Navigation and SLAM" (Brossard et al., 2019)
• New nonlinear KF framework: "A New Framework for Nonlinear Kalman Filters" (Jiang et al., 2024)