KalmanNet: Hybrid State Estimation

• KalmanNet is a hybrid state estimation framework that combines the model-based Kalman filter recursion with data-driven RNN learning.
• It adapts to unknown noise statistics and nonlinear system behaviors by directly learning the Kalman gain from data.
• The architecture delivers robust performance in diverse applications such as control, robotics, and real-time signal processing.

KalmanNet is a family of hybrid model-based/data-driven state estimation architectures that integrates the recursive structure of the classical Kalman filter (KF) with a compact deep recurrent neural network (RNN) to learn the Kalman gain directly from data. KalmanNet was originally proposed to address the limitations of analytic KFs in settings with unknown or mismatched noise statistics, partial system knowledge, model nonlinearities, or practical constraints such as high-dimensionality, non-Gaussianity, or quantized observations. Unlike black-box RNN approaches, KalmanNet preserves the interpretability, inductive bias, and data efficiency of classical filtering while adapting to complex regimes through end-to-end training.

1. Hybrid Model-Based/Data-Driven Structure

KalmanNet is built upon the discrete-time state-space model

$$x_t = F x_{t-1} + w_t, \quad w_t \sim \mathcal{N}(0,Q)$$

$y_t = H x_t + v_t, \quad v_t \sim \mathcal{N}(0,R)$

where $F$ and $H$ are known system matrices, while the noise covariances $Q$ and $R$ are typically unknown or mismatched in realistic applications (Revach et al., 2021, Revach et al., 2021).

The architecture follows the two-step recursion of the classical KF:

• Model-based prediction: Computes predicted state and observation using (possibly nonlinear) models:

$$\hat{x}_{t|t-1} = F \hat{x}_{t-1}, \quad \hat{y}_{t|t-1} = H \hat{x}_{t|t-1}$$

• Data-driven update: The innovation $\Delta y_t = y_t - \hat{y}_{t|t-1}$, together with auxiliary features (such as previous innovations or predicted states), is input to a small RNN (typically a GRU). The RNN, with parameters $\theta$, produces an estimate of the Kalman gain $K_t(\theta)$. The update is:

$$\hat{x}_t = \hat{x}_{t|t-1} + K_t(\theta) \Delta y_t$$

This setup mirrors the recursive flow of the original KF but bypasses the need for closed-form covariances (Revach et al., 2021, Revach et al., 2021).

2. Training Schemes: Supervised, Unsupervised, and Online Adaptation

KalmanNet can operate in supervised and unsupervised regimes:

• Supervised Training: Minimizes the MSE between predicted and ground-truth states over observed sequences:

$$\mathcal{L}(\theta) = \frac{1}{T} \sum_{t=1}^T \|x_t - \hat{x}_t(\theta)\|_2^2 + \gamma \|\theta\|_2^2$$

where $\gamma$ is an $\ell_2$ regularization weight (Revach et al., 2021).

• Unsupervised Training: Exploits the KF structure, where predicted observations $\hat{y}_{t|t-1}$ are available internally. The RNN is trained to minimize the $\ell_2$ error between predicted and actual observations (innovation loss), without requiring state labels:

$$\tilde{\ell}_i(\theta) = \frac{1}{T_i}\sum_{t=1}^{T_i} \|\hat{y}_{t|t-1}^{(i)}(\theta) - y_t^{(i)}\|_2^2 + \gamma \|\theta\|_2^2$$

This approach enables genuine unsupervised learning for state estimation and model adaptation (Revach et al., 2021).

• Online/Adaptive Training: Parameters can be updated online by re-training on the most recent data window, adapting to time-varying or abruptly changing models ($F$, $H$) and noise ($Q$, $R$). The innovation loss rapidly drives the network to minimize prediction errors under evolving dynamics (Revach et al., 2021).

3. Kalman Gain Learning: RNN Design and Feature Engineering

The core of KalmanNet is the replacement of the analytic computation for $K_t$ by a data-driven module:

• RNN Input Features: Typical choices include current and previous innovations, predicted observations, and optionally the previous RNN hidden state.
• RNN Architecture: Compact variants with one or several GRU or LSTM layers, often sandwiched between input/output linear layers. In multi-factor formulations (e.g., Split-KalmanNet), separate RNNs learn different factors corresponding to process and measurement statistics.
• Parameterization: The gain output $K_t(\theta)$ has the shape appropriate to the system (e.g., $m \times n$).

The RNN is trained end-to-end through time using backpropagation-through-time (BPTT), regardless of whether state labels are available (Revach et al., 2021, Revach et al., 2021, Choi et al., 2022).

4. Theoretical and Empirical Performance

KalmanNet has been extensively validated across diverse regimes:

• Linear SSM, Known Topology: KalmanNet with supervised learning matches the minimum MSE of the ideal KF. Unsupervised KalmanNet achieves identical accuracy, validating its use without explicit noise statistics or state labels (Revach et al., 2021).
• Model Mismatch: For unknown or drifting $Q$, $R$, or time-varying $F$, $H$, online adaptation of KalmanNet tracks optimal MSE as the noise regime changes, outperforming a fixed-parameter filter (Revach et al., 2021).
• Nonlinear Models: KalmanNet extends directly to nonlinear (EKF-type) systems by propagating first moments through the known $f(\cdot)$, $h(\cdot)$ and learning the correction step, bypassing Jacobians, analytic covariance, and inverses (Revach et al., 2021, Revach et al., 2021). For chaotic and mismatched systems (e.g., the Lorenz attractor), KalmanNet reduces MSE by $\sim$0.5 dB compared to EKF and is notably faster at inference time.
• Convergence: Supervised KalmanNet requires fewer epochs (~20) than unsupervised (~80) for convergence to optimal MSE (Revach et al., 2021).

Summary of representative results:

Model/Scenario	KalmanNet (Unsupervised)	Classical KF	EKF
Linear SSM (Known)	Matches MMSE	MMSE	–
Linear SSM (Drift)	Approaches MMSE	Fails	–
Lorenz Attractor	$\sim$0.5 dB worse	–	Reference

(Revach et al., 2021)

5. Extensions: Robustness and Generalizations

The KalmanNet framework has spawned a broad array of extensions:

• Split-KalmanNet: Decomposes gain estimation into separate RNNs for process and innovation covariances, with the structure

$K_t = G_t^{1} H_t G_t^{2}$

This enhances robustness to mismatches in $Q$, $R$, and enables alternate optimization of decoupled modules (Choi et al., 2022).

• Adaptive KalmanNet (AKNet): Employs a small hypernetwork for conditional modulation, enabling the core RNN's weights to adapt continuously to time-varying noise by modulating on the ratio SoW$_k = q_k^2 / r_k^2$. AKNet generalizes well to unseen noise regimes and accommodates abrupt noise jumps without retraining (Ni et al., 2023).
• Recursive KalmanNet: Enhances uncertainty quantification by propagating covariance matrices via the Joseph's recursion. A twin-RNN architecture jointly outputs both the gain and an error-covariance Cholesky factor, with a negative-log-likelihood loss, ensuring consistency between predicted uncertainty and observed errors (Mortada et al., 13 Jun 2025).
• GSP-KalmanNet: Generalizes KalmanNet to tracking time-varying graph signals by leveraging the graph spectral domain, producing scalable, robust graph-filtered gain matrices, and improving computation for high-dimensional systems (Buchnik et al., 2023).
• Latent-KalmanNet: Integrates a learned encoder for high-dimensional observations (e.g., images) with KalmanNet operating in the induced latent space, optimizing the downstream MSE end-to-end (Buchnik et al., 2023).
• Bussgang-aided KalmanNet: Combines Bussgang decomposition with KalmanNet RNN gain learning to enable state estimation under severe quantization (1-bit observations), retaining robustness even with extreme quantization and model mismatch (Jung et al., 23 Jul 2025).
• Q-Net, SIKNet, and others: Variants have been employed in transfer learning for traffic state estimation and multi-object tracking, integrating semantic-independence, structured weight sharing, and local feature grouping for high interpretability and transferability, corroborated by empirical improvements in domain-specific metrics (Gao et al., 29 Sep 2025, Song et al., 14 Sep 2025).

6. Limitations and Theoretical Implications

KalmanNet's strengths and shortcomings are well delineated:

Advantages:

• Recovers or exceeds performance of the analytic KF under partial/noise mismatch, nonlinear dynamics, and partial observability.
• No ground-truth states needed in unsupervised modes; gains can adapt to real-time nonstationarities (Revach et al., 2021).
• Preserves inductive bias via model-based prediction, minimizing network size and training data requirements.
• Architecture is interpretable, with the learned gain conforming to the Kalman update structure (Revach et al., 2021).

Limitations:

• Convergence may be significantly slower in unsupervised versus supervised training.
• Correct topology ($F$, $H$) must be known; the unsupervised approach is not robust to large model mismatches in system matrices (Revach et al., 2021).
• Covariance consistency is not guaranteed unless specific procedures (e.g., Joseph's formula, Bayesian sampling) are introduced (Mortada et al., 13 Jun 2025, Dahan et al., 2023).
• Real-world safety-critical deployment is limited by reliability and robustness issues relative to strong model-based schemes such as IMM filters, especially under fast maneuvers or data regimes with limited prior knowledge (Mehrfard et al., 2024).

Potential extensions include joint learning of the system matrices and gain, integration with variational or particle-based methods for non-Gaussian models, and mixture/expert-based architectures for switching or time-varying regimes (Revach et al., 2021).

7. Impact, Applications, and Future Research

KalmanNet and its derivatives have demonstrated efficacy and versatility across diverse fields:

• Control and Robotics: Robust filtering under model and sensor uncertainty, simultaneous localization and mapping (SLAM) (Choi et al., 2022).
• Autonomous Vehicles: State estimation for localization and tracking—though rigorous reliability for safety-critical scenarios remains a challenge (Mehrfard et al., 2024).
• Traffic Systems: Queue length estimation with strong transferability and interpretability, outperforming classical and black-box methods (Gao et al., 29 Sep 2025).
• Signal Processing: Tracking and uncertainty quantification under high-dimensional, graph-based, quantized, or non-Gaussian regimes (Buchnik et al., 2023, Jung et al., 23 Jul 2025).

Ongoing research explores (1) seamless integration with physical priors and adaptive uncertainty estimates for safety, (2) interpretable, scalable architectures for high-dimensional and structured signals, (3) robust online adaptation to abrupt, nonstationary environments, and (4) principled end-to-end training objectives that couple state accuracy with credible uncertainty quantification.

KalmanNet has established a paradigm for hybrid filtering that unifies model-based interpretability with data-driven adaptability, setting a benchmark for further developments in learning-augmented state estimation (Revach et al., 2021, Revach et al., 2021, Choi et al., 2022, Ni et al., 2023, Dahan et al., 2023, Buchnik et al., 2023, Buchnik et al., 2023, Jung et al., 23 Jul 2025, Mortada et al., 13 Jun 2025, Gao et al., 29 Sep 2025, Song et al., 14 Sep 2025, Piperigkos et al., 2024, Mehrfard et al., 2024).