acceptodds
Under review as a conference paper at ICLR 2027

Learning Meta-Dynamics for Off-Dynamics Kalman Filtering

Abstract

Recursive state estimation becomes challenging when the transition model itself evolves during filtering: each recursion provides only one new noisy observation, yet the transition must be revised before state and covariance propagation. We formulate this setting as off-dynamics Kalman filtering and view the causal evolution of the transition model as meta-dynamics. We introduce the Dynamics-Adaptive Kalman Network (DyAKNet), which represents dynamics adaptation explicitly in a sparse coefficient space. DyAKNet converts the diagnostic innovation under the preceding model into coefficient-space evidence, aggregates evidence across recursions, and separately updates signed coefficient values and active support before prediction. The revised transition is then immediately reused within the same filtering recursion. Under local smoothness, boundedness, and contraction conditions, we derive a finite-horizon mean-square error bound that characterizes the contributions of vector-field mismatch and stochastic noise. Experiments span coefficient evolution, structural changes within a prescribed candidate representation, unseen noise and switching patterns, and real-world localization. Across these settings, DyAKNet consistently improves state-estimation accuracy over the evaluated adaptive and learning-based baselines, while retaining an explicit and interpretable dynamics representation. These results support same-recursion dynamics adaptation as an effective strategy for filtering under evolving model mismatch. Code is available at https://anonymous.4open.science/r/DyAKNet-191A.

Primary area:
applications to physical sciences (physics, chemistry, biology, etc.)
Keywords:
State Estimation, Neural Kalman Filtering, Sparse Dynamics Modeling

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_bxEcOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes DyAKNet for recursive state estimation in systems with continually evolving dynamics. It combines innovation-based coefficient attribution with GRU-based recurrent memory to update sparse dynamics coefficients and their soft support within a predefined library, then uses the updated model in the same EKF recursion. The paper provides a local filtering-error bound and evaluates the method on synthetic systems, target tracking, and NCLT localization, reporting improved state-estimation accuracy over the evaluated baselines.

Critical strengths
  1. DyAKNet provides an explicit mechanism for adapting dynamics before the current prediction and correction, combining innovation-based coefficient attribution with recurrent memory and separate value/support updates. This is a meaningful contribution to filtering under evolving dynamics.

  2. The experiments is extensive, covering both simulated systems and the real-world NCLT dataset. The method achieves the lowest reported mean MSE across all nine nonlinear-system settings and the lowest reported MSE on NCLT. This consistency across different mismatch conditions supports the practical value of the approach.

Critical weaknesses
  1. Dependence on supervision and limited evidence of practical generalization. DyAKNet uses ground-truth coefficient supervision in the synthetic experiments and true state supervision on NCLT. Such labels may be costly or unavailable in practical applications, yet the dependence on their availability and quality is not quantified.

  2. The NCLT evaluation includes only five short test trajectories from a single recording, leaving transfer to longer trajectories and different recording sessions without additional labeled training insufficiently validated.

  3. All synthetic experiments use full-state observations, h(x)=Ixh(x)=Ix, leaving coefficient-space attribution under partial observations insufficiently tested. Although NCLT uses velocity-only measurements, it provides limited evidence for adaptation in partially observed nonlinear systems.

Questions for authors
  1. How dependent is performance on ground-truth coefficient or state supervision? Can the authors quantify performance when supervision is reduced or removed?

  2. Do the NCLT improvements persist on longer trajectories or recordings from different sessions, without additional state-supervised training on the target session?

  3. How effective is coefficient-space attribution in nonlinear systems where only a subset of the states is observed?

  4. How sensitive is DyAKNet to misspecified noise covariances QQ and RR, given that the reported noise-generalization experiments assume the current covariances are known?

I would be open to revising my rating upward if the authors’ clarifications and supporting evidence adequately address these concerns.

Reviewer_XP2nOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper considers the problem of filtering when the underlying dynamics model changes.

The state dynamics is assumed to be a non-linear dynamics of the SINDy form (roughly), xt+1=xt+Ξtθ(xt)x_{t+1} = x_t + \Xi_t \theta(x_t), where θ\theta is a non-linear lifting of the state, and Ξt\Xi_t is a linear map. In standard models Ξt\Xi_t is independent of tt.

The state xtx_t is observed indirectly, via the linear observation function hh, i.e. the observations are yt=hxt+noisey_t = h x_t + \text{noise}.

When the dynamics Ξ\Xi is fixed, one of the standard approaches to the problem is via the extended Kalman filter, EKF.

This paper proposes to learn the Ξt\Xi_t. Two settings are considered. The first is when the true Ξt\Xi_t are known in training, along with the noise variances. Then the proposed method is to run the standard EKF, where Ξt\Xi_t are provided by a GRU network taking previous innovations and the current filtered state as inputs. The GRU is trained with the objective of matching the ground truth Ξt\Xi_t. In inference, one then runs the EKF with the predicted Ξt\Xi_t.

In the other setting, the NCLT dataset, the special problem structure allows part of the state to serve as labels, which are used to train the GRU instead of the ground truth Ξt\Xi_t.

Evaluation is provided on four synthetic systems and on the NCLT dataset.

Critical strengths

The problem of filtering when the underlying dynamics model changes is important.

The proposal in this paper, using observed true dynamics of the model during training to learn a transition model for the dynamics itself, is reasonable, if one accepts the assumption that the ground truth dynamics Ξt\Xi_t (or another sort of privileged information suitable as labels) is available during training.

The experiments show improvement over baselines.

Critical weaknesses

There are considerable issues with novelty, scope of the paper (i.e. strong assumptions), and evaluation.

Novelty.

  • The problem itself is well known, and the closest recent work, EKF-SINDy (Rosafalco et al., 2025), is compared. A second 2025 paper on Kalman filtering of time-varying SINDy coefficients (Pillonetto, Yazdani and Aravkin, "On-line learning of dynamic systems: sparse regression meets Kalman filtering", arXiv 2511.11178) and the learned Kalman filter line that followed KalmanNet are not cited.

  • More importantly, the technical recipe itself, modelling an unknown quantity with a GRU in the EKF setting, has been introduced before, in KalmanNet, where it was used to model unknown covariances. The code here is a straightforward modification of their code.

Scope.

  • Assumed access to privileged data. As discussed in the summary, this method assumes knowledge of the ground truth dynamics in training, or at least some sort of labels to guide the learning. While this may be reasonable in some cases, the most common case has training data which is of the same form as inference data. There is no consideration of this case in the paper. This restricts the generality of the approach.

  • Noise variances are assumed to be known in this paper. Noise variances are an important part of the dynamics, and assuming they are known further restricts the generality.

As presented, this is not a general filtering method, and I consider this the decisive issue.

Evaluation.

  • The evaluation is done on toy synthetic examples. The examples of Table 1 have linear dynamics of Ξ\Xi, and it is the same dynamics on all trajectories, and it is also the same one both on train and test sets. The experiments in Tables 2 and 3 are likewise toy synthetic examples. Given the novelty discussion above, I would expect the paper's contribution to come mainly from evaluation, which should be much more thorough than this.

  • For the NCLT example, a comparison with KalmanNet is missing. Given that KalmanNet introduced this task and obtained some results on it, it is not clear why it is not included.

  • No code for how most of the benchmarks were run was provided.

Questions for authors

Please discuss the items in the weakness section.

Reviewer_e51rOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper concerns the problem of designing a filter that is able incorporate adaptive dynamics into it. It is a common occurrence in practice that when a fixed filter is used, the transition dynamics imposed in the model can no longer reflect the data generating process accurately. It is therefore a sensible approach to consider the transition process between transition models themselves. This is termed meta-dynamics by the authors. The authors' method is centered around learning a causal dynamics-update operator which can be used upon the availability of an observation, and hence for state and covariance propagation. The method is tested on standard tracking problems with promising results.

Critical strengths

This kind of learning of a dynamical correction to the transition dynamics is sensible and seems like an approach that could produce an adaptive behaviour to transitions, provided that the learned network that modifies the dynamics is sufficiently flexible and well-trained.

Critical weaknesses

The major weaknesses of this method are 1) the fact that it has to be trained on complete trajectories and 2) it is not presented in a way that makes it usable online. The main purpose of methods that can change the transition is to adapt to changing dynamics as the data comes in, without necessarily having access to large amounts of data for model pre-training (and subsequent amortization).

Questions for authors

How well would the method do in an online setting? Can this method in principle be adapted to an online variant?

The experiments as presented are unclear. Are you 1) generating complete trajectories, 2) training on them, and then testing your method on held-out trajectories from the data generating process? What are the trajectory lengths? How many do you need to train your model?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.