pith. sign in

arxiv: 2605.18704 · v1 · pith:IZKTV3NUnew · submitted 2026-05-18 · 📡 eess.SP · cs.LG

Learned Memory Attenuation in Sage-Husa Kalman Filters for Robust UAV State Estimation

Kenan Majewski , Marcin \.Zugaj This is my paper

Pith reviewed 2026-05-20 08:08 UTC · model grok-4.3

classification 📡 eess.SP cs.LG
keywords Sage-Husa Kalman Filterrecurrent neural networkUAV state estimationadaptive filteringmemory attenuationsensor outagesinnovation sequence
0
0 comments X

The pith

A hierarchical recurrent network learns a vector memory attenuation policy to replace the fixed forgetting factor in Sage-Husa Kalman filters for UAVs.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper shows how to adapt the Sage-Husa Kalman Filter for non-stationary conditions by training a recurrent network to produce a vector of memory weights instead of using one fixed scalar. The network processes whitened innovation sequences through shallow layers for sudden anomalies and deeper layers for ongoing trends, with an extra reconstruction loss to keep the features useful. End-to-end training directly reduces state error while preserving the filter's recursive covariance structure. Tests on chaotic attractors and real UAV recordings indicate the method stays stable and accurate when sensors drop out or noise statistics shift.

Core claim

The N-Deep Recurrent Sage-Husa Filter replaces the static scalar forgetting factor of the Sage-Husa Kalman Filter with a vector-valued memory attenuation policy produced by a bifurcated recurrent network operating on whitened innovation sequences. Shallow recurrent states capture instantaneous sensor anomalies while deeper states encode sustained dynamic trends; an auxiliary reconstruction objective prevents feature collapse. The complete filter, including its recursive covariance updates, is trained end-to-end through backpropagation through time to minimize state estimation error.

What carries the argument

The N-Deep Recurrent Sage-Husa Filter (NDR-SHKF), a bifurcated recurrent network that generates a learned vector memory attenuation policy for online noise-statistic estimation inside the Sage-Husa recursion.

If this is right

  • The learned policy generalizes across topologically distinct chaotic attractors where purely data-driven estimators diverge.
  • The filter maintains accuracy during transitions to proprioceptive dead reckoning when external sensors are lost.
  • It outperforms classical adaptive estimators on recorded real-world UAV datasets during sensor outages.
  • End-to-end differentiability allows the covariance recursion to remain stable while the attenuation policy adapts.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • The same learned-attenuation idea could be inserted into other adaptive filters that rely on a forgetting factor.
  • Training on a wider variety of flight regimes might further reduce the need for manual tuning of process-noise models.
  • The whitened-innovation input representation may transfer to estimation tasks outside UAVs, such as robot arm tracking under vibration.

Load-bearing premise

The bifurcated recurrent architecture together with the auxiliary reconstruction objective produces a memory attenuation policy that improves accuracy without causing instability in the recursive covariance updates.

What would settle it

Observation of covariance-matrix divergence or higher state estimation error than the classical Sage-Husa filter on additional UAV flight segments containing similar telemetry outages would falsify the claim.

Figures

Figures reproduced from arXiv: 2605.18704 by Kenan Majewski, Marcin \.Zugaj.

Figure 1
Figure 1. Figure 1: The NDR-SHKF architecture, illustrating the flow from the whitened innovation [PITH_FULL_IMAGE:figures/full_fig_p008_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Side-by-side 3D state trajectories of the Lorenz and Rössler attractors. The plots [PITH_FULL_IMAGE:figures/full_fig_p020_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Coordinate-wise tracking performance (x, y, z) over time for both the Lorenz and Rössler dynamical systems. The plots compare the ground truth against the proposed NDR-SHKF architectures and various classical, adaptive, and learning-based baselines. The NDR-SHKF rejects instantaneous measurement spikes while maintaining tracking during continuous out-of-distribution maneuvers. All methods exhibit a baselin… view at source ↗
Figure 4
Figure 4. Figure 4: Cumulative Root Mean Square Error (CRMSE) evaluated over the 600-step [PITH_FULL_IMAGE:figures/full_fig_p022_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Distribution of Rössler out-of-distribution ARMSE across 100 training seeds [PITH_FULL_IMAGE:figures/full_fig_p023_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: The Quanser QDrone experimental platform within the OptiTrack motion cap [PITH_FULL_IMAGE:figures/full_fig_p027_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: UAV Position tracking (X, Y, Z) during the benchmark flight. The magenta [PITH_FULL_IMAGE:figures/full_fig_p037_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: UAV Attitude tracking (Roll, Pitch, Yaw). The magenta shaded regions denote [PITH_FULL_IMAGE:figures/full_fig_p038_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: UAV Position and Attitude RMSE over time. The magenta shaded regions [PITH_FULL_IMAGE:figures/full_fig_p039_9.png] view at source ↗
read the original abstract

Unmanned Aerial Vehicles in dynamic environments face telemetry outages, structural vibrations, and regime-dependent noise that invalidate the stationary covariance assumptions of classical Kalman filters. The Sage-Husa Kalman Filter (SHKF) estimates noise statistics online, but its reliance on a static, scalar forgetting factor forces a strict compromise between steady-state stability and transient responsiveness. We introduce the N-Deep Recurrent Sage-Husa Filter (NDR-SHKF), which replaces this scalar parameter with a vector-valued memory attenuation policy learned by a hierarchical recurrent network operating on whitened innovation sequences. A bifurcated architecture routes shallow recurrent states to capture instantaneous sensor anomalies and deep states to encode sustained dynamic trends, while an auxiliary reconstruction objective prevents feature collapse. The complete filter, including recursive covariance updates, is trained end-to-end via backpropagation through time to directly minimize state estimation error. Evaluations on topologically distinct chaotic attractors demonstrate cross-domain generalization, outperforming purely data-driven baselines that diverge under out-of-distribution dynamics. Furthermore, evaluations on recorded real-world UAV flight datasets validate the framework's practical viability, demonstrating its capacity to bridge transitions into proprioceptive dead reckoning and outperform classical adaptive estimators during sensor outages.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

1 major / 2 minor

Summary. The manuscript proposes the N-Deep Recurrent Sage-Husa Filter (NDR-SHKF) for robust UAV state estimation under non-stationary noise, telemetry outages, and vibrations. It replaces the scalar forgetting factor of the classical Sage-Husa Kalman filter with a vector-valued memory attenuation policy produced by a bifurcated hierarchical recurrent network operating on whitened innovation sequences. Shallow recurrent states capture instantaneous anomalies while deep states encode sustained trends; an auxiliary reconstruction objective is added to avoid feature collapse. The full filter, including the recursive covariance updates, is trained end-to-end via backpropagation through time to minimize state estimation error. Evaluations are reported on topologically distinct chaotic attractors (claiming cross-domain generalization) and on recorded real-world UAV flight data (claiming improved bridging to dead reckoning and outperformance versus classical adaptive estimators during outages).

Significance. If the stability and generalization claims hold, the work would demonstrate a practical integration of recurrent learning with recursive Bayesian estimation, potentially improving adaptability of Kalman-type filters in regimes where stationary noise assumptions fail. The end-to-end training directly on estimation error and the attempt to evaluate across chaotic systems and real UAV traces are positive features that could influence future hybrid learned-classical estimators.

major comments (1)
  1. [Recursive covariance updates] Recursive covariance updates (methods description of NDR-SHKF): the vector attenuation weights produced by the bifurcated recurrent network are inserted directly into the Sage-Husa recursions for the process and measurement noise covariances Q and R. No projection, clipping, or other explicit constraint is stated to guarantee that the resulting matrices remain positive definite or that their traces remain bounded. Because the central robustness claims concern performance precisely when innovation statistics shift (sensor outages, OOD UAV regimes), the absence of such safeguards is load-bearing; a policy optimized on training attractors can still produce unstable updates under distribution shift.
minor comments (2)
  1. [Abstract] Abstract: performance advantages are asserted on chaotic attractors and real UAV data, yet no numerical metrics, baseline descriptions, or statistical tests are supplied, making the magnitude of improvement difficult to gauge from the summary alone.
  2. Notation: the precise mapping from the network outputs to the per-component attenuation factors applied inside the Sage-Husa update equations should be written explicitly (e.g., as a diagonal matrix or element-wise multiplication) to allow reproduction of the covariance recursion.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the constructive feedback and for recognizing the potential of hybrid recurrent-recursive estimators. We address the single major comment below.

read point-by-point responses

  1. Referee: [Recursive covariance updates] Recursive covariance updates (methods description of NDR-SHKF): the vector attenuation weights produced by the bifurcated recurrent network are inserted directly into the Sage-Husa recursions for the process and measurement noise covariances Q and R. No projection, clipping, or other explicit constraint is stated to guarantee that the resulting matrices remain positive definite or that their traces remain bounded. Because the central robustness claims concern performance precisely when innovation statistics shift (sensor outages, OOD UAV regimes), the absence of such safeguards is load-bearing; a policy optimized on training attractors can still produce unstable updates under distribution shift.

    Authors: We acknowledge that the manuscript does not describe any explicit projection, clipping, or other constraint to enforce positive definiteness or bounded traces on the updated Q and R matrices. The present implementation relies on end-to-end training that directly minimizes state estimation error; empirical results on both topologically distinct chaotic attractors and recorded UAV flights show stable behavior without observed divergence. To strengthen the robustness argument under distribution shift, we will revise the methods section to incorporate a lightweight projection step: after each Sage-Husa update we add a small diagonal loading if any eigenvalue falls below a threshold and clip the matrix trace to a conservative upper bound derived from the maximum expected innovation norm. These safeguards will be stated explicitly and their effect on training will be reported in the revised manuscript. revision: yes

Circularity Check

0 steps flagged

No circularity: learned policy optimized end-to-end against estimation error

full rationale

The paper replaces the scalar forgetting factor with a vector policy produced by a bifurcated recurrent network trained via BPTT to minimize state estimation error on innovation sequences. This is an empirical optimization procedure whose outputs are not equivalent to its inputs by construction; the architecture, auxiliary reconstruction loss, and recursive covariance updates are explicitly defined and differentiated through the training loop. No self-citations, uniqueness theorems, or ansatzes from prior author work are invoked to justify the core mechanism, and the derivation chain consists of standard differentiable filtering equations plus a data-driven training objective. Performance claims rest on empirical generalization across attractors and UAV datasets rather than tautological reduction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The approach rests on the assumption that whitened innovation sequences contain the necessary information for learning an effective memory policy and that end-to-end training of the recursive filter equations is stable and beneficial.

free parameters (1)
  • Recurrent network weights
    Learned end-to-end via backpropagation through time to directly minimize state estimation error rather than through a separate loss on noise statistics.
axioms (1)
  • domain assumption Whitened innovation sequences are suitable inputs for capturing both instantaneous sensor anomalies and sustained dynamic trends
    The architecture description states that the network operates on whitened innovation sequences.
invented entities (1)
  • Bifurcated hierarchical recurrent network for memory attenuation no independent evidence
    purpose: To produce a vector-valued forgetting policy that separates quick and slow dynamics while avoiding feature collapse via auxiliary reconstruction
    New architectural component introduced to replace the static scalar forgetting factor of classical SHKF.

pith-pipeline@v0.9.0 · 5738 in / 1478 out tokens · 63644 ms · 2026-05-20T08:08:19.685554+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

  • IndisputableMonolith/Cost/FunctionalEquation.lean washburn_uniqueness_aczel unclear
    ?
    unclear

    Relation between the paper passage and the cited Recognition theorem.

    replaces this scalar parameter with a vector-valued memory attenuation policy learned by a hierarchical recurrent network... bifurcated architecture routes shallow recurrent states to capture instantaneous sensor anomalies and deep states to encode sustained dynamic trends

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

63 extracted references · 63 canonical work pages · 8 internal anchors

  1. [1]

    M. L. Ireland, D. Anderson, Development of navigation algorithms for nap-of-the-earth UAV flight in a constrained urban environment, 2012

    work page 2012

  2. [2]

    Wahbah, M

    M. Wahbah, M. Chehadeh, M. Hamandi, L. Seneviratne, Y. Zweiri, Real-time adaptive dynamics based state estimation scheme for un- manned aircrafts, IEEE Sensors Journal 22 (14) (2022) 14397–14414. doi:10.1109/JSEN.2022.3183187

    work page doi:10.1109/jsen.2022.3183187 2022

  3. [3]

    B. Yang, E. Yang, H. Shi, L. Yu, C. Niu, Adaptive square-root cubature Kalman filter based low cost UAV positioning in dark and GPS-denied environments, IEEE Transactions on Intelligent Vehicles 10 (5) (2025) 3587–3599. doi:10.1109/TIV.2024.3457678

    work page doi:10.1109/tiv.2024.3457678 2025

  4. [4]

    X. Wang, X. Yan, Q. Luo, Adaptive square-root extended cu- bature Kalman filter based on Huber M-estimation for multi- AUV cooperative navigation, Measurement 249 (2025) 117035. doi:https://doi.org/10.1016/j.measurement.2025.117035

    work page doi:10.1016/j.measurement.2025.117035 2025

  5. [5]

    Hajiyev, H

    C. Hajiyev, H. E. Soken, Robust adaptive Kalman filter for es- timation of UAV dynamics in the presence of sensor/actuator faults, Aerospace Science and Technology 28 (1) (2013) 376–383. doi:https://doi.org/10.1016/j.ast.2012.12.003

    work page doi:10.1016/j.ast.2012.12.003 2013

  6. [6]

    Y.Bar-Shalom, X.R.Li, T.Kirubarajan, StateEstimationforNonlinear Dynamic Systems, John Wiley & Sons, Ltd, 2002, Ch. 10, pp. 371–420. doi:https://doi.org/10.1002/0471221279.ch10

    work page doi:10.1002/0471221279.ch10 2002

  7. [7]

    Foehn, E

    P. Foehn, E. Kaufmann, A. Romero, R. Penicka, S. Sun, L. Bauers- feld, T. Laengle, G. Cioffi, Y. Song, A. Loquercio, D. Scara- muzza, Agilicious: Open-source and open-hardware agile quadrotor for vision-based flight, Science Robotics 7 (67) (2022) eabl6259. arXiv:https://www.science.org/doi/pdf/10.1126/scirobotics.abl6259, doi:10.1126/scirobotics.abl6259. 42

    work page doi:10.1126/scirobotics.abl6259 2022

  8. [8]

    Merchant, S

    E. Kaufmann, L. Bauersfeld, A. Loquercio, M. Müller, V. Koltun, D. Scaramuzza, Champion-level drone racing using deep reinforcement learning, Nature 620 (7976) (2023) 982–987. doi:10.1038/s41586-023- 06419-4. URLhttps://doi.org/10.1038/s41586-023-06419-4

    work page doi:10.1038/s41586-023- 2023

  9. [9]

    Brossard, A

    M. Brossard, A. Barrau, S. Bonnabel, AI-IMU dead-reckoning, IEEE Transactions on Intelligent Vehicles 5 (4) (2020) 585–595. doi:10.1109/TIV.2020.2980758

    work page doi:10.1109/tiv.2020.2980758 2020

  10. [10]

    T. Qin, P. Li, S. Shen, VINS-Mono: A robust and versatile monocular visual-inertial state estimator, IEEE Transactions on Robotics 34 (4) (2018) 1004–1020. doi:10.1109/TRO.2018.2853729

    work page doi:10.1109/tro.2018.2853729 2018

  11. [11]

    Mehra, Approaches to adaptive filtering, IEEE Transactions on Au- tomatic Control 17 (5) (1972) 693–698

    R. Mehra, Approaches to adaptive filtering, IEEE Transactions on Au- tomatic Control 17 (5) (1972) 693–698. doi:10.1109/TAC.1972.1100100

    work page doi:10.1109/tac.1972.1100100 1972

  12. [12]

    Duník, O

    J. Duník, O. Straka, O. Kost, J. Havlík, Noise covariance ma- trices in state-space models: A survey and comparison of estimation methods—part I, International Journal of Adap- tive Control and Signal Processing 31 (11) (2017) 1505–1543. arXiv:https://onlinelibrary.wiley.com/doi/pdf/10.1002/acs.2783, doi:https://doi.org/10.1002/acs.2783

    work page doi:10.1002/acs.2783 2017

  13. [13]

    A. P. Sage, G. W. Husa, Adaptive filtering with unknown prior statistics, 1969, pp. 760–769

    work page 1969

  14. [14]

    Analysis and Modeling of Inertial Sensors Using Allan Variance,

    N. El-Sheimy, H. Hou, X. Niu, Analysis and modeling of inertial sen- sors using allan variance, IEEE Transactions on Instrumentation and Measurement 57 (1) (2008) 140–149. doi:10.1109/TIM.2007.908635

    work page doi:10.1109/tim.2007.908635 2008

  15. [15]

    R. C. Leishman, J. C. Macdonald, R. W. Beard, T. W. McLain, Quadrotors and accelerometers: State estimation with an improved dy- namic model, IEEE Control Systems Magazine 34 (1) (2014) 28–41. doi:10.1109/MCS.2013.2287362

    work page doi:10.1109/mcs.2013.2287362 2014

  16. [16]

    doi:10.1007/s001900050236

    A.H.Mohamed, K.P.Schwarz, AdaptiveKalmanfilteringforINS/GPS, Journal of Geodesy 73 (4) (1999) 193–203. doi:10.1007/s001900050236. URLhttps://doi.org/10.1007/s001900050236 43

    work page doi:10.1007/s001900050236 1999

  17. [17]

    Juston, S

    M. Juston, S. Gupta, S. Mathur, W. R. Norris, D. Nottage, A. Soyle- mezoglu, Robust error state Sage-Husa adaptive Kalman filter for UWB localization, IEEE Sensors Journal 25 (9) (2025) 16034–16049. doi:10.1109/JSEN.2025.3549315

    work page doi:10.1109/jsen.2025.3549315 2025

  18. [18]

    X. Wang, A. Wang, D. Wang, Y. Xiong, B. Liang, Y. Qi, A mod- ified Sage-Husa adaptive Kalman filter for state estimation of elec- tric vehicle servo control system, Energy Reports 8 (2022) 20–27, iCPE 2021 - The 2nd International Conference on Power Engineering. doi:https://doi.org/10.1016/j.egyr.2022.02.105

    work page doi:10.1016/j.egyr.2022.02.105 2022

  19. [19]

    Y. Fan, S. Qiao, G. Wang, H. Zhang, An improved Sage- Husa variational robust adaptive Kalman filter with uncertain noise covariances, IEEE Sensors Journal 24 (18) (2024) 28921–28930. doi:10.1109/JSEN.2024.3421271

    work page doi:10.1109/jsen.2024.3421271 2024

  20. [20]

    Abbeel, A

    P. Abbeel, A. Coates, M. Montemerlo, A. Ng, S. Thrun, Discriminative training of Kalman filters, 2005, pp. 289–296. doi:10.15607/RSS.2005.I.038

    work page doi:10.15607/rss.2005.i.038 2005

  21. [21]

    Q. Xia, M. Rao, Y. Ying, X. Shen, Adaptive fading Kalman filter with an application, Automatica 30 (8) (1994) 1333–1338. doi:https://doi.org/10.1016/0005-1098(94)90112-0

    work page doi:10.1016/0005-1098(94)90112-0 1994

  22. [22]

    Backprop KF: Learning Discriminative Deterministic State Estimators

    T. Haarnoja, A. Ajay, S. Levine, P. Abbeel, Backprop KF: Learning Dis- criminative Deterministic State Estimators (2017). arXiv:1605.07148. URLhttps://arxiv.org/abs/1605.07148

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  23. [23]

    Differentiable Particle Filters: End-to-End Learning with Algorithmic Priors

    R. Jonschkowski, D. Rastogi, O. Brock, Differentiable particle filters: End-to-end learning with algorithmic priors (2018). arXiv:1805.11122. URLhttps://arxiv.org/abs/1805.11122

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  24. [24]

    H. Wen, X. Chen, G. Papagiannis, C. Hu, Y. Li, End-to-end semi- supervised learning for differentiable particle filters, in: 2021 IEEE In- ternational Conference on Robotics and Automation (ICRA), 2021, pp. 5825–5831. doi:10.1109/ICRA48506.2021.9561889

    work page doi:10.1109/icra48506.2021.9561889 2021

  25. [25]

    Cohen, I

    N. Cohen, I. Klein, Inertial navigation meets deep learning: A survey of current trends and future directions, Results in Engineering 24 (2024) 103565. doi:https://doi.org/10.1016/j.rineng.2024.103565. 44

    work page doi:10.1016/j.rineng.2024.103565 2024

  26. [26]

    Shlezinger, G

    N. Shlezinger, G. Revach, A. Ghosh, S. Chatterjee, S. Tang, T. Im- biriba, J. Dunik, O. Straka, P. Closas, Y. C. Eldar, Artificial intelligence-aided kalman filters: AI-augmented designs for Kalman- type algorithms, IEEE Signal Processing Magazine 42 (3) (2025) 52–76. doi:10.1109/MSP.2025.3569395

    work page doi:10.1109/msp.2025.3569395 2025

  27. [27]

    Revach, N

    G. Revach, N. Shlezinger, X. Ni, A. L. Escoriza, R. J. G. van Sloun, Y. C. Eldar, KalmanNet: Neural network aided kalman filtering for partially known dynamics, IEEE Transactions on Signal Processing 70 (2022) 1532–1547. doi:10.1109/TSP.2022.3158588

    work page doi:10.1109/tsp.2022.3158588 2022

  28. [28]

    URL http://dx.doi.org/10.23919/ EUSIPCO63237.2025.11226601

    H. Mortada, C. Falcon, Y. Kahil, M. Clavaud, J.-P. Michel, Recursive KalmanNet: Deep learning-augmented kalman filtering for state esti- mation with consistent uncertainty quantification, in: 2025 33rd Eu- ropean Signal Processing Conference (EUSIPCO), 2025, pp. 885–889. doi:10.23919/EUSIPCO63237.2025.11226444

    work page doi:10.23919/eusipco63237.2025.11226444 2025

  29. [29]

    Zheng, Y

    T. Zheng, Y. Yao, F. He, X. Zhang, An RNN-based learn- able extended Kalman filter design and application, in: 2019 18th European Control Conference (ECC), 2019, pp. 3304–3309. doi:10.23919/ECC.2019.8796088

    work page doi:10.23919/ecc.2019.8796088 2019

  30. [30]

    Z. Song, Z. Zhang, F. Zhang, W. Bao, RG-ESKF: Residual-gated error-state extended Kalman filter for learning-based spacecraft motion state estimation, Aerospace Science and Technology 176 (2026) 112090. doi:https://doi.org/10.1016/j.ast.2026.112090

    work page doi:10.1016/j.ast.2026.112090 2026

  31. [31]

    X. Gao, D. You, S. Katayama, Seam tracking monitoring based on adap- tive Kalman filter embedded Elman neural network during high-power fiber laser welding, IEEE Transactions on Industrial Electronics 59 (11) (2012) 4315–4325. doi:10.1109/TIE.2012.2193854

    work page doi:10.1109/tie.2012.2193854 2012

  32. [32]

    E. Kim, S. Kim, S. Jung, Battery SOC and SOH estima- tion based on Sage-Husa extended Kalman filter and feedfor- ward neural network, IEEE Access 13 (2025) 174044–174056. doi:10.1109/ACCESS.2025.3617967

    work page doi:10.1109/access.2025.3617967 2025

  33. [33]

    X. Meng, H. Tan, P. Yan, Q. Zheng, G. Chen, J. Jiang, A GNSS/INS integrated navigation compensation method based on CNN–GRU + IRAKF hybrid model during GNSS outages, IEEE 45 Transactions on Instrumentation and Measurement 73 (2024) 1–15. doi:10.1109/TIM.2024.3369131

    work page doi:10.1109/tim.2024.3369131 2024

  34. [34]

    J. Li, Z. Zhang, H. Zhao, H. Li, C. Zhai, ROV localization in the nuclear reactor pressure vessel using LSTM and an im- proved adaptive Kalman filter, Measurement 242 (2025) 116206. doi:https://doi.org/10.1016/j.measurement.2024.116206

    work page doi:10.1016/j.measurement.2024.116206 2025

  35. [35]

    S. M. Hosseini, M. Jalili, An online incremental learning support vector regression for INS/GPS integrated navigation system during long-time gps outage, GPS Solutions 29 (4) (2025) 164. doi:10.1007/s10291-025- 01900-1

    work page doi:10.1007/s10291-025- 2025

  36. [36]

    URL https: //doi.org/10.1109/TPAMI.2021.3079209

    T. Hospedales, A. Antoniou, P. Micaelli, A. Storkey, Meta- learning in neural networks: A survey, IEEE Transactions on Pat- tern Analysis and Machine Intelligence 44 (9) (2022) 5149–5169. doi:10.1109/TPAMI.2021.3079209

    work page doi:10.1109/tpami.2021.3079209 2022

  37. [37]

    C. Finn, P. Abbeel, S. Levine, Model-agnostic meta-learning for fast adaptation of deep networks (2017). arXiv:1703.03400. URLhttps://arxiv.org/abs/1703.03400

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  38. [38]

    K. Cho, B. van Merrienboer, C. Gulcehre, D. Bahdanau, F. Bougares, H. Schwenk, Y. Bengio, Learning phrase representations using RNN encoder-decoder for statistical machine translation (2014). arXiv:1406.1078. URLhttps://arxiv.org/abs/1406.1078

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  39. [39]

    J. X. Wang, Z. Kurth-Nelson, D. Tirumala, H. Soyer, J. Z. Leibo, R. Munos, C. Blundell, D. Kumaran, M. Botvinick, Learning to re- inforcement learn (2017). arXiv:1611.05763. URLhttps://arxiv.org/abs/1611.05763

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  40. [40]

    Z. Chen, C. Heckman, S. Julier, N. Ahmed, Weak in the NEES?: Auto- tuning Kalman filters with bayesian optimization, in: 2018 21st Inter- national Conference on Information Fusion (FUSION), 2018, pp. 1072–

    work page 2018

  41. [41]

    doi:10.23919/ICIF.2018.8454982

    work page doi:10.23919/icif.2018.8454982 2018

  42. [42]

    Werbos, Backpropagation through time: what it does and how to do it, Proceedings of the IEEE 78 (10) (1990) 1550–1560

    P. Werbos, Backpropagation through time: what it does and how to do it, Proceedings of the IEEE 78 (10) (1990) 1550–1560. doi:10.1109/5.58337. 46

    work page doi:10.1109/5.58337 1990

  43. [43]

    Differentiation of the Cholesky decomposition

    I. Murray, Differentiation of the Cholesky decomposition (2016). arXiv:1602.07527. URLhttps://arxiv.org/abs/1602.07527

    work page internal anchor Pith review Pith/arXiv arXiv 2016

  44. [44]

    Cheng, H

    J. Cheng, H. Chen, Z. Xue, Y. Huang, Y. Zhang, An online ex- ploratory maximum likelihood estimation approach to adaptive Kalman filtering, IEEE/CAA Journal of Automatica Sinica PP (2025) 1–27. doi:10.1109/JAS.2024.125001

    work page doi:10.1109/jas.2024.125001 2025

  45. [45]

    E. N. Lorenz, Deterministic nonperiodic flow, Journal of the Atmo- spheric Sciences 20 (2) (1963) 130–141

    work page 1963

  46. [46]

    Strogatz, Nonlinear Dynamics and Chaos: With Applications to Physics, Biology, Chemistry, and Engineering, CRC Press, 2018

    S. Strogatz, Nonlinear Dynamics and Chaos: With Applications to Physics, Biology, Chemistry, and Engineering, CRC Press, 2018. URLhttps://books.google.pl/books?id=A0paDwAAQBAJ

    work page 2018

  47. [47]

    Glorot, A

    X. Glorot, A. Bordes, Y. Bengio, Deep sparse rectifier neural networks, in: International Conference on Artificial Intelligence and Statistics, JMLR Workshop and Conference Proceedings, 2011, pp. 315–323

    work page 2011

  48. [48]

    D. P. Kingma, J. Ba, Adam: A method for stochastic optimization (2017). arXiv:1412.6980. URLhttps://arxiv.org/abs/1412.6980

    work page internal anchor Pith review Pith/arXiv arXiv 2017

  49. [49]

    Rössler, An equation for continuous chaos, Physics Letters A 57 (5) (1976) 397–398

    O. Rössler, An equation for continuous chaos, Physics Letters A 57 (5) (1976) 397–398. doi:https://doi.org/10.1016/0375-9601(76)90101-8

    work page doi:10.1016/0375-9601(76)90101-8 1976

  50. [50]

    Huang, Y

    Y. Huang, Y. Zhang, Z. Wu, N. Li, J. Chambers, A novel adaptive Kalman filter with inaccurate process and measurement noise covariance matrices, IEEE Transactions on Automatic Control 63 (2) (2018) 594–

    work page 2018

  51. [51]

    doi:10.1109/TAC.2017.2730480

    work page doi:10.1109/tac.2017.2730480 2017

  52. [52]

    NaturalPoint, Inc., OptiTrack motion capture systems, https://optitrack.com, accessed: 2026-04-15 (2025)

    work page 2026

  53. [53]

    URLhttps://www.quanser.com/products/qdrone/ 47

    Quanser Inc., Physical System Parameters Drone Parametrization v0.4, accessed: 2026-04-10 (2024). URLhttps://www.quanser.com/products/qdrone/ 47

    work page 2026

  54. [54]

    A. Yu, I. Kolotylo, H. A. Hashim, A. E. E. Eltoukhy, Electronic warfare cyberattacks, countermeasures, and modern defensive strate- gies of uav avionics: A survey, IEEE Access 13 (2025) 68660–68681. doi:10.1109/ACCESS.2025.3561068

    work page doi:10.1109/access.2025.3561068 2025

  55. [55]

    F. R. Matzke, T. J. Barot, S. Sridharan, N. Ammann, C. Kessler, H.- G. Maas, F. H. P. Fitzek, A. Eltner, Addressing GNSS vulnerabilities in AAM: A multi-modal UAV testbed for redundant and reliable nav- igation, The International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences XLVIII-2/W11-2025 (2025) 211–218. doi:10.5194/i...

    work page doi:10.5194/isprs-archives-xlviii-2-w11-2025-211-2025 2025

  56. [56]

    H. A. Hashim, Advances in UAV avionics systems architec- ture, classification and integration: A comprehensive review and future perspectives, Results in Engineering 25 (2025) 103786. doi:https://doi.org/10.1016/j.rineng.2024.103786

    work page doi:10.1016/j.rineng.2024.103786 2025

  57. [57]

    Silveria, K

    D. Silveria, K. Cabral, P. Jardine, S. Givigi, Lie group control architec- tures for UAVs: a comparison of se2(3)-based approaches in simulation and hardware (2025). arXiv:2511.15023. URLhttps://arxiv.org/abs/2511.15023

    work page arXiv 2025

  58. [58]

    P. J. Huber, Robust estimation of a location parameter, Springer, 1992, pp. 492–518. doi:10.1007/978-1-4612-4380-9_35

    work page doi:10.1007/978-1-4612-4380-9_35 1992

  59. [59]

    M. Roth, E. Özkan, F. Gustafsson, A student’s t filter for heavy tailed process and measurement noise, in: 2013 IEEE International Confer- ence on Acoustics, Speech and Signal Processing, 2013, pp. 5770–5774. doi:10.1109/ICASSP.2013.6638770

    work page doi:10.1109/icassp.2013.6638770 2013

  60. [60]

    Strapdown Inertial Navigation Technology,

    D. Titterton, J. Weston, Strapdown Inertial Navigation Technology, 2nd Edition, The Institution of Engineering and Technology, 2004. arXiv:https://digital-library.theiet.org/doi/pdf/10.1049/PBRA017E, doi:10.1049/PBRA017E

    work page doi:10.1049/pbra017e 2004

  61. [61]

    O. J. Woodman, An introduction to inertial navigation, Tech. rep., University of Cambridge, Computer Laboratory (2007). URLhttps://www.cl.cam.ac.uk/techreports/UCAM-CL-TR-696.pdf 48

    work page 2007

  62. [62]

    L. Lai, N. Suda, V. Chandra, CMSIS-NN: Efficient neural network ker- nels for arm cortex-M CPUs (2018). arXiv:1801.06601. URLhttps://arxiv.org/abs/1801.06601

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  63. [63]

    Stone, M

    A. Stone, M. Petersen, C. Peterson, Real-time thermal-inertial odom- etry on embedded hardware for high-speed GPS-denied flight (2026). arXiv:2603.02114. URLhttps://arxiv.org/abs/2603.02114 49

    work page arXiv 2026