Pith. sign in

REVIEW 2 major objections 4 minor 39 references

Physics-Regularized Machine Learning for Proprioceptive Vehicle Localization Using Onboard Sensors

T0 review · 2 major / 4 minor · reviewed 2026-07-11 · grok-4.5

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

Pith's one-line read Training a machine-learning model through a differentiable Kalman filter regularizes it with vehicle physics, cutting onboard-sensor dead-reckoning error and improving low-friction generalization.

desk verdict Solid hybrid onboard-sensor odometry with a real low-µ dataset and clear gains; the “physics regularizer” story is useful but rests on a simple kinematic prior that is only lightly stress-tested under slip. read the letter →

arxiv 2607.05663 v1 pith:WBO4TTG3 submitted 2026-07-06 cs.RO cs.AIcs.LG

Abinav Kalyanasundaram , Karthikeyan Chandra Sekaran , Wolfgang Utschick , Michael Botsch This is my paper · ORCID
classification cs.ROcs.AIcs.LG
keywords proprioceptivelocalizationdifferentiableKalmanfilterphysics-informedmachinelearningonboardsensorsvehicleodometrylow-frictiondatasetdeadreckoningtransformer
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

When satellite signals drop out, vehicles must keep track of where they are using only the sensors already built into production cars—wheel speeds, steering angle, yaw rate, and the like. Pure learning models can map those signals to motion estimates but often produce physically inconsistent trajectories that drift or fail on new surfaces. This paper shows that back-propagating pose error through a differentiable extended Kalman filter forces the learned measurement model to stay consistent with a kinematic motion model, acting as a physics regularizer. The resulting hybrid system, PRML2, achieves lower 60-second position error than prior onboard-sensor methods and generalizes better to snowy low-friction conditions while still running in real time. A new public low-friction dataset is released so others can test the same claim.

What carries the argument

PRML2: a transformer that maps a short window of onboard sensor readings to vehicle dynamic states (velocities, accelerations, yaw rate) together with their uncertainties, followed by a fully differentiable extended Kalman filter whose constant-acceleration process model and adaptive measurement covariance serve both as the pose estimator and as the physics regularizer during training.

What would settle it

Retrain PRML2 and the ablated pretraining-only variant on the same dry-road data, then measure 60 s dead-reckoning RMSE on a held-out low-friction set; if the physics-regularized version no longer outperforms the ablated version (or any strong non-physics baseline), the claimed regularization benefit under low friction is falsified.

Watch

Extended reading notes

Core claim

End-to-end training of a transformer-based measurement model through a differentiable extended Kalman filter turns the filter into a physics regularizer: the loss on posterior pose estimates penalizes physical and temporal inconsistencies in the learned vehicle-dynamic states. On a public road dataset this yields 1.5 m RMSE position error over 60 s dead-reckoning intervals (versus 2.1 m for the strongest prior baseline) and, without any low-friction training data, 2.5 m RMSE on an unseen snow dataset, while remaining real-time capable on an automotive edge device.

Load-bearing premise

The simple constant-acceleration planar motion model used inside the Kalman filter still supplies useful physical inductive bias even when tires are sliding on snow, rather than actively distorting the learned measurements.

Editorial extensions

If this is right

  • Production vehicles can maintain usable short-term localization during GNSS outages using only the sensors already present, without adding IMUs or cameras.
  • Physics-regularized end-to-end training improves generalization to out-of-distribution surfaces such as snow without requiring low-friction training data.
  • The same differentiable-filter regularizer can be reused for other proprioceptive platforms once a suitable kinematic process model is supplied.
  • A new public low-friction dataset becomes available for benchmarking future onboard-odometry methods.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • The same end-to-end filter regularization could be applied to IMU-only or wheel-IMU hybrid pipelines to reduce their quadratic drift growth.
  • If the planar assumption is relaxed to a full 3-D process model, the method may extend to vehicles with significant roll or pitch without changing the learning architecture.
  • Online fine-tuning whenever GNSS is briefly available could keep the measurement model adapted to changing tire wear or vehicle load.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

2 major / 4 minor

Summary. The paper proposes PRML2, a hybrid architecture that maps standard onboard vehicle sensors (wheel speeds, steering angle, yaw rate, brake pressure, lateral acceleration) to vehicle dynamic states via a transformer, then fuses those states as adaptive measurements inside a differentiable EKF whose process model is a constant-acceleration kinematic model. End-to-end training back-propagates a pose loss through the EKF, which the authors interpret as a physics regularizer that improves both 60 s dead-reckoning accuracy and generalization. Results on the public ReV-StED test set and a newly collected low-friction snow dataset show lower position RMSE than five baselines (including prior differentiable-filter and RNN-EKF methods), an ablation isolating uncertainty estimation and the EKF regularizer, and real-time inference (~30 Hz) on an NVIDIA Orin. A novel low-µ dataset and code are released.

Significance. If the claimed gains hold under broader validation, the work supplies a practical, low-cost proprioceptive localization pipeline that does not require IMUs or exteroceptive sensors and remains real-time capable on automotive edge hardware. The public release of both code and a challenging low-friction dataset is a concrete community contribution. The formal loss decomposition that frames the differentiable EKF as a physics regularizer is a useful conceptual clarification, even if the numerical advantage ultimately rests on empirical tables rather than a parameter-free derivation. The combination of transformer measurement model, physics-guard bounds, and adaptive EKF is a solid engineering advance for GNSS-denied short-term odometry.

major comments (2)
  1. [§V-B.3, Eq. (9), Table II] The central generalization claim (Table II, snow/low-µ column; abstract and §V-B.3) attributes the OOD improvement (2.5 m RMSE vs 3.4–3.8 m for baselines) to physics regularization by the constant-acceleration kinematic model of Eq. (9). Under high-slip low-µ regimes that model is known to be incomplete (sideslip, tire-force saturation). The manuscript supplies neither residual statistics of Eq. (9) on the snow trajectories nor an ablation that replaces the process model with a weaker (e.g., constant-velocity) or stronger (e.g., bicycle with friction) alternative. Without that evidence the observed gain could equally be explained by end-to-end pose supervision or temporal smoothing rather than genuine physics-informed inductive bias. This link is load-bearing for the paper’s strongest claim and should be tested or the claim tempered.
  2. [§III-C, Eqs. (14)–(16)] The loss decomposition in Eqs. (14)–(15) is presented as theoretical validation that the Kalman filter acts as a physics regularizer. The algebra is an exact rewriting of the standard EKF residual; it does not by itself guarantee that the learned measurements remain on the physically consistent manifold when the process model is misspecified. The empirical support is the PRML2 vs PRML2* comparison, yet that comparison confounds the regularizer with the additional pose-supervision term in Eq. (16). A cleaner isolation (identical pose loss with and without the recursive EKF path) would strengthen the theoretical claim.
minor comments (4)
  1. [Fig. 2] Fig. 2 caption and body text contain several typos (“trasformer”, “a posteriorivehicle”, “Correvit”). A careful proof-read is needed.
  2. [§V-C, Table III] The physics-guard activation rate is reported as <2 %; the corresponding ablation row in Table III therefore shows almost no effect. It would be clearer to state that the guard is retained mainly for safety rather than for accuracy.
  3. [§IV-B] Baselines DL-AVL and RNN-EKF were originally designed for IMU inputs; a short paragraph justifying the fairness of the adaptation to onboard-sensor-only inputs would help readers.
  4. [§V-A.1] The hyper-parameter β = 8 and the 70 : 30 pre-train/fine-tune split are stated without sensitivity analysis. A one-sentence note on robustness to these choices would be useful.

Circularity Check

1 steps flagged · score 1.0 of 10

No load-bearing circularity; the EKF loss rewrite is an algebraic identity, results are measured against external RTK ground truth, and self-citations are non-essential.

  1. self citation load bearing [Sec. IV-A / Ref. [36] and Sec. II / Ref. [5]]
    "the publicly available Real-world Vehicle State Estimation Dataset (ReV-StED) [36] is used to evaluate the proposed method. ... inspired by prior attention-based designs [5]."

    The evaluation dataset and the transformer backbone are taken from the authors’ own prior publications. This is ordinary self-citation and does not force the localization numbers or the regularizer interpretation; the snow set is new and the GT is external RTK, so the circularity is only minor and non-load-bearing.

full rationale

The paper's central claim (end-to-end training of a transformer measurement model through a differentiable EKF yields lower 60 s dead-reckoning error and better OOD generalization on the new low-µ snow set) is an empirical result, not a derivation forced by construction. The loss decomposition in Eqs. (14)–(15) simply expands the standard EKF posterior residual ||x_{t|t} - x_gt||^{2} = ||prior residual + K·innovation||^{2} and labels the cross terms a “physics regularizer”; this is interpretive bookkeeping, not a self-definitional claim that the numerical accuracy is guaranteed by the rewrite. Training still optimizes an explicit position-supervision term (Eq. 16) against external RTK-grade ground-truth trajectories that never enter the model as inputs. The snow evaluation is fully out-of-distribution (models trained only on ReV-StED). Baselines are independent published methods; the only self-references are the authors’ own prior ReV-StED dataset paper and a related sideslip estimator, neither of which supplies a uniqueness theorem or forces the present architecture. No fitted parameter is later re-labeled a prediction, no ansatz is smuggled via citation, and no known empirical pattern is merely renamed. Minor self-citation of the authors’ own dataset is present but not load-bearing, hence score 1 rather than 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a standard kinematic motion model, Gaussian noise assumptions, and a handful of hand-chosen hyperparameters and friction bounds. No new physical entities are postulated; the transformer and differentiable EKF are architectural choices rather than invented ontological objects. Free parameters are the usual ML and filter knobs that any such hybrid system must set.

free parameters (5)
  • loss scaling β = 8
    Set to 8 after a hyperparameter study; controls the relative weight of the position residual inside the tanh-smoothed total loss (Eq. 16).
  • process noise covariance Q
    Hand-tuned on the validation set for all EKF-based methods, including PRML2; never learned.
  • physics-guard limits (a_max, v_max, ω_max) = a_max=10/5, v_max=30, ω_max=1.5
    a_max = 10 m/s^{2} (dry) or 5 m/s^{2} (low-µ), v_max = 30 m/s, ω_max = 1.5 rad/s; chosen from domain knowledge / friction coefficient assumptions rather than data fit.
  • pretrain/finetune epoch ratio and total epochs = 20 epochs, 70:30
    70:30 split of 20 epochs; chosen for reliable convergence rather than exhaustive search.
  • transformer dimensions (L, H, d, layers, heads) = L=250,H=150,d=512,2x8
    L=250, H=150, d=512, 2 layers, 8 heads; architectural choices that affect capacity and therefore final accuracy.
assumptions (4)
  • domain assumption Vehicle motion obeys a constant-acceleration kinematic model whose state transition is given by Eq. (9) with Euler-angle rotation matrix.
    Used both as the EKF process model and as the source of the physics-regularization inductive bias; validity under high-slip low-µ conditions is not proven.
  • domain assumption Onboard-sensor noise and residual model error are zero-mean Gaussian, allowing the NLL loss (Eq. 6) and the EKF update (Eq. 13).
    Standard filtering assumption; real sensor noise can be biased or heavy-tailed.
  • domain assumption Motion is planar (2-D localization); roll/pitch are small enough that Euler angles do not suffer gimbal lock.
    Stated explicitly because height is unavailable from the chosen onboard sensors.
  • standard math Differentiable EKF prediction and Joseph-form update preserve a valid computational graph for back-propagation through time.
    Relies on analytic Jacobians and Cholesky factorization of the innovation covariance; standard numerical linear algebra.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Physics-Regularized Machine Learning for Proprioceptive Vehicle Localization Using Onboard Sensors." pith.science (2026). https://pith.science/paper/WBO4TTG3

@misc{pith2026260705663,
  author       = {Pith},
  title        = {Pith review of: Physics-Regularized Machine Learning for Proprioceptive Vehicle Localization Using Onboard Sensors},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WBO4TTG3}},
  note         = {Machine review of arXiv:2607.05663}
}
read the original abstract

Accurate and robust localization is essential for autonomous mobility systems in real-world environments. While fusing Inertial Measurement Unit (IMU) data with satellite-based correction signals provides precise vehicle pose estimates, performance degrades substantially during outages. Recent studies indicate that Machine Learning (ML) can improve IMU-based proprioceptive localization, highlighting untapped potential for onboard sensors readily available in production vehicles. This paper introduces Physics-Regularized Machine Learning for Localization (PRML2), a hybrid framework that combines the complementary strengths of Kalman filtering and data-driven learning to estimate vehicle pose directly from onboard sensors. A key aspect of PRML2 is its physics-regularized learning, enabled by end-to-end training of an ML model through a differentiable Kalman filter. This improves consistency with vehicle motion models, thereby enhancing both localization accuracy and generalization across driving conditions. We evaluate the performance limits of ML-enhanced onboard odometry on a publicly available dataset and show that PRML2 achieves superior localization accuracy and demonstrates real-time capability. This work also introduces a novel dataset to support vehicle localization research under low-friction conditions. The proposed framework provides a robust and cost-effective solution for vehicle localization under degraded sensing conditions by integrating learning with physics-based priors.

Figures

Figures reproduced from arXiv: 2607.05663 by the authors.

Figure 1
Figure 1. Overview of the test vehicle configuration and the [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The PRML2 architecture contains two core components: A trasformer based ML model and a differentiable Extended [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Vehicle localization performance over 60 s under longitudinal, lateral, and mixed dynamic conditions in the ReV-StED dataset. The Snow (low-µ) indicates generalization performance on the proposed low-friction dataset without training. TABLE II: Horizontal position error (m) and its standard deviation over 60 s dead-reckoning using onboard sensors. ReV-StED Snow (low-µ) Models RMSE MAX RMSE MAX OSD-Baseline 5.3 ± 2.1… view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: Drift evolution during dead-reckoning in ReV-StED. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

39 extracted references · 28 canonical work pages

  1. [1]

    Perception, planning, control, and coordination for autonomous vehicles,

    S. D. Pendleton, H. Andersen, and X. Du, “Perception, planning, control, and coordination for autonomous vehicles,”Machines, 2017

    2017

  2. [2]

    Gnss/ins integration methods,

    A. Angrisano, “Gnss/ins integration methods,” PhD dissertation, Parthenope University of Naples, Naples, Italy, 2010

    2010

  3. [3]

    An improved ukf for imu state estimation based on modulation lstm neural network,

    J. Luo, K. Wu, and Wang, “An improved ukf for imu state estimation based on modulation lstm neural network,”IEEE Transactions on Intelligent Transportation Systems, 2024

    2024

  4. [4]

    A survey on localization for au- tonomous vehicles,

    D. Kumar and N. Muhammad, “A survey on localization for au- tonomous vehicles,”IEEE Access, vol. 11, pp. 115 865–115 883, 2023

    2023

  5. [5]

    Uncertainty- aware hybrid machine learning in virtual sensors for vehicle sideslip angle estimation,

    A. Kalyanasundaram, W. Utschick, and M. Botsch, “Uncertainty- aware hybrid machine learning in virtual sensors for vehicle sideslip angle estimation,” in2025 IEEE Intelligent Vehicles Symposium (IV)

    work page

  6. [6]

    Three-dimensional vehicle dynamics state estimation for high-speed race cars under varying signal quality,

    S. Goblirsch, M. Weinmann, and J. Betz, “Three-dimensional vehicle dynamics state estimation for high-speed race cars under varying signal quality,” in2024 IEEE/RSJ IROS. IEEE, 2024

    2024

  7. [7]

    A survey of vehicle localization: Performance analysis and challenges,

    X. Shan, A. Cabani, and H. Chafouk, “A survey of vehicle localization: Performance analysis and challenges,”IEEE Access, vol. 11, 2023

    2023

  8. [8]

    Deep learning for inertial positioning: A survey,

    C. Chen and X. Pan, “Deep learning for inertial positioning: A survey,” IEEE transactions on intelligent transportation systems, vol. 25, 2024

    2024

Show all 39 references
  1. [9]

    Inertial navigation systems for mobile robots,

    B. Barshan and H. Durrant-Whyte, “Inertial navigation systems for mobile robots,”IEEE Transactions on Robotics and Automation, vol. 11, no. 3, pp. 328–342, 1995

    1995

  2. [10]

    Ai-imu dead-reckoning,

    M. Brossard, A. Barrau, and S. Bonnabel, “Ai-imu dead-reckoning,” IEEE Transactions on Intelligent Vehicles, vol. 5, no. 4, 2020

    2020

  3. [11]

    Imu dead-reckoning localization with rnn-iekf algorithm,

    H. Zhou, Y . Zhaoet al., “Imu dead-reckoning localization with rnn-iekf algorithm,” in2022 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2022

    2022

  4. [12]

    Deep learning-based approach for au- tonomous vehicle localization: application and experimental analysis,

    N. Mark ´o, E. Horv ´athet al., “Deep learning-based approach for au- tonomous vehicle localization: application and experimental analysis,” Machines, vol. 11, no. 12, p. 1079, 2023

    2023

  5. [13]

    A hybrid physics-data driven approach for vehicle dynamics state estimation,

    Q. Li, B. Zhanget al., “A hybrid physics-data driven approach for vehicle dynamics state estimation,”Mechanical Systems and Signal Processing, vol. 225, p. 112249, 2025

    2025

  6. [14]

    Advances in neural information processing systems,

    A. Vaswani, N. Shazeeret al., “Advances in neural information processing systems,” vol. 30. Curran Associates, Inc., 2017

    2017

  7. [15]

    On the identification of noise covariances and adaptive kalman filtering: A new look at a 50 year-old problem,

    L. Zhang, D. Sidotiet al., “On the identification of noise covariances and adaptive kalman filtering: A new look at a 50 year-old problem,” IEEE Access, vol. 8, pp. 59 362–59 388, 2020

    2020

  8. [16]

    Wheel-INS: A wheel-mounted MEMS imu-based dead reckoning system,

    X. Niu, Y . Wu, and J. Kuang, “Wheel-INS: A wheel-mounted MEMS imu-based dead reckoning system,”IEEE Transactions on Vehicular Technology, vol. 70, no. 10, pp. 9814–9825, 2021

    2021

  9. [17]

    Improved vehicle localization using on-board sensors and vehicle lateral velocity,

    L. Gao, L. Xionget al., “Improved vehicle localization using on-board sensors and vehicle lateral velocity,”IEEE Sensors Journal, vol. 22, no. 7, pp. 6818–6831, 2022

    2022

  10. [18]

    Learning wheel odometry and imu errors for localization,

    M. BROSSARD, “Learning wheel odometry and imu errors for localization,” in2019 International Conference on Robotics and Au- tomation (ICRA), 2019, pp. 291–297

    2019

  11. [19]

    Llio: Lightweight learned inertial odometer,

    Y . Wang, J. Kuanget al., “Llio: Lightweight learned inertial odometer,” IEEE Internet of Things Journal, vol. 10, no. 3, pp. 2508–2518, 2023

    2023

  12. [20]

    A2dio: Attention-driven deep inertial odometry for pedestrian localization based on 6d imu,

    Y . Wang, H. Cheng, and M. Q.-H. Meng, “A2dio: Attention-driven deep inertial odometry for pedestrian localization based on 6d imu,” in2022 International Conference on Robotics and Automation (ICRA)

  13. [21]

    A hybrid adaptive velocity aided navigation filter with application to ins/dvl fusion,

    B. Or and I. Klein, “A hybrid adaptive velocity aided navigation filter with application to ins/dvl fusion,” inOCEANS 2022, Hampton Roads, 2022, pp. 1–5

    2022

  14. [22]

    Calib-net: Calibrating the low-cost imu via deep convolutional neural network,

    X. Li, B. Liet al., “Calib-net: Calibrating the low-cost imu via deep convolutional neural network,”Frontiers in Robotics and AI, 2022

    2022

  15. [23]

    An unscented kalman filter-informed neural network for vehicle sideslip angle estimation,

    A. Bertipaglia, M. Alirezaei, and R. Happee, “An unscented kalman filter-informed neural network for vehicle sideslip angle estimation,” IEEE Transactions on Vehicular Technology, vol. 73, no. 9, 2024

    2024

  16. [24]

    End-to-end neural network for vehicle dynamics modeling,

    L. Hermansdorfer and Trauth, “End-to-end neural network for vehicle dynamics modeling,” in2020 6th IEEE Congress on Information Science and Technology (CiSt), 2020

    2020

  17. [25]

    Backprop kf: Learning discriminative deterministic state estimators,

    T. Haarnoja, A. Ajayet al., “Backprop kf: Learning discriminative deterministic state estimators,”Advances in neural information pro- cessing systems, vol. 29, 2016

    2016

  18. [26]

    Particle filter recurrent neural networks,

    X. Ma, P. Karkuset al., “Particle filter recurrent neural networks,” in The Thirty-Fourth AAAI Conference on Artificial Intelligence, 2020

    2020

  19. [27]

    Multimodal sensor fusion with differentiable filters,

    M. A. Lee, B. Yiet al., “Multimodal sensor fusion with differentiable filters,” in2020 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2020, pp. 10 444–10 451

    2020

  20. [28]

    Deep learning for time series clas- sification and extrinsic regression: A current survey,

    N. M. Foumani, L. Milleret al., “Deep learning for time series clas- sification and extrinsic regression: A current survey,”ACM Comput. Surv., 2024

    2024

  21. [29]

    C. M. Bishop,Pattern Recognition and Machine Learning, 2006

    2006

  22. [30]

    Botsch and W

    M. Botsch and W. Utschick,Fahrzeugsicherheit und automatisiertes Fahren. Munchen, Germany: Carl Hanser Verlag GmbH, Jun. 2020

    2020

  23. [31]

    Physics-informed machine learning,

    G. E. Karniadakis, I. G. Kevrekidiset al., “Physics-informed machine learning,”Nature Reviews Physics, vol. 3, no. 6, pp. 422–440, 2021

    2021

  24. [32]

    Kalmannet: Data-driven kalman filtering,

    G. Revach, N. Shlezingeret al., “Kalmannet: Data-driven kalman filtering,” inICASSP 2021 - 2021 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), 2021

    2021

  25. [33]

    An expert’s guide to training physics-informed neural networks (2023),

    S. Wang, S. Sankaranet al., “An expert’s guide to training physics-informed neural networks (2023),”Preprint at https://arxiv. org/pdf/2308.08468. pdf, 2023

    2023 arXiv

  26. [34]

    Self-scalable tanh (stan): Multi-scale solutions for physics-informed neural networks,

    R. Gnanasambandam, B. Shenet al., “Self-scalable tanh (stan): Multi-scale solutions for physics-informed neural networks,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 45, no. 12, pp. 15 588–15 603, 2023

    2023

  27. [35]

    Are we ready for autonomous driving? the kitti vision benchmark suite,

    A. Geiger, P. Lenz, and R. Urtasun, “Are we ready for autonomous driving? the kitti vision benchmark suite,” inConference on Computer Vision and Pattern Recognition (CVPR), 2012

    2012

  28. [36]

    Real-world vehicle state estimation dataset (ReV- StED), IEEE IV 2025,

    A. Kalyanasundaram, “Real-world vehicle state estimation dataset (ReV- StED), IEEE IV 2025,” Apr. 2025, available:https://doi.org/ 10.5281/zenodo.15270060

    2025 doi

  29. [37]

    Adma-g pro+ high-precision gnss/inertial sys- tem,

    Genesys Offenburg, “Adma-g pro+ high-precision gnss/inertial sys- tem,” https://genesys-offenburg.de/adma-g/, accessed: 2025-Nov

    2025

  30. [38]

    Fault-diagnosis systems. an introduction from fault detection to fault tolerance,

    R. Isermann, “Fault-diagnosis systems. an introduction from fault detection to fault tolerance,” Jul. 2006, place: Germany

    2006

  31. [39]

    Nvidia jetson orin developer kit,

    “Nvidia jetson orin developer kit,” 2023, available:https://developer. nvidia.com/embedded/jetson-orin

    2023

Pith tools

Reviewed July 11, 2026 · model on record in the stance chip above.