Pith. sign in

REVIEW 5 major objections 6 minor 36 references

An Inertial Sequence Learning Framework for Vehicle Speed Estimation via Smartphone IMU

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read DVSE claims that splitting the inertial speed-estimation problem into a noise-compensation network and a motion-transformation network, trained with random-rotation augmentation and min-alignment loss, gives the best reported…

desk verdict Useful incremental systems paper; the architecture and ablations are sound, but the core accuracy claim over DeepTrack is held up by a small margin on a private GNSS-supervised dataset. read the letter →

arxiv 2505.18490 v1 pith:WWXZTYQW submitted 2025-05-24 cs.RO

Xuan Xiao , Xiaotong Ren , Haitao Li This is my paper · ORCID
classification cs.RO
keywords vehiclespeedestimationsmartphoneIMUinertialsequencelearningnoisecompensationnetworkmotiontransformationrandomrotationaugmentationGNSSsupervisiontimestampalignmentloss
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

The paper proposes that a smartphone can estimate a car's forward speed from its built-in IMU alone, supervised by GNSS, using a split sequence-learning architecture called DVSE. The central idea is to separate the problem into two learnable parts: a network that compensates sensor noise and bias, and a network that aligns the phone's coordinate frame with the vehicle's. On a crowdsourced dataset of about 200 hours from 300 drivers, the authors report that DVSE reaches a 60-second velocity MAE of 2.35 m/s and distance MAE of 50.84 m, improving on the previous DeepTrack baseline without needing any initialization. The claim matters because GNSS degrades in tunnels, parking structures, and dense city canyons, exactly where phone-only speed estimation is needed.

What carries the argument

The split architecture is the load-bearing mechanism: a noise compensation network (a GRU with embedding and regression layers) fits the aggregate disturbance term $N$ in the forward-velocity propagation equation, and a motion transformation network (a dilated causal TCN) outputs the three Euler angles that rotate phone-frame acceleration into the vehicle frame. A random-rotation data augmentation rotates each input window by a randomly generated rotation matrix to mimic different phone poses, and a min-alignment SmoothL1 loss compares the model output with either the time-aligned GNSS target or the target shifted by one second, whichever gives the smaller error. This separation lets each network learn a simpler function, and the min-loss absorbs GNSS latency without explicit time-delay calibration.

What would settle it

Collect a test drive through a tunnel or urban canyon while recording an independent wheel-speed or OBD-II reference, run DVSE on the same phone IMU, and compare. If DVSE's velocity MAE over 60 seconds is close to its reported 2.35 m/s, the claim stands; if error grows markedly relative to a GNSS comparison on the same drive, the assumption that GNSS supervision is adequate is violated.

Watch

Extended reading notes

Core claim

DVSE's central claim is that explicitly decomposing the inertial speed integration of Eq. (5) into a learned noise term $N$ and a learned phone-to-vehicle rotation $R_v^p$ makes supervised speed estimation both more accurate and more generalizable than learning the whole pipeline end-to-end. A GRU-based noise compensation network estimates the disturbance term from one-second feature summaries of accelerometer and gyroscope data plus a reference velocity, while a TCN-based motion transformation network estimates Euler angles from pre-integrated acceleration and gravity. Random rotation of input windows simulates arbitrary phone placements, and a loss that takes the minimum between aligned and one-second-shifted target comparisons absorbs GNSS timestamp delay. The paper reports that this combination achieves its best results without initialization parameters, and ablation experiments show each module contributes, with data augmentation giving the largest single accuracy gain.

Load-bearing premise

The load-bearing premise is that GNSS-reported speed is accurate enough to serve as both the training target and the evaluation ground truth, despite the paper's motivating scenarios being exactly those where GNSS is degraded or blocked.

Editorial extensions

If this is right

  • DVSE can be applied with no initialization or per-driver setup, unlike PCA-based or magnetometer-based baselines.
  • Models trained with trajectory-level splits degrade less on unseen phones and poses: DVSE's 60-second velocity MAE rose to 2.26 m/s versus DeepTrack's 4.15 m/s.
  • Each module is removable, and ablation shows removing data augmentation hurts most, establishing pose variation as a key generalization bottleneck.
  • The loss-matching mechanism tolerates GNSS delays up to about one second, reducing the need for precise timestamp filtering in training data.
  • The chosen GRU noise network runs on commodity smartphones with modest memory and inference time, supporting on-device use.

Reading between the lines

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

  • The paper leaves implicit that the same split of pose estimation and noise compensation could transfer to other vehicle-mounted inertial tasks, such as lateral speed or attitude estimation, since Eq. (5) is specific to forward speed.
  • An untested but natural extension is to evaluate on GNSS-denied segments such as tunnels and parking garages against wheel-speed or OBD ground truth; the reported GNSS-supervised MAE may understate true error where GNSS itself is inaccurate.
  • Because data augmentation gave the largest gain, this suggests that pose diversity, not hardware diversity, is the dominant generalization bottleneck in this crowdsourced regime; whether the same holds for a wider device distribution remains open.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

5 major / 6 minor

Summary. The paper proposes DVSE, a deep-learning framework for estimating vehicle speed from smartphone IMU data using GNSS-derived speed as supervision. The framework decomposes the task into a GRU-based noise compensation network and a TCN-based motion transformation network, adds random-rotation data augmentation to simulate arbitrary phone poses, and uses a min-of-two loss-matching scheme to tolerate timestamp misalignment between IMU and GNSS. The method is evaluated on a private 200-hour crowdsourced dataset under trajectory-split and traditional-split protocols, compared against AI-IMU, VeTorch, DeepVIP, and DeepTrack, and ablated module by module. The authors report that DVSE achieves the best accuracy without initialization, with a 60-second velocity MAE of 2.35 m/s versus 2.51 m/s for DeepTrack, and also report deployment results on Android phones via ONNX Runtime.

Significance. If the reported results are reliable, the paper makes a useful engineering contribution: the modular separation of pose estimation and noise compensation is well motivated by the motion model in Eqs. (1)-(5), the ablation in Table III shows that every component contributes, and the deployment measurements in Table II and Fig. 9 demonstrate a practical efficiency advantage. The random-rotation augmentation and the sliding loss-matching idea are simple and potentially transferable to other IMU-learning tasks. However, the central claims rest on a private dataset, single-run evaluations without error bars or significance tests, and GNSS-derived speed as both training target and evaluation ground truth. The margins over the strongest baseline are small, and some reported numbers are internally inconsistent. These issues must be resolved before the accuracy and generalization claims can be accepted.

major comments (5)
  1. [Sec. III-A, Sec. VII-C] The entire supervision and evaluation use GNSS-derived speed as ground truth, but the paper motivates GNSS-degraded environments (tunnels, urban canyons, underground parking) and never quantifies the GNSS error in the 200-hour crowdsourced dataset. No filtering by GNSS quality indicators (e.g., number of satellites, dilution of precision, C/N0) is reported, and no independent reference such as OBD-II/CAN, RTK, or wheel-speed data is used. As a result, the reported MAEs may reflect fidelity to GNSS artifacts (delays, multipath, interpolation) rather than true vehicle speed, especially since the 60-second advantage over DeepTrack is only 0.16 m/s and the 80th-percentile velocity errors are essentially tied (3.62 vs. 3.63 m/s). Please add a validation subset with independent ground truth, or at minimum report GNSS error statistics and quality-filtered results, and show that the main conclusions are stable under plausible GNSS noise.
  2. [Table I, Table III, Fig. 8] No error bars, confidence intervals, or significance tests are reported for any comparison. The central claim of 'highest accuracy' rests on a 0.16 m/s margin in 60-second velocity MAE, which is smaller than typical seed-to-seed variation for deep inertial models, and the 80th-percentile errors are practically identical. Please report means and variances over multiple training runs (at least five seeds) and perform paired significance tests for the comparisons in Table I, Table III, and Fig. 8. Without this, the claimed superiority over DeepTrack is not statistically established.
  3. [Sec. VII-D2, Table I] The baseline evaluation protocol is not controlled enough to support the accuracy comparison. AI-IMU results are reported after 'removing the segments that converge during evaluation,' but the number and nature of removed segments are not given, which creates a risk of cherry-picking. VeTorch and DeepTrack are initialized using the first 100 seconds of each test trajectory and use GNSS to identify stationary and forward segments, so Table I does not isolate the claimed advantage 'without the need for initialization parameters'; that advantage is instead supported by Fig. 8, whose numbers do not match Table I. Please specify the exact protocol for every baseline, report how many segments are removed and why, and provide a no-initialization variant for each baseline.
  4. [Sec. VII-D2, Fig. 8, Table I] The reported results are internally inconsistent. Fig. 8 reports 60-second trajectory-split DVSE errors of 2.26 m/s (velocity) and 49.75 m (distance) and DeepTrack errors of 4.15 m/s and 93.44 m, while Table I reports 60-second values of 2.35/50.84 for DVSE and 2.51/58.56 for DeepTrack under what appears to be the same trajectory-split setting. If the settings differ (e.g., with vs. without initialization, different random splits, or different input configurations), the difference must be stated explicitly; otherwise the reader cannot determine which numbers support the paper's central claim.
  5. [Sec. IV-C, Sec. VI-B] The reference velocity input Vr creates a train/inference mismatch that is not analyzed. During training Vr is the first-frame GNSS velocity, while during inference it is the model's own previous estimate, so estimation errors can compound over the 30-60 second evaluation horizons. No experiment varies the quality of Vr or measures recursive error growth. Please add an analysis (e.g., oracle Vr versus estimated Vr, and error trajectories over longer horizons) to show that the reported accuracy does not depend on this teacher-forcing shortcut.
minor comments (6)
  1. [Throughout] There are numerous typographical and grammatical errors that should be corrected in a copyedit: 'introduct', 'drving', 'tenporal', 'varients', 'conducte', 'incorrecttion', 'donate', 'mehthod', and 'Receptive F ield' in Eq. (7).
  2. [Sec. VI-A, Eq. (13)] The loss weight lambda is set to 0.7 without any sensitivity analysis; please report how the results vary with lambda in a small range around this value.
  3. [Sec. V-C] The data augmentation description is underspecified: the sentence 'it is important to use the same rotation matrix for each batch of data' is ambiguous, and the distribution/range of the random three-axis Euler angles (alpha_r, beta_r, gamma_r) is never given, even though it is a free hyperparameter of the method.
  4. [Sec. VII-D2] The phrase 'we remove the segments that converge during evaluation' for AI-IMU is confusing; if the intended meaning is 'segments that do not converge', it should be stated that way, along with the fraction of data removed.
  5. [Sec. VII-B2] The hardware is described as 'GeForce GTX 2080 Ti'; if the GPU is an RTX 2080 Ti, the name should be corrected. Also, reference [34] is cited for cosine annealing, but the cited paper is about decoupled weight decay; this reference should be replaced or corrected.
  6. [Table II] The column header 'RAMMAE' appears to be a formatting error; the RAM column should be separated from the MAE columns, and RAM units should be stated consistently (the text mentions KB and the table appears to list KB).

Circularity Check

0 steps flagged · score 2.0 of 10

No circular derivation; DVSE's predictions are learned on held-out IMU data, and self-cited baselines are empirical comparisons, not load-bearing premises.

full rationale

The paper's derivation chain is a standard supervised learning pipeline: Eq. (3)-(5) decompose velocity change into an integration term and a noise term N, a GRU network is trained to fit N from IMU features and a reference velocity, and a motion-transformation network estimates the phone-to-vehicle pose; the whole model is trained end-to-end against GNSS-derived velocity changes. Training and evaluation use disjoint trajectory splits (Sec. VII-B1), so the reported errors are not fits of the test targets. The teacher-forcing reference velocity Vr uses GNSS only at the first frame of a training window and switches to the model's own estimate during inference (Sec. IV-C), so the predicted velocity is not copied from the input by construction. The min-alignment loss of Eq. (14) selects between two temporal alignments to the GNSS target; it does not define the target as the model output. References [4] and [6] are prior works by the same group and are used as baselines and as motivation for the split design; they do not supply an unverified uniqueness or ansatz premise on which the derivation rests, and the comparison in Table I is empirical rather than definitional. The main concerns--GNSS error not quantified, possible unfairness in baseline initialization, and the use of GNSS as both supervision and evaluation ground truth--are correctness/validity issues, not circularity. Thus no circular step is identified; the minor self-citation of prior group work is not load-bearing.

Assumptions & free parameters 3 free parameters · 6 assumptions · 0 invented entities

The framework rests on GNSS as a trusted reference, a rigid phone-vehicle coupling, a bounded timestamp delay, an unvalidated augmentation distribution, and an interpretation of the GRU residual as noise. None of these is demonstrated independently in the paper.

free parameters (3)
  • lambda (loss weight) = 0.7
    Hand-chosen weight balancing the delta-velocity loss and the integrated-velocity loss in Eq. 13. No sensitivity analysis or search procedure is reported.
  • Random rotation augmentation range = not specified
    Section V-C describes random Euler angles alpha_r, beta_r, gamma_r but does not give their distribution or magnitude, even though the ablation shows this module yields the largest accuracy gain.
  • Input window length = 10 seconds
    The model uses a 10-second input window and 1-second feature and pre-integration intervals; no ablation on these design choices is reported.
assumptions (6)
  • domain assumption GNSS-derived speed is accurate enough to act as training target and evaluation ground truth.
    Invoked in Sec. III-A ('their accuracy is generally sufficient for practical applications') and in all loss and metric computations (Sec. VI and VII-C).
  • domain assumption The smartphone is relatively fixed to the vehicle during each estimation window.
    Sec. III-A states that when the phone is relatively fixed, its motion can capture the vehicle's motion; the pose network estimates a single rotation per window.
  • domain assumption The GNSS-IMU timestamp mismatch is at most one second.
    Sec. VI-B limits the sliding matching to two alignments because 'based on observations, this delay usually is at most one second when the GNSS signal is strong'.
  • ad hoc to paper Randomly rotating IMU windows simulates valid smartphone poses without corrupting vehicle motion semantics.
    Sec. V-C introduces the augmentation but provides no validation that the augmented pose distribution covers real-world phone placements.
  • ad hoc to paper The learned residual N represents sensor noise rather than absorbing arbitrary modeling error.
    Sec. IV claims the GRU fits the noise term N of Eq. 5, but N is a flexible learned function with no independent noise measurement.
  • domain assumption One-second pre-integration of acceleration retains enough information for pose estimation.
    Sec. V-B1 integrates accelerometer data over each second to shorten sequences; the paper does not study the information loss from this downsampling.

how reviews work

0 comments
Cite this review

Pith. "Pith review of An Inertial Sequence Learning Framework for Vehicle Speed Estimation via Smartphone IMU." pith.science (2026). https://pith.science/paper/WWXZTYQW

@misc{pith2026250518490,
  author       = {Pith},
  title        = {Pith review of: An Inertial Sequence Learning Framework for Vehicle Speed Estimation via Smartphone IMU},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/WWXZTYQW}},
  note         = {Machine review of arXiv:2505.18490}
}
read the original abstract

Accurately estimating vehicle velocity via smartphone is critical for mobile navigation and transportation. This paper introduces a cutting-edge framework for velocity estimation that incorporates temporal learning models, utilizing Inertial Measurement Unit (IMU) data and is supervised by Global Navigation Satellite System (GNSS) information. The framework employs a noise compensation network to fit the noise distribution between sensor measurements and actual motion, and a pose estimation network to align the coordinate systems of the phone and the vehicle. To enhance the model's generalizability, a data augmentation technique that mimics various phone placements within the car is proposed. Moreover, a new loss function is designed to mitigate timestamp mismatches between GNSS and IMU signals, effectively aligning the signals and improving the velocity estimation accuracy. Finally, we implement a highly efficient prototype and conduct extensive experiments on a real-world crowdsourcing dataset, resulting in superior accuracy and efficiency.

Figures

Figures reproduced from arXiv: 2505.18490 by the authors.

Figure 1
Figure 1. Left: Driving with the assistance of navigation application. Right: [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Coordinate system transformation from the phone to the vehicle. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The architecture of our model. The model consists three parts: [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Model architecture of the noise compensation block. 2x means that [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 5
Figure 5. Figure 5: The architecture of the motion transformation network. The TCN [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]
Figure 6
Figure 6. Figure 6: An example of the delay of GNSS. The IMU ( [PITH_FULL_IMAGE:figures/full_fig_p006_6.png]
Figure 7
Figure 7. Figure 7: Two alignments for loss matching. We utilize a sliding window to achieve a more accurate match between the output and the target. As mentioned before, [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 8
Figure 8. Figure 8: Comparison of DeepTrack and DVSE to different dataset split method. [PITH_FULL_IMAGE:figures/full_fig_p008_8.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

36 extracted references · 27 canonical work pages

  1. [1]

    A comprehensive review of gnss/ins integration techniques for land and air vehicle applications,

    N. Boguspayev, D. Akhmedov, A. Raskaliyev, A. Kim, and A. Sukhenko, “A comprehensive review of gnss/ins integration techniques for land and air vehicle applications,” Applied Sciences, vol. 13, no. 8, p. 4819, 2023

    work page 2023

  2. [2]

    Towards robust wifi fingerprint-based vehicle tracking in dynamic indoor parking environments: An online learning framework,

    K. Liu, F. Jin, J. Hu, R. Xie, F. Gu, S. Guo, and J. Luo, “Towards robust wifi fingerprint-based vehicle tracking in dynamic indoor parking environments: An online learning framework,” IEEE Transactions on Mobile Computing, vol. 22, no. 12, pp. 6970–6984, 2023

    work page 2023

  3. [3]

    Ai-imu dead-reckoning,

    M. Brossard, A. Barrau, and S. Bonnabel, “Ai-imu dead-reckoning,” IEEE Transactions on Intelligent Vehicles , vol. 5, no. 4, pp. 585–595, 2020

    work page 2020

  4. [4]

    Glow in the dark: Smartphone inertial odometry for vehicle tracking in gps blocked environments,

    R. Gao, X. Xiao, S. Zhu, W. Xing, C. Li, L. Liu, L. Ma, and H. Chai, “Glow in the dark: Smartphone inertial odometry for vehicle tracking in gps blocked environments,” IEEE Internet of Things Journal , vol. 8, no. 16, pp. 12 955–12 967, 2021

    work page 2021

  5. [5]

    Deepvip: Deep learning-based vehicle indoor positioning using smartphones,

    B. Zhou, Z. Gu, F. Gu, P. Wu, C. Yang, X. Liu, L. Li, Y . Li, and Q. Li, “Deepvip: Deep learning-based vehicle indoor positioning using smartphones,” IEEE Transactions on Vehicular Technology , vol. 71, no. 12, pp. 13 299–13 309, 2022

    work page 2022

  6. [6]

    Vehicle inertial tracking via mobile crowdsensing: Experience and enhancement,

    Y . Tong, S. Zhu, X. Ren, Q. Zhong, D. Tao, C. Li, L. Liu, and R. Gao, “Vehicle inertial tracking via mobile crowdsensing: Experience and enhancement,” IEEE Transactions on Instrumentation and Measurement, vol. 71, pp. 1–13, 2022

    work page 2022

  7. [7]

    High-integrity imm-ekf-based road vehicle navigation with low-cost gps/sbas/ins,

    R. Toledo-Moreo, M. A. Zamora-Izquierdo, B. Ubeda-Minarro, and A. F. Gomez-Skarmeta, “High-integrity imm-ekf-based road vehicle navigation with low-cost gps/sbas/ins,” IEEE Transactions on Intelligent Transportation Systems, vol. 8, no. 3, pp. 491–511, 2007

    work page 2007

  8. [8]

    Longitudinal vehicle speed estimation for four-wheel-independently-actuated electric vehicles based on multi-sensor fusion,

    X. Ding, Z. Wang, L. Zhang, and C. Wang, “Longitudinal vehicle speed estimation for four-wheel-independently-actuated electric vehicles based on multi-sensor fusion,” IEEE Transactions on Vehicular Technology , vol. 69, no. 11, pp. 12 797–12 806, 2020

    work page 2020

Show all 36 references
  1. [9]

    Sensing vehicle dynamics for determining driver phone use,

    Y . Wang, J. Yang, H. Liu, Y . Chen, M. Gruteser, and R. P. Martin, “Sensing vehicle dynamics for determining driver phone use,” in Pro- ceeding of the 11th annual international conference on Mobile systems, applications, and services , 2013, pp. 41–54

    2013

  2. [10]

    Estimating speed using a side- looking single-radar vehicle detector,

    S.-L. Jeng, W.-H. Chieng, and H.-P. Lu, “Estimating speed using a side- looking single-radar vehicle detector,” IEEE Transactions on Intelligent Transportation Systems, vol. 15, no. 2, pp. 607–614, 2014

    2014

  3. [11]

    Some practical vehicle speed estimation methods by a single traffic magnetic sensor,

    H. Li, H. Dong, L. Jia, D. Xu, and Y . Qin, “Some practical vehicle speed estimation methods by a single traffic magnetic sensor,” in 2011 14th International IEEE Conference on Intelligent Transportation Systems (ITSC). IEEE, 2011, pp. 1566–1573

    2011

  4. [12]

    Mag- monitor: Vehicle speed estimation and vehicle classification through a magnetic sensor,

    Y . Feng, G. Mao, B. Chen, C. Li, Y . Hui, Z. Xu, and J. Chen, “Mag- monitor: Vehicle speed estimation and vehicle classification through a magnetic sensor,” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 2, pp. 1311–1322, 2022

    2022

  5. [13]

    Tightly coupled integration of gnss/uwb/vio for reliable and seamless positioning,

    T. Liu, B. Li, G. Chen, L. Yang, J. Qiao, and W. Chen, “Tightly coupled integration of gnss/uwb/vio for reliable and seamless positioning,” IEEE Transactions on Intelligent Transportation Systems , vol. 25, no. 2, pp. 2116–2128, 2024

    2024

  6. [14]

    Vehicle positioning using wifi fingerprinting in urban environment,

    C.-W. Ang, “Vehicle positioning using wifi fingerprinting in urban environment,” in 2018 IEEE 4th World Forum on Internet of Things (WF-IoT), 2018, pp. 652–657

    2018

  7. [15]

    Doumiati, A

    M. Doumiati, A. Charara, A. Victorino, and D. Lechner, Vehicle dynam- ics estimation using Kalman filtering: experimental validation . John Wiley & Sons, 2012

    2012

  8. [16]

    Tightly-coupled fusion of vins and motion constraint for autonomous vehicle,

    Z. Yu, L. Zhu, and G. Lu, “Tightly-coupled fusion of vins and motion constraint for autonomous vehicle,” IEEE Transactions on Vehicular Technology, vol. 71, no. 6, pp. 5799–5810, 2022

    2022

  9. [17]

    Smartphone- based real time vehicle tracking in indoor parking structures,

    R. Gao, M. Zhao, T. Ye, F. Ye, Y . Wang, and G. Luo, “Smartphone- based real time vehicle tracking in indoor parking structures,” IEEE Transactions on Mobile Computing, vol. 16, no. 7, pp. 2023–2036, 2017

    2023

  10. [18]

    Ionet: Learning to cure the curse of drift in inertial odometry,

    C. Chen, X. Lu, A. Markham, and N. Trigoni, “Ionet: Learning to cure the curse of drift in inertial odometry,” in Proceedings of the AAAI Conference on Artificial Intelligence , vol. 32, no. 1, 2018

    2018

  11. [19]

    Ronin: Robust neural inertial navigation in the wild: Benchmark, evaluations, & new methods,

    S. Herath, H. Yan, and Y . Furukawa, “Ronin: Robust neural inertial navigation in the wild: Benchmark, evaluations, & new methods,” in 2020 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2020, pp. 3146–3152

    2020

  12. [20]

    A learning-based model predictive trajectory planning controller for automated driving in unstructured dynamic environments,

    Z. Li, P. Zhao, C. Jiang, W. Huang, and H. Liang, “A learning-based model predictive trajectory planning controller for automated driving in unstructured dynamic environments,” IEEE Transactions on Vehicular Technology, 2022

    2022

  13. [21]

    A review of recurrent neural networks: Lstm cells and network architectures,

    Y . Yu, X. Si, C. Hu, and J. Zhang, “A review of recurrent neural networks: Lstm cells and network architectures,” Neural computation , vol. 31, no. 7, pp. 1235–1270, 2019

    2019

  14. [22]

    Long short-term memory,

    S. Hochreiter and J. Schmidhuber, “Long short-term memory,” Neural computation, vol. 9, no. 8, pp. 1735–1780, 1997

    1997

  15. [23]

    Empirical evaluation of gated recurrent neural networks on sequence modeling,

    J. Chung, C. Gulcehre, K. Cho, and Y . Bengio, “Empirical evaluation of gated recurrent neural networks on sequence modeling,” arXiv preprint arXiv:1412.3555, 2014

    2014 arXiv

  16. [24]

    An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,

    S. Bai, J. Z. Kolter, and V . Koltun, “An empirical evaluation of generic convolutional and recurrent networks for sequence modeling,” arXiv preprint arXiv:1803.01271, 2018

    2018 arXiv

  17. [25]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems , vol. 30, 2017. 11

    2017

  18. [26]

    Informer: Beyond efficient transformer for long sequence time-series forecasting,

    H. Zhou, S. Zhang, J. Peng, S. Zhang, J. Li, H. Xiong, and W. Zhang, “Informer: Beyond efficient transformer for long sequence time-series forecasting,” in Proceedings of the AAAI conference on artificial intel- ligence, vol. 35, no. 12, 2021, pp. 11 106–11 115

    2021

  19. [27]

    Noise modeling and analysis of an imu- based attitude sensor: improvement of performance by filtering and sensor fusion,

    K. Nirmal, A. Sreejith, J. Mathew, M. Sarpotdar, A. Suresh, A. Prakash, M. Safonova, and J. Murthy, “Noise modeling and analysis of an imu- based attitude sensor: improvement of performance by filtering and sensor fusion,” in Advances in optical and mechanical technologies for...

    2016

  20. [28]

    Maximum likelihood identification of inertial sensor noise model parameters,

    J. Nikolic, P. Furgale, A. Melzer, and R. Siegwart, “Maximum likelihood identification of inertial sensor noise model parameters,” IEEE Sensors Journal, vol. 16, no. 1, pp. 163–176, 2015

    2015

  21. [29]

    Teacher forcing recovers reward func- tions for text generation,

    Y . Hao, Y . Liu, and L. Mou, “Teacher forcing recovers reward func- tions for text generation,” Advances in Neural Information Processing Systems, vol. 35, pp. 12 594–12 607, 2022

    2022

  22. [30]

    Spatial transformer networks,

    M. Jaderberg, K. Simonyan, A. Zisserman et al. , “Spatial transformer networks,” Advances in neural information processing systems , vol. 28, 2015

    2015

  23. [31]

    Vins-mono: A robust and versatile monocular visual-inertial state estimator,

    T. Qin, P. Li, and S. Shen, “Vins-mono: A robust and versatile monocular visual-inertial state estimator,” IEEE Transactions on Robotics , vol. 34, no. 4, pp. 1004–1020, 2018

    2018

  24. [32]

    A comprehensive survey of regres- sion based loss functions for time series forecasting,

    A. Jadon, A. Patil, and S. Jadon, “A comprehensive survey of regres- sion based loss functions for time series forecasting,” arXiv preprint arXiv:2211.02989, 2022

    2022 arXiv

  25. [33]

    Digging into self-supervised monocular depth estimation,

    C. Godard, O. Mac Aodha, M. Firman, and G. J. Brostow, “Digging into self-supervised monocular depth estimation,” in Proceedings of the IEEE/CVF international conference on computer vision, 2019, pp. 3828– 3838

    2019

  26. [34]

    Decoupled weight decay regularization,

    I. Loshchilov and F. Hutter, “Decoupled weight decay regularization,” arXiv preprint arXiv:1711.05101 , 2017

    2017 arXiv

  27. [35]

    tf.train.exponentialmovingaverage

    TensorFlow, “tf.train.exponentialmovingaverage.” [Online]. Available: https://www.tensorflow.org/api docs/python/tf/train/ ExponentialMovingAverage

  28. [36]

    Mixed precision training,

    P. Micikevicius, S. Narang, J. Alben, G. Diamos, E. Elsen, D. Garcia, B. Ginsburg, M. Houston, O. Kuchaiev, G. Venkatesh et al. , “Mixed precision training,” arXiv preprint arXiv:1710.03740 , 2017

    2017 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.