Pith. sign in

REVIEW 4 major objections 4 minor 50 references

Uncertainty-Aware Ankle Exoskeleton Control

T0 review · 4 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read An ankle exoskeleton can learn when to stand down, and this paper shows how.

desk verdict A genuinely useful feasibility study with an honest online test, but the stair results undercut the 'safe disengagement' claim and the conclusion oversells it. read the letter →

arxiv 2508.21221 v2 pith:QIOZU3EO submitted 2025-08-28 cs.RO

Fatima Mumtaza Tourk , Bishoy Galoaa , Sanat Shajan , Aaron J. Young , Michael Everett , Max K. Shepherd This is my paper
classification cs.RO
keywords ankleexoskeletonuncertaintyestimationout-of-distributiondetectiongaitphaseensemblemodelswearableroboticssafetyreal-timecontrol
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 argues that exoskeletons can be made safer in the real world by giving them an explicit sense of what they do not know. It proposes an uncertainty estimator that runs alongside a normal gait-phase controller and continuously asks whether incoming sensor data looks like the walking, jogging, and inclines the controller was trained on. If the data does not look familiar, the exoskeleton switches off assistance instead of risking an ill-timed torque. The strongest version of this estimator, a seven-model ensemble that tracks gait phase, was tested offline and then online on a new user navigating an outdoor course, where it switched assistance on and off correctly for most tasks. The paper's central claim is that this 'do nothing when unsure' layer is a practical path to safe, task-agnostic exoskeleton control.

What carries the argument

The ensemble of gait phase estimators: seven temporal convolutional networks, each predicting the sine of gait phase for both legs from a one-second window of IMU, encoder, and ankle-velocity data. The uncertainty score is the average variance across the seven predictions; high variance indicates out-of-distribution input. The system uses a threshold set only from training data (99.5th percentile) plus a causal median filter, so no out-of-distribution samples are needed to calibrate it.

What would settle it

Have a new user ascend stairs with the same shallow, ramp-like geometry that fooled the online test while the exoskeleton runs the ensemble; if the uncertainty score stays below the threshold for the entire ascent, the detector fails to disengage for a near-distribution out-of-distribution task, and the safety guarantee does not cover tasks that resemble the training distribution.

Watch

Extended reading notes

Core claim

The central discovery is that the variance across an ensemble of temporal convolutional networks predicting gait phase is a usable real-time uncertainty signal for human movement. Trained only on sensor data from walking and jogging on inclines, seven networks agree on familiar cyclic motions and disagree on unfamiliar ones; the averaged variance of their left/right gait-phase predictions becomes an anomaly score. A fixed threshold set at the 99.5th percentile of training-data scores separates in-distribution from out-of-distribution action, and when the score exceeds the threshold the exoskeleton's actuators unspool to zero impedance. Offline, this ensemble outperformed an autoencoder with

Load-bearing premise

The paper assumes that the variance across ensemble gait-phase predictions is a reliable signal that incoming sensor data is outside the training distribution; the stair results show that out-of-distribution movements that biomechanically resemble training data can escape detection.

Editorial extensions

If this is right

  • An existing exoskeleton controller can be wrapped in a safety layer that decides when to actuate, without retraining the base controller.
  • The uncertainty layer can run in real time on onboard hardware, as demonstrated by the paper's end-to-end 105 Hz operation.
  • The threshold recipe using only training data means a lab only needs to collect in-distribution data to set up the safety switch for a new controller.
  • Future controllers trained on non-cyclic tasks could use the same framework if a label such as a synthetic correlation target replaces gait phase.
  • Clearly dangerous actions, like feet leaving the ground during jumping, produce high uncertainty, so the system is most conservative at the moments when misfires cause tripping.

Reading between the lines

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

  • Because the stair section failed on shallow, ramp-like steps, safety coverage is likely limited to out-of-distribution tasks that are clearly distinct from training data; a complementary detector would be needed for tasks that resemble known movements.
  • The binary on/off behavior could be extended to continuous torque scaling based on uncertainty magnitude, which might reduce trust loss and transition jolts, an extension the paper did not test.
  • The ensemble's dependence on gait-phase labels confines its in-distribution set to cyclic actions; closing the gap with the synthetic-target ensemble would extend the framework to non-cyclic in-distribution tasks.
  • The fixed 99.5th-percentile threshold is sensitive to which subjects and trials are in the training set; an adaptive or per-user threshold may be needed for broader deployment.
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

4 major / 4 minor

Summary. This paper proposes an uncertainty-aware gating layer for ankle exoskeletons. An ensemble of seven TCN gait-phase estimators is trained on walking/jogging/incline data; the variance of ensemble predictions is used as an uncertainty score, and a threshold fixed at the 99.5th percentile of training-only uncertainty scores decides when the exoskeleton should disengage. The authors compare this ensemble with autoencoder and GAN alternatives offline, then deploy the ensemble online on a novel subject completing an outdoor circuit. The reported overall online F1 is 89.2 with J-statistic 73.5, but stair ascent/descent accuracy is only 58.7%. The paper argues that this framework enables safe operation across diverse real-world activities by turning off assistance on out-of-distribution movements.

Significance. If the central safety claim is supported, this is a valuable systems contribution: it is among the first real-time uncertainty estimators used to gate wearable-robot assistance, and the decision to set the threshold using only training data is methodologically principled. The paper also provides a useful three-way architecture comparison and detailed model specifications, and the online test is a genuine external check with a new subject and new tasks. However, the current evidence does not yet support the safety claim: the detector's worst online performance occurs on stairs, a safety-critical near-OOD task, and the Discussion's mitigation rests on an untested assumption.

major comments (4)
  1. [Section V-C, Table III; Section VI] The paper's own online results undermine the central safety claim. Stair accuracy is 58.7% because, as the authors state, flatter, wider stairs 'likely provided less of a distinction from ramps,' causing much of stair ascent/descent to be labeled in-distribution. Stairs are safety-critical: incorrect ankle actuation during swing can cause trips or falls. The Discussion's assertion that false positives on biomechanically similar OOD tasks are 'less likely to be safety critical' is unsupported; the base controller was trained on walking/jogging/ramps, and its torque profiles are not validated on stairs. To support the safety claim, the authors need either to demonstrate that the phase-based controller remains safe on stair-like OOD inputs, or to narrow the claim to clearly distinct OOD tasks and provide a separate mechanism for near-OOD cyclic tasks.
  2. [Section IV-A] Fixing the threshold at the 99.5th percentile of training uncertainty scores is an independent design choice, but it entrenches a blind spot for moderate-variance OOD inputs. The authors themselves note that the threshold 'may be highly dependent on a few outlier movements in the training set' and that the threshold was not derived from OOD data. The paper provides no sensitivity analysis showing how classification of near-OOD tasks changes with the threshold, nor any characterization of ensemble variance on stair-like inputs. Since the central mechanism relies on variance separating OOD from in-distribution, this missing analysis is load-bearing rather than cosmetic.
  3. [Section V-B] The online validation uses a single novel subject, and the ground-truth labels are subjective judgments from three external observers with no reported inter-rater agreement. Given that the stair section achieves near-chance accuracy and that transition regions are labeled by the same process, the reported F1 and J-statistic are not robust enough to support the conclusion of 'robust real-time uncertainty estimation.' The authors should report agreement statistics (e.g., Cohen's kappa), provide per-subject results if available, and temper the conclusion to reflect the n=1 nature of the online validation.
  4. [Section IV-A and Section VI] The Discussion's claim that the model worked 'out of the box' and that a validation set 'may not be needed' is not supported by the methodology. Section IV-A states that D_ID,OOD_val was used for model development and hyperparameter tuning; architecture selection, ensemble size, latent-space size, GAN hyperparameters, and median-filter length were all chosen with access to OOD validation data. Only the threshold was fixed using training data alone. The 'out of the box' statement should be limited to the threshold-setting procedure, not the entire model-selection pipeline.
minor comments (4)
  1. [Section III-C] In the GAN description, the sentence 'the discriminator learns to determine when it receives real in-distribution data and when it receives fabricated data from the discriminator' should read 'from the generator.'
  2. [Section II] The data-preparation description says windows were generated 'using every tenth window.' Please clarify the stride and overlap, since this affects potential temporal leakage between training and validation windows.
  3. [Figure 3] The uncertainty-score distributions use green and red bars; if the journal is printed in grayscale or readers are color-blind, the separation will be hard to read. Consider adding patterns or labels.
  4. [Section IV-C] The sentence 'This approach has the advantage of being able to learn complex patterns but is often unstable in training' appears in the GAN section and is repeated conceptually in the Discussion; it could be consolidated.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the uncertainty signal is an ensemble variance, the threshold is fixed from training data only, and the online ground-truth labels come from external observers.

full rationale

The paper's central claim -- that an uncertainty estimator can switch assistance on and off for in-distribution vs out-of-distribution movements -- is not forced by construction. The uncertainty score is the variance across seven independently initialized gait-phase TCNs (Eq. 2), trained only on the in-distribution dataset. The decision threshold is explicitly fixed as the 99.5th percentile of training-only uncertainty scores, not optimized on out-of-distribution labels, so the online F1=89.2 is not a refit of the threshold. Online ground-truth labels were provided by three external observers instructed to mark walking/jogging as in-distribution and everything else as out-of-distribution; this label definition is independent of the uncertainty score itself. The base TCN hyperparameters are inherited from the authors' prior work [37], but that is a standard reuse of an independently published architecture rather than a circular reduction, and the prior work was externally evaluated. The observed stair misclassification (58.7% accuracy, Table III) is an empirical failure of the ensemble-variance assumption for near-distribution out-of-distribution tasks; it weakens the safety claim but is not an instance of the result being equivalent to its inputs. No equation or fitted parameter is renamed as a prediction, and no load-bearing argument reduces to a self-citation chain. The derivation is therefore self-contained with respect to circularity.

Assumptions & free parameters 7 free parameters · 5 assumptions · 0 invented entities

The framework introduces no new physical entities or forces. It relies on a small set of free design choices, chiefly the threshold, filter length, ensemble size, and architecture hyperparameters, several of which are tuned on the validation set. The main domain assumptions are that ensemble variance acts as a valid OOD signal and that the chosen training set defines the safe operating envelope.

free parameters (7)
  • Uncertainty threshold = 99.5th percentile of training uncertainty scores
    Chosen as a design target for a 0.5% false negative rate on training data, not fitted to OOD labels. The authors note it may be sensitive to a few outlier training movements.
  • Median filter length = 88 points (~0.5 s)
    Selected to smooth spikes while limiting transition delay to approximately 0.25 s, based on an assumption about human movement frequency content.
  • Number of ensemble models = 7
    The authors state that including more than 7 models did not further improve performance, implying this was tuned on the validation set.
  • TCN hyperparameters = 30 filters/layer, kernel size 20, 3 layers
    Taken from prior work (Shetty et al.), not re-tuned in this paper.
  • Autoencoder latent space size = 28 (7 time steps x 4 filter dims)
    Swept over a range to maximize J-statistic on the offline validation set, as shown in Fig. 4.
  • GAN hyperparameters = spectral norm, G LR=2e-4, D LR=5e-5, 5 G updates per 1 D update
    Tuned manually by monitoring generator/discriminator strength, using the validation set.
  • Architecture selection (ensemble vs AE vs GAN) = Ensemble (gait phase)
    The best architecture was selected based on offline validation performance, introducing selection bias into the online evaluation.
assumptions (5)
  • domain assumption Ensemble prediction variance is a reliable OOD indicator
    Core premise: models that agree on in-distribution data will diverge on OOD data. Empirically validated on the offline set, but the stair misclassification shows it does not hold for all near-distribution OOD tasks (Section V-C).
  • domain assumption Training data (9 subjects walking/jogging on inclines) represents the full in-distribution class
    All in-distribution labels in offline and online tests are defined relative to this set. The paper assumes this set is the correct reference for safe assistance.
  • domain assumption A 0.5% false negative rate on training data is a safe operating point
    The threshold is set to admit only 0.5% of training windows as OOD. There is no safety analysis justifying this exact rate.
  • domain assumption Gait phase is a well-defined label for cyclic training actions
    The ensemble method requires ground-truth gait phase for training, which only exists for cyclic actions. This limits the in-distribution set to cyclic ambulation.
  • domain assumption The offline validation set is representative of real-world OOD tasks
    The authors acknowledge that the full validation set was used for model development, and the online circuit provides new OOD tasks, but there is no guarantee the validation set covers the space of safety-relevant movements.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Uncertainty-Aware Ankle Exoskeleton Control." pith.science (2026). https://pith.science/paper/QIOZU3EO

@misc{pith2026250821221,
  author       = {Pith},
  title        = {Pith review of: Uncertainty-Aware Ankle Exoskeleton Control},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/QIOZU3EO}},
  note         = {Machine review of arXiv:2508.21221}
}
read the original abstract

Lower limb exoskeletons show promise to assist human movement, but their utility is limited by controllers designed for discrete, predefined actions in controlled environments, restricting their real-world applicability. We present an uncertainty-aware control framework that enables ankle exoskeletons to operate safely across diverse scenarios by automatically disengaging when encountering unfamiliar movements. Our approach uses an uncertainty estimator to classify movements as similar (in-distribution) or different (out-of-distribution) relative to actions in the training set. We evaluated three architectures (model ensembles, autoencoders, and generative adversarial networks) on an offline dataset and tested the strongest performing architecture (ensemble of gait phase estimators) online. The online test demonstrated the ability of our uncertainty estimator to turn assistance on and off as the user transitioned between in-distribution and out-of-distribution tasks (F1: 89.2). This new framework provides a path for exoskeletons to safely and autonomously support human movement in unstructured, everyday environments.

Figures

Figures reproduced from arXiv: 2508.21221 by the authors.

Figure 1
Figure 1. Overview of the uncertainty classifier system for exoskeleton control. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Neural network architectures for uncertainty classification: (A) Autoencoder approach with TCN-based encoder and decoder to measure reconstruction [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Offline test results: Distribution of uncertainty scores across the full duration of in-distribution (green) trials and out-of-distribution (red) trials, shown [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (3 more)
Figure 4
Figure 4. Figure 4: From left to right, a) Latent Space Size vs. Reconstruction MSE of Training, In Distribution Validation, and Out of Distribution Dataset. b) 3D plot [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Test route around Northeastern University’s Krentzman Quadrangle: [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Participant performing various in- and out-of-distribution actions on the outdoor course, with labeled actions and corresponding anomaly score plotted [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 39 canonical work pages

  1. [1]

    The role of lower limb exoskeletons in rehabilitation: a scoping review,

    P. Abery, E. F.D. Canetti, and W. Hing, “The role of lower limb exoskeletons in rehabilitation: a scoping review,” Physical Therapy Reviews , pp. 1–19, Feb. 2025. [Online]. Available: https: //www.tandfonline.com/doi/full/10.1080/10833196.2025.2465937

    arXiv 2025

  2. [2]

    Exoskeletons for Lower Limb Applications: A Review,

    M. S. Alqahtani, G. Cooper, C. Diver, and P. J. B ´artolo, “Exoskeletons for Lower Limb Applications: A Review,” in Bio-Materials and Prototyping Applications in Medicine, P. J. B´artolo and B. Bidanda, Eds. Cham: Springer International Publishing, 2021, pp. 139–164. [Online]. Available: http://link.springer.com/10.1007/978-3-030-35876-1 8

    work page doi:10.1007/978-3-030-35876-1 2021

  3. [3]

    Review of control strategies for lower-limb exoskeletons to assist gait,

    R. Baud, A. R. Manzoori, A. Ijspeert, and M. Bouri, “Review of control strategies for lower-limb exoskeletons to assist gait,” Journal of NeuroEngineering and Rehabilitation , vol. 18, no. 1, p. 119, Jul. 2021. [Online]. Available: https://doi.org/10.1186/s12984-021-00906-3

    work page doi:10.1186/s12984-021-00906-3 2021

  4. [4]

    Human-in-the-loop optimization of exoskeleton assistance during walking,

    J. Zhang, P. Fiers, K. A. Witte, R. W. Jackson, K. L. Poggensee, C. G. Atkeson, and S. H. Collins, “Human-in-the-loop optimization of exoskeleton assistance during walking,” Science, vol. 356, no. 6344, pp. 1280–1284, 2017

    2017

  5. [5]

    On human-in-the-loop optimization of human–robot interaction,

    P. Slade, C. Atkeson, J. M. Donelan, H. Houdijk, K. A. Ingraham, M. Kim, K. Kong, K. L. Poggensee, R. Riener, M. Steinert et al., “On human-in-the-loop optimization of human–robot interaction,” Nature, vol. 633, no. 8031, pp. 779–788, 2024

    2024

  6. [6]

    The role of user preference in the customized control of robotic exoskeletons,

    K. A. Ingraham, C. D. Remy, and E. J. Rouse, “The role of user preference in the customized control of robotic exoskeletons,” Science robotics, vol. 7, no. 64, p. eabj3487, 2022

    work page 2022

  7. [7]

    Leveraging user preference in the design and evaluation of lower-limb exoskeletons and prostheses,

    K. A. Ingraham, M. Tucker, A. D. Ames, E. J. Rouse, and M. K. Shepherd, “Leveraging user preference in the design and evaluation of lower-limb exoskeletons and prostheses,”Current Opinion in Biomedical Engineering, vol. 28, p. 100487, 2023

    work page 2023

  8. [8]

    Human-in-the-loop optimization of hip exoskeleton assistance during stair climbing,

    D. Park, J. An, D. Lee, I. Kang, and A. J. Young, “Human-in-the-loop optimization of hip exoskeleton assistance during stair climbing,” IEEE Transactions on Biomedical Engineering , 2025

    work page 2025

Show all 50 references
  1. [9]

    Human-in-the- loop optimization of hip assistance with a soft exosuit during walking,

    Y . Ding, M. Kim, S. Kuindersma, and C. J. Walsh, “Human-in-the- loop optimization of hip assistance with a soft exosuit during walking,” Science robotics, vol. 3, no. 15, p. eaar5438, 2018

    2018

  2. [10]

    Opportunities and challenges in the development of exoskeletons for locomotor assistance,

    C. Siviy, L. M. Baker, B. T. Quinlivan, F. Porciuncula, K. Swaminathan, L. N. Awad, and C. J. Walsh, “Opportunities and challenges in the development of exoskeletons for locomotor assistance,” Nature Biomedical Engineering, vol. 7, no. 4, pp. 456–472, Dec. 2022. [Online]. Avai...

    2022

  3. [11]

    Real-time gait phase and task estimation for controlling a powered ankle exoskeleton on extremely uneven terrain,

    R. L. Medrano, G. C. Thomas, C. G. Keais, E. J. Rouse, and R. D. Gregg, “Real-time gait phase and task estimation for controlling a powered ankle exoskeleton on extremely uneven terrain,” IEEE Transactions on Robotics, vol. 39, no. 3, pp. 2170–2182, 2023

    2023

  4. [12]

    Deep Learning Enables Exoboot Control to Augment Variable-Speed Walking,

    M. K. Shepherd, D. D. Molinaro, G. S. Sawicki, and A. J. Young, “Deep Learning Enables Exoboot Control to Augment Variable-Speed Walking,” IEEE Robotics and Automation Letters , vol. 7, no. 2, pp. 3571–3577, Apr. 2022. [Online]. Available: https://ieeexplore.ieee.org/document/9699101/

    2022

  5. [13]

    A versatile knee exoskeleton mitigates quadriceps fatigue in lifting, lowering, and carrying tasks,

    N. V . Divekar, G. C. Thomas, A. R. Yerva, H. B. Frame, and R. D. Gregg, “A versatile knee exoskeleton mitigates quadriceps fatigue in lifting, lowering, and carrying tasks,” Science Robotics , vol. 9, no. 94, p. eadr8282, 2024

    2024

  6. [14]

    Deep learning approaches for enhanced lower-limb exoskeleton control: A review,

    M. Belal, N. Alsheikh, A. Aljarah, and I. Hussain, “Deep learning approaches for enhanced lower-limb exoskeleton control: A review,” IEEE Access, vol. 12, pp. 143 883–143 907, 2024

    2024

  7. [15]

    Real-time neural network- based gait phase estimation using a robotic hip exoskeleton,

    I. Kang, P. Kunapuli, and A. J. Young, “Real-time neural network- based gait phase estimation using a robotic hip exoskeleton,” IEEE Transactions on Medical Robotics and Bionics , vol. 2, no. 1, pp. 28–37, 2019

    2019

  8. [16]

    Personalizing exoskeleton assistance while walking in the real world,

    P. Slade, M. J. Kochenderfer, S. L. Delp, and S. H. Collins, “Personalizing exoskeleton assistance while walking in the real world,” Nature, vol. 610, no. 7931, pp. 277–282, Oct. 2022. [Online]. Available: https://www.nature.com/articles/s41586-022-05191-1

    2022

  9. [17]

    Real-time experimental validation of an adaptive gait event detection algorithm,

    J. A. Strick, J. J. Wiebrecht, R. J. Farris, and J. T. Sawicki, “Real-time experimental validation of an adaptive gait event detection algorithm,” IEEE Sensors Journal , 2025

    2025

  10. [18]

    Robustification of bayesian-inference-based gait estimation for lower-limb wearable robots,

    T.-W. Hsu, R. D. Gregg, and G. C. Thomas, “Robustification of bayesian-inference-based gait estimation for lower-limb wearable robots,” IEEE robotics and automation letters , vol. 9, no. 3, pp. 2104– 2111, 2024

    2024

  11. [19]

    Oscillator-based assistance of cyclical movements: model-based and model-free approaches,

    R. Ronsse, T. Lenzi, N. Vitiello, B. Koopman, E. Van Asseldonk, S. M. M. De Rossi, J. Van Den Kieboom, H. Van Der Kooij, M. C. Carrozza, and A. J. Ijspeert, “Oscillator-based assistance of cyclical movements: model-based and model-free approaches,” Medical & bio- logical engin...

    2011

  12. [20]

    A lower limb wearable exosuit for improved sitting, standing, and walking efficiency,

    X. Zhang, E. Tricomi, X. Ma, M. Gomez-Correa, A. Ciaramella, F. Missiroli, L. Miˇskovi´c, H. Su, and L. Masia, “A lower limb wearable exosuit for improved sitting, standing, and walking efficiency,” IEEE Transactions on Robotics , 2024

    2024

  13. [21]

    Task-agnostic exoskeleton JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 11 control via biological joint moment estimation,

    D. D. Molinaro, K. L. Scherpereel, E. B. Schonhaut, G. Evangelopoulos, M. K. Shepherd, and A. J. Young, “Task-agnostic exoskeleton JOURNAL OF LATEX CLASS FILES, VOL. 14, NO. 8, AUGUST 2021 11 control via biological joint moment estimation,” Nature, vol. 635, no. 8038, pp. 337–...

    2021

  14. [22]

    Impact of imperfect exoskeleton algorithms on step characteristics, task performance, and perception of exoskeleton performance,

    M. I. Wu and L. Stirling, “Impact of imperfect exoskeleton algorithms on step characteristics, task performance, and perception of exoskeleton performance,” in 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2023, pp. 4088–4093

    2023

  15. [23]

    Measuring Trust for Exoskeleton Systems,

    L. Stirling, M. I. Wu, and X. Peng, “Measuring Trust for Exoskeleton Systems,” 2024, version Number: 1. [Online]. Available: https://arxiv.org/abs/2407.07200

    2024 arXiv

  16. [24]

    Anomaly detection in time series: a comprehensive evaluation,

    S. Schmidl, P. Wenig, and T. Papenbrock, “Anomaly detection in time series: a comprehensive evaluation,” Proceedings of the VLDB Endowment, vol. 15, no. 9, pp. 1779–1797, May 2022. [Online]. Available: https://dl.acm.org/doi/10.14778/3538598.3538602

    2022

  17. [25]

    A review on outlier/anomaly detection in time series data,

    A. Bl ´azquez-Garc´ıa, A. Conde, U. Mori, and J. A. Lozano, “A review on outlier/anomaly detection in time series data,” Feb. 2020, arXiv:2002.04236 [cs]. [Online]. Available: http://arxiv.org/abs/2002. 04236

    2020 arXiv

  18. [26]

    Anomaly detection in multivariate time series data using deep ensemble models,

    A. Iqbal, R. Amin, F. S. Alsubaei, and A. Alzahrani, “Anomaly detection in multivariate time series data using deep ensemble models,” PLOS ONE, vol. 19, no. 6, p. e0303890, Jun. 2024. [Online]. Available: https://dx.plos.org/10.1371/journal.pone.0303890

    2024 doi

  19. [27]

    Deep learning for anomaly detection in multivariate time series: Approaches, applications, and challenges,

    G. Li and J. J. Jung, “Deep learning for anomaly detection in multivariate time series: Approaches, applications, and challenges,” Information Fusion, vol. 91, pp. 93–102, Mar. 2023. [Online]. Available: https://linkinghub.elsevier.com/retrieve/pii/S1566253522001774

    2023

  20. [28]

    Deep Learning for Anomaly Detection in Time-Series Data: Review, Analysis, and Guidelines,

    K. Choi, J. Yi, C. Park, and S. Yoon, “Deep Learning for Anomaly Detection in Time-Series Data: Review, Analysis, and Guidelines,” IEEE Access , vol. 9, pp. 120 043–120 065, 2021. [Online]. Available: https://ieeexplore.ieee.org/document/9523565/

    2021

  21. [29]

    Safe Reinforcement Learning with Model Uncertainty Estimates,

    B. L ¨utjens, M. Everett, and J. P. How, “Safe Reinforcement Learning with Model Uncertainty Estimates,” Mar. 2019, arXiv:1810.08700 [cs]. [Online]. Available: http://arxiv.org/abs/1810.08700

    2019 arXiv

  22. [30]

    A General Framework for Uncertainty Estimation in Deep Learning,

    A. Loquercio, M. Seg `u, and D. Scaramuzza, “A General Framework for Uncertainty Estimation in Deep Learning,” IEEE Robotics and Automation Letters , vol. 5, no. 2, pp. 3153–3160, Apr. 2020, arXiv:1907.06890 [cs]. [Online]. Available: http://arxiv.org/abs/1907. 06890

    2020 arXiv

  23. [31]

    Uncertainty- Aware Reinforcement Learning for Collision Avoidance,

    G. Kahn, A. Villaflor, V . Pong, P. Abbeel, and S. Levine, “Uncertainty- Aware Reinforcement Learning for Collision Avoidance,” Feb. 2017, arXiv:1702.01182 [cs]. [Online]. Available: http://arxiv.org/abs/1702. 01182

    2017 arXiv

  24. [32]

    Variational Autoencoder based Anomaly Detection using Reconstruction Probability,

    J. An and S. Cho, “Variational Autoencoder based Anomaly Detection using Reconstruction Probability,” SNU Data Mining Center , vol. 2015-2 Special Lecture on IE, Dec. 2015. [Online]. Available: http://dm.snu.ac.kr/static/docs/TR/SNUDM-TR-2015-03.pdf

    2015

  25. [33]

    A Survey on GANs for Anomaly Detection,

    F. D. Mattia, P. Galeone, M. D. Simoni, and E. Ghelfi, “A Survey on GANs for Anomaly Detection,” Sep. 2021, arXiv:1906.11632 [cs]. [Online]. Available: http://arxiv.org/abs/1906.11632

    2021 arXiv

  26. [34]

    Ensemble of Autoencoders for Anomaly Detection in Biomedical Data: A Narrative Review,

    A. Nawaz, S. S. Khan, and A. Ahmad, “Ensemble of Autoencoders for Anomaly Detection in Biomedical Data: A Narrative Review,” IEEE Access , vol. 12, pp. 17 273–17 289, 2024. [Online]. Available: https://ieeexplore.ieee.org/document/10418130/

    2024

  27. [35]

    Using a V AE-SOM architecture for anomaly detection of flexible sensors in limb prosthesis,

    Z. Zhu, P. Su, S. Zhong, J. Huang, S. Ottikkutti, K. N. Tahmasebi, Z. Zou, L. Zheng, and D. Chen, “Using a V AE-SOM architecture for anomaly detection of flexible sensors in limb prosthesis,” Journal of Industrial Information Integration , vol. 35, p. 100490, Oct

  28. [36]

    A Systematic Method for Outlier Detection in Human Gait Data,

    B. Hobbs and P. Artemiadis, “A Systematic Method for Outlier Detection in Human Gait Data,” in 2022 International Conference on Rehabilitation Robotics (ICORR) . Rotterdam, Netherlands: IEEE, Jul. 2022, pp. 1–6. [Online]. Available: https://ieeexplore.ieee.org/document/ 9896411/

    2022

  29. [37]

    Ankle Exoskeleton Control via Data-Driven Gait Estimation for Walking, Running, and Inclines,

    P. R. Shetty, J. A. Menezes, S. Song, A. J. Young, and M. K. Shepherd, “Ankle Exoskeleton Control via Data-Driven Gait Estimation for Walking, Running, and Inclines,” IEEE Robotics and Automation Letters, vol. 10, no. 6, pp. 5855–5862, Jun. 2025, publisher: Institute of Electr...

    2025

  30. [38]

    Subject-Independent, Biological Hip Moment Estimation During Multimodal Overground Ambulation Using Deep Learning,

    D. D. Molinaro, I. Kang, J. Camargo, M. C. Gombolay, and A. J. Young, “Subject-Independent, Biological Hip Moment Estimation During Multimodal Overground Ambulation Using Deep Learning,” IEEE Transactions on Medical Robotics and Bionics , vol. 4, no. 1, pp. 219–229, Feb. 2022....

    2022

  31. [39]

    Real-Time Gait Phase Estimation for Robotic Hip Exoskeleton Control During Multimodal Locomotion,

    I. Kang, D. D. Molinaro, S. Duggal, Y . Chen, P. Kunapuli, and A. J. Young, “Real-Time Gait Phase Estimation for Robotic Hip Exoskeleton Control During Multimodal Locomotion,” IEEE Robotics and Automation Letters , vol. 6, no. 2, pp. 3491–3497, Apr. 2021. [Online]. Available: ...

    2021

  32. [40]

    Real-Time Neural Network- Based Gait Phase Estimation Using a Robotic Hip Exoskeleton,

    I. Kang, P. Kunapuli, and A. J. Young, “Real-Time Neural Network- Based Gait Phase Estimation Using a Robotic Hip Exoskeleton,” IEEE Transactions on Medical Robotics and Bionics , vol. 2, no. 1, pp. 28–37, Feb. 2020. [Online]. Available: https://ieeexplore.ieee.org/ document/8941004/

    2020

  33. [41]

    Contrastive autoen- coder for anomaly detection in multivariate time series,

    H. Zhou, K. Yu, X. Zhang, G. Wu, and A. Yazidi, “Contrastive autoen- coder for anomaly detection in multivariate time series,” Information Sciences, vol. 614, pp. 214–229, 2022

    2022

  34. [42]

    Autoencoders for anomaly detection in an industrial multivariate time series dataset,

    K. Asimakis, N. Karampasis, T. Kotsiopoulos, A. Manikas, P. Sarigianni- dis, and T. Lagkas, “Autoencoders for anomaly detection in an industrial multivariate time series dataset,” Machine Learning and Knowledge Extraction, vol. 6, no. 1, pp. 23–40, 2024

    2024

  35. [43]

    Temporal convolutional autoencoder for unsupervised anomaly detection in time series,

    M. Thill, W. Konen, H. Wang, and T. B ¨ack, “Temporal convolutional autoencoder for unsupervised anomaly detection in time series,” Applied Soft Computing, vol. 112, p. 107756, 2021

    2021

  36. [44]

    Generative Adversarial Networks,

    I. J. Goodfellow, J. Pouget-Abadie, M. Mirza, B. Xu, D. Warde-Farley, S. Ozair, A. Courville, and Y . Bengio, “Generative Adversarial Networks,” 2014, version Number: 1. [Online]. Available: https: //arxiv.org/abs/1406.2661

    2014 arXiv

  37. [45]

    D. A. Winter, Biomechanics and motor control of human movement . John wiley & sons, 2009

    2009

  38. [46]

    Index for rating diagnostic tests,

    W. J. Youden, “Index for rating diagnostic tests,” Cancer, vol. 3, no. 1, pp. 32–35, 1950. [Online]. Available: https://acsjournals.onlinelibrary. wiley.com/doi/abs/10.1002/1097-0142%281950%293%3A1%3C32% 3A%3AAID-CNCR2820030106%3E3.0.CO%3B2-3

    1950 doi

  39. [47]

    Youden Index and Optimal Cut-Point Estimated from Observations Affected by a Lower Limit of Detection,

    M. D. Ruopp, N. J. Perkins, B. W. Whitcomb, and E. F. Schisterman, “Youden Index and Optimal Cut-Point Estimated from Observations Affected by a Lower Limit of Detection,” Biometrical Journal, vol. 50, no. 3, pp. 419–430, Jun. 2008. [Online]. Available: https://onlinelibrary.w...

    2008 doi

  40. [48]

    Biomechanical models in the lower-limb exoskeletons development: A review,

    V . Firouzi, A. Seyfarth, S. Song, O. von Stryk, and M. Ahmad Sharbafi, “Biomechanical models in the lower-limb exoskeletons development: A review,”Journal of NeuroEngineering and Rehabilitation, vol. 22, no. 1, p. 12, 2025

    2025

  41. [49]

    Neuromechanical model-based adaptive control of bilateral ankle ex- oskeletons: Biological joint torque and electromyogram reduction across walking conditions,

    G. Durandau, W. F. Rampeltshammer, H. van der Kooij, and M. Sartori, “Neuromechanical model-based adaptive control of bilateral ankle ex- oskeletons: Biological joint torque and electromyogram reduction across walking conditions,” IEEE Transactions on Robotics, vol. 38, no. 3,...

    2022

  42. [2023]

    Available: https://linkinghub.elsevier.com/retrieve/pii/ S2452414X23000638

    [Online]. Available: https://linkinghub.elsevier.com/retrieve/pii/ S2452414X23000638

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.