pith. sign in

arxiv: 2606.07902 · v1 · pith:ER2OUWDXnew · submitted 2026-06-05 · 💻 cs.RO

End-to-End Control of a Powered Knee-Ankle Prosthesis Towards Unified, Tuning-Free Assistance

John Shim , Christoph Nuesslein , Sixu Zhou , Hanjun kim , Kinsey Herrin , Aaron Young This is my paper

Pith reviewed 2026-06-27 21:24 UTC · model grok-4.3

classification 💻 cs.RO
keywords powered prosthesisend-to-end controltemporal convolutional networklocomotion modesknee-ankle prosthesisamputee assistancemode-adaptive controlreal-time deployment
0
0 comments X

The pith

An end-to-end neural controller for powered knee-ankle prostheses reproduces training-data scalings for ankle torque, knee flexion, and resistive torque across five locomotion modes without subject-specific tuning or mode classifiers.

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

The paper shows that a neural network can map onboard sensor readings directly to continuous knee and ankle actuator commands for a powered prosthesis. Training occurred on a dataset covering level ground, ramps, and stairs collected from 18 individuals with transfemoral amputation. When deployed in real time, the same model produced actuator outputs that matched the scaling relationships seen in the original training data for walking speed, ramp grade, and descent resistance. It also generated smooth transitions during stair ascent and descent even though the training set contained only one limb-leading sequence. These behaviors held for both able-bodied and amputee test participants, removing the need for separate intent detectors or manual impedance adjustments.

Core claim

A temporal convolutional network trained end-to-end on multi-terrain sensor data from 18 transfemoral amputees can be deployed in real time to estimate continuous actuator signals that reproduce the training-data scaling of peak ankle torque with walking speed (0.85 Nm/kg per m/s), knee pre-flexion with ramp grade (2.92 deg/deg), and resistive knee torque on descent (+0.16 Nm/kg), while producing seamless stair transitions for both leading-limb sequences, all without explicit mode classification or subject-specific tuning.

What carries the argument

Temporal Convolutional Networks that take onboard sensor inputs and output continuous knee and ankle actuator signals, trained across multiple locomotion modes to replace separate classifiers and impedance parameters.

If this is right

  • A single model can supply actuator commands for level walking, ramp ascent, ramp descent, stair ascent, and stair descent.
  • Scaling relationships observed in the training data for torque and angle are preserved during real-time use.
  • Stair transitions remain continuous even when the training data contain only one limb-leading sequence.
  • Both able-bodied and amputee users can receive assistance from the same deployed controller.

Where Pith is reading between the lines

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

  • Fitting time for new users could decrease if the controller generalizes beyond the tested participants.
  • The same end-to-end approach might extend to additional sensors or more variable environments such as uneven ground.
  • Long-term stability of the reproduced scalings could be checked by repeated deployments over multiple sessions.

Load-bearing premise

That performance observed on four test participants after training on 18 amputees will generalize to new amputee users without any subject-specific tuning or retraining.

What would settle it

Deploy the controller on additional transfemoral amputees not in the training set and check whether the measured ankle torque versus speed slope, knee pre-flexion versus grade slope, and descent torque offset remain statistically consistent with the training-data values.

Figures

Figures reproduced from arXiv: 2606.07902 by Aaron Young, Christoph Nuesslein, Hanjun Kim, John Shim, Kinsey Herrin, Sixu Zhou.

Figure 1
Figure 1. Figure 1: Proposed end-to-end control framework. A deep neural network maps onboard sensor measurements—joint encoders (position, velocity), IMU (3-axis [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: (a) Onboard sensor configuration of the Open-Source Leg. (b) Training [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Offline swing-phase trajectory estimation R² across prediction horizons [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: (a) Peak stance ankle torque across walking speeds, with subject-level data points and the subject-averaged real-time mean and training-data mean [PITH_FULL_IMAGE:figures/full_fig_p005_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Ramp adaptation results. (a) Peak knee pre-flexion vs. ramp grade [PITH_FULL_IMAGE:figures/full_fig_p005_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Terrain transition behavior for TF01 across full stair circuits. Shaded regions mark swing phases throughout (indicating joint angles, light lines); [PITH_FULL_IMAGE:figures/full_fig_p006_6.png] view at source ↗
read the original abstract

Powered prostheses conventionally rely on impedance controllers that require extensive manual tuning and explicit mode classification. In this work, we present real-time deployment of an end-to-end prosthesis controller that estimates continuous actuator signals from onboard sensors, eliminating the need for intent classifiers and subject-specific tuning. Temporal Convolutional Networks were trained on a multi-terrain dataset from 18 individuals with transfemoral amputation and deployed in real time across five locomotion modes. Four participants (three able-bodied, one with transfemoral amputation) ambulated across level ground, ramp ascent and descent, and stair ascent and descent. During level walking, the deployed controller reproduced the training-data scaling of peak ankle torque with walking speed (deployed 0.85 Nm/kg per m/s, p = 0.001; training 0.96 Nm/kg per m/s, 95% CI [0.42, 1.50], p = 0.002), after excluding one outlier traced to atypical prosthesis loading. During ramp ascent, the controller scaled knee pre-flexion with grade (deployed 2.92 deg/deg, p = 0.027; training 3.30 deg/deg, 95% CI [1.83, 4.77], p < 0.001). During ramp descent, the controller increased resistive knee torque relative to level walking (deployed +0.16 Nm/kg, p < 0.001; training +0.16 Nm/kg, p = 0.008). Seamless stair transitions were generated for both intact- and prosthetic-side-leading sequences in ascent and descent, despite the training data containing only one limb-leading sequence. These results provide initial evidence towards end-to-end control that can provide unified, mode-adaptive prosthetic assistance without subject-specific tuning.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 2 minor

Summary. The paper trains Temporal Convolutional Networks on multi-terrain locomotion data from 18 transfemoral amputees to produce an end-to-end controller that maps onboard sensor inputs directly to continuous knee and ankle actuator commands. It reports real-time deployment on four participants (three able-bodied, one transfemoral amputee) across level walking, ramps, and stairs, claiming reproduction of training-data scalings (e.g., ankle torque vs. speed, knee pre-flexion vs. grade) and seamless stair transitions without subject-specific tuning or mode classification.

Significance. If the generalization claim holds, the approach could substantially reduce the clinical burden of impedance tuning and explicit mode switching in powered prostheses. The work provides concrete real-time deployment metrics and statistical comparisons to training data, which are strengths, but the small deployment cohort and population mismatch limit the strength of evidence for tuning-free assistance on new amputee users.

major comments (3)
  1. [Abstract / Deployment results] Abstract and deployment results: The central claim of evidence for 'unified, mode-adaptive prosthetic assistance without subject-specific tuning' on new amputee users rests on deployment data from only one transfemoral amputee (plus three able-bodied participants). Able-bodied users have intact neuromuscular control and different socket/loading mechanics, so their results do not directly test generalization to the target population; this is load-bearing for the tuning-free claim.
  2. [Abstract] Abstract: Post-hoc exclusion of one outlier participant in the level-walking torque scaling analysis (attributed to atypical prosthesis loading) is reported without a priori criteria or sensitivity analysis showing results with/without exclusion. This affects defensibility of the reported p=0.001 match to training data (0.85 vs. 0.96 Nm/kg per m/s).
  3. [Abstract / Results] Abstract / Results: No comparison is made to a subject-specifically tuned impedance controller (the conventional baseline) on the same deployment participants or tasks. Without this, it is not possible to quantify whether the end-to-end controller matches or exceeds performance while eliminating tuning.
minor comments (2)
  1. [Abstract] The abstract states 'seamless stair transitions were generated for both intact- and prosthetic-side-leading sequences' despite training data containing only one limb-leading sequence; clarify how this extrapolation was quantified (e.g., via transition timing or torque profiles) in the results section.
  2. [Results] Clarify the exact number of strides or trials per condition in the deployment cohort to allow assessment of statistical power for the reported p-values.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for their constructive comments on our manuscript. We address each major comment below, indicating planned revisions where appropriate.

read point-by-point responses

  1. Referee: [Abstract / Deployment results] Abstract and deployment results: The central claim of evidence for 'unified, mode-adaptive prosthetic assistance without subject-specific tuning' on new amputee users rests on deployment data from only one transfemoral amputee (plus three able-bodied participants). Able-bodied users have intact neuromuscular control and different socket/loading mechanics, so their results do not directly test generalization to the target population; this is load-bearing for the tuning-free claim.

    Authors: We acknowledge that the deployment cohort is small, with only one transfemoral amputee participant. The three able-bodied participants were included to demonstrate the controller's real-time performance and generalization across different users and loading conditions, but we agree they do not substitute for additional amputee data. The results from the single amputee participant do show reproduction of key scalings without tuning. We will revise the abstract and discussion to more clearly state the limitations of the current deployment sample size and emphasize that these results provide initial evidence rather than definitive proof of generalization to new amputee users. revision: partial

  2. Referee: [Abstract] Abstract: Post-hoc exclusion of one outlier participant in the level-walking torque scaling analysis (attributed to atypical prosthesis loading) is reported without a priori criteria or sensitivity analysis showing results with/without exclusion. This affects defensibility of the reported p=0.001 match to training data (0.85 vs. 0.96 Nm/kg per m/s).

    Authors: The exclusion was based on observed atypical prosthesis loading during the experiment, which was noted post-deployment. We agree that a priori criteria would strengthen the analysis. We will add a sensitivity analysis to the results section showing the scaling with and without the outlier, and revise the abstract to note the exclusion criterion more explicitly. revision: yes

  3. Referee: [Abstract / Results] Abstract / Results: No comparison is made to a subject-specifically tuned impedance controller (the conventional baseline) on the same deployment participants or tasks. Without this, it is not possible to quantify whether the end-to-end controller matches or exceeds performance while eliminating tuning.

    Authors: We did not include a direct comparison to a tuned impedance controller in this study, as the primary aim was to demonstrate that the end-to-end controller can reproduce key biomechanical scalings from the training data without any subject-specific tuning or mode classification. A full comparison would require additional experiments with expert-tuned baselines on the same tasks and participants. This is a valid point, and we will add discussion of this limitation in the manuscript, noting that future work should include such benchmarks. revision: partial

Circularity Check

0 steps flagged

No significant circularity; results are empirical deployment matches rather than derived reductions

full rationale

The paper trains Temporal Convolutional Networks on a dataset from 18 amputees and deploys the resulting controller on four new participants, reporting measured scalings (e.g., ankle torque vs. speed slopes of 0.85 vs. 0.96 Nm/kg per m/s) and transition behavior as empirical outcomes. No equations, self-citations, or ansatzes are invoked that reduce these reported quantities to the training inputs by construction; the central claim rests on experimental reproduction of learned patterns on held-out users rather than any self-definitional or fitted-input renaming step. This is standard supervised learning validation and remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 0 invented entities

The central claim rests on standard supervised learning assumptions plus the untested generalization from 18 training amputees to new users. No new physical entities are postulated. The network weights constitute fitted parameters but are not ad-hoc scalars chosen by hand.

free parameters (1)
  • TCN hyperparameters and weights
    Learned from the 18-person dataset; the claim depends on these parameters generalizing without retuning.
axioms (2)
  • domain assumption Sensor signals from the prosthesis are sufficient to infer continuous actuator commands across locomotion modes
    Invoked when the end-to-end mapping is trained and deployed without explicit intent classification.
  • domain assumption Training distribution from 18 amputees is representative for new users
    Required for the 'without subject-specific tuning' claim when deploying on four new participants.

pith-pipeline@v0.9.1-grok · 5878 in / 1550 out tokens · 16098 ms · 2026-06-27T21:24:00.980529+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

27 extracted references · 1 canonical work pages · 1 internal anchor

  1. [1]

    Physical capacity and walking ability after lower limb amputation: a systematic review,

    J. M. van Velzen, C. A. M. van Bennekom, W. Polomski, J. R. Slootman, L. H. V . van der Woude, and H. Houdijk, “Physical capacity and walking ability after lower limb amputation: a systematic review,”Clinical Rehabilitation, vol. 20, no. 11, pp. 999–1016, Nov. 2006

    2006

  2. [2]

    Self-selected walking speed in individuals with transfemoral amputation: recovery, economy and rehabilitation index,

    R. L. Bona, N. A. Gomenuka, J. L. L. Storniolo, A. Bonezi, and C. M. Biancardi, “Self-selected walking speed in individuals with transfemoral amputation: recovery, economy and rehabilitation index,”European Journal of Physiotherapy, vol. 22, no. 3, pp. 133–140, May 2020

    2020

  3. [3]

    Review of secondary physical conditions associated with lower-limb amputation and long-term prosthesis use,

    R. Gailey, “Review of secondary physical conditions associated with lower-limb amputation and long-term prosthesis use,”Journal of Reha- bilitation Research and Development, vol. 45, no. 1, pp. 15–30, Dec. 2008

    2008

  4. [4]

    Biomechanical evaluation of stair ambulation using impedance control on an active prosthesis,

    J. Camargo, K. Bhakta, K. Herrin, and A. Young, “Biomechanical evaluation of stair ambulation using impedance control on an active prosthesis,”Journal of Biomechanical Engineering, vol. 145, no. 2, p. 021007, Feb. 2023

    2023

  5. [5]

    Bionic ankle–foot prosthesis normalizes walking gait for persons with leg amputation,

    H. M. Herr and A. M. Grabowski, “Bionic ankle–foot prosthesis normalizes walking gait for persons with leg amputation,”Proceedings of the Royal Society B: Biological Sciences, vol. 279, no. 1728, pp. 457–464, Jul. 2011

    2011

  6. [6]

    Design and control of a pow- ered transfemoral prosthesis,

    F. Sup, A. Bohara, and M. Goldfarb, “Design and control of a pow- ered transfemoral prosthesis,”The International Journal of Robotics Research, vol. 27, no. 2, pp. 263–273, Feb. 2008

    2008

  7. [7]

    Continuous-context, user-independent, real-time intent recognition for powered lower-limb prostheses,

    K. Bhaktaet al., “Continuous-context, user-independent, real-time intent recognition for powered lower-limb prostheses,”Journal of Biomechan- ical Engineering, vol. 147, no. 2, p. 021009, Feb. 2025

    2025

  8. [8]

    Mode-unified intent estimation of a robotic prosthesis using deep learning,

    H. Kim, D. Lee, J. Y . Maldonado-Contreras, S. Zhou, K. R. Herrin, and A. J. Young, “Mode-unified intent estimation of a robotic prosthesis using deep learning,”IEEE Robotics and Automation Letters, vol. 10, no. 4, pp. 3206–3213, Apr. 2025

    2025

  9. [9]

    Configuring a Powered Knee and Ankle Prosthesis for Transfemoral Amputees within Five Specific Ambulation Modes,

    A. M. Simon, K. A. Ingraham, N. P. Fey, S. B. Finucane, R. D. Lipschutz, A. J. Young, and L. J. Hargrove, “Configuring a Powered Knee and Ankle Prosthesis for Transfemoral Amputees within Five Specific Ambulation Modes,” vol. 9, no. 6, p. e99387

  10. [10]

    Impedance control strategies for enhancing sloped and level walking ca- pabilities for individuals with transfemoral amputation using a powered 7 multi-joint prosthesis,

    K. Bhakta, J. Camargo, P. Kunapuli, L. Childers, and A. Young, “Impedance control strategies for enhancing sloped and level walking ca- pabilities for individuals with transfemoral amputation using a powered 7 multi-joint prosthesis,”Military Medicine, vol. 185, no. Supplement 1, pp. 490–499, Jan. 2020

    2020

  11. [11]

    Physical activity in indi- viduals with lower extremity amputations: a narrative review,

    M. Halsne, M. Czerniecki, and B. J. Hafner, “Physical activity in indi- viduals with lower extremity amputations: a narrative review,”Physical Therapy Reviews, 2017

    2017

  12. [12]

    Effects of locomotion mode recog- nition errors on volitional control of powered above-knee prostheses,

    F. Zhang, M. Liu, and H. Huang, “Effects of locomotion mode recog- nition errors on volitional control of powered above-knee prostheses,” IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 23, no. 1, pp. 64–72, Jan. 2015

    2015

  13. [13]

    Online rein- forcement learning control for the personalization of a robotic knee prosthesis,

    Y . Wen, J. Si, A. Brandt, X. Gao, and H. H. Huang, “Online rein- forcement learning control for the personalization of a robotic knee prosthesis,”IEEE Transactions on Cybernetics, vol. 50, no. 6, pp. 2346– 2356, Jun. 2020

    2020

  14. [14]

    Data-driven variable impedance control of a powered knee–ankle prosthesis for adaptive speed and incline walking,

    T. K. Best, C. G. Welker, E. J. Rouse, and R. D. Gregg, “Data-driven variable impedance control of a powered knee–ankle prosthesis for adaptive speed and incline walking,”IEEE Transactions on Robotics, vol. 39, no. 3, pp. 2151–2169, Jun. 2023

    2023

  15. [15]

    Unified control of a powered knee-ankle prosthesis enables walking, stairs, transitions, and other daily ambulation activities,

    L. M. Sullivan, M. Cowan, L. Gabert, and T. Lenzi, “Unified control of a powered knee-ankle prosthesis enables walking, stairs, transitions, and other daily ambulation activities,”IEEE Transactions on Neural Systems and Rehabilitation Engineering, vol. 33, pp. 3024–3039, 2025

    2025

  16. [16]

    Task-agnostic exoskeleton 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 control via biological joint moment estimation,”Nature, vol. 635, no. 8038, pp. 337–344, Nov. 2024

    2024

  17. [17]

    Seamless and intuitive control of a powered prosthetic leg using deep neural network for transfemoral amputees,

    M. Kim, A. M. Simon, and L. J. Hargrove, “Seamless and intuitive control of a powered prosthetic leg using deep neural network for transfemoral amputees,”Wearable Technologies, vol. 3, p. e24, 2022

    2022

  18. [18]

    A deep learning framework for end- to-end control of powered prostheses,

    C. P. O. Nuesslein and A. J. Young, “A deep learning framework for end- to-end control of powered prostheses,”IEEE Robotics and Automation Letters, vol. 9, no. 5, pp. 3988–3994, May 2024

    2024

  19. [19]

    Comparing the biomechanical response of users of an open-source powered knee and ankle prosthesis versus a passive prosthesis during ramp and stair ambulation,

    S. Zhou, S. Kestur, J. Maldonado, K. Herrin, N. Fey, and A. Young, “Comparing the biomechanical response of users of an open-source powered knee and ankle prosthesis versus a passive prosthesis during ramp and stair ambulation,”Journal of Biomechanics, vol. 186, p. 112732, Jun. 2025

    2025

  20. [20]

    Real-time adaptation of deep learning walking speed estimators enables biomimetic assistance modulation in an open-source bionic leg,

    J. Maldonado-Contreraset al., “Real-time adaptation of deep learning walking speed estimators enables biomimetic assistance modulation in an open-source bionic leg,”IEEE Transactions on Medical Robotics and Bionics, pp. 1–1, 2025

    2025

  21. [21]

    Design and clinical implementation of an open-source bionic leg,

    A. F. Azocar, L. M. Mooney, J.-F. Duval, A. M. Simon, L. J. Hargrove, and E. J. Rouse, “Design and clinical implementation of an open-source bionic leg,”Nature Biomedical Engineering, vol. 4, no. 10, pp. 941–953, Oct. 2020

    2020

  22. [22]

    A compensated open-loop impedance controller evaluated on the second-generation Open-Source Leg prosthesis,

    T. K. Best, G. C. Thomas, S. R. Ayyappan, R. D. Gregg, and E. J. Rouse, “A compensated open-loop impedance controller evaluated on the second-generation Open-Source Leg prosthesis,”IEEE/ASME Trans- actions on Mechatronics, pp. 1–13, 2024

    2024

  23. [23]

    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, Apr. 2018

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  24. [24]

    Prosthetic gait of unilateral transfemoral amputees: A kinematic study,

    S. M. H. J. Jaegers, J. H. Arendzen, and H. J. de Jongh, “Prosthetic gait of unilateral transfemoral amputees: A kinematic study,”Archives of Physical Medicine and Rehabilitation, vol. 76, no. 8, pp. 736–743, Aug. 1995

    1995

  25. [25]

    Axiomatic attribution for deep networks,

    M. Sundararajan, A. Taly, and Q. Yan, “Axiomatic attribution for deep networks,” inProceedings of the 34th International Conference on Machine Learning - Volume 70, ser. ICML’17, pp. 3319–3328

  26. [26]

    Ground reaction forces at different speeds of human walking and running,

    J. Nilsson and A. Thorstensson, “Ground reaction forces at different speeds of human walking and running,” vol. 136, no. 2, pp. 217–227

  27. [27]

    Scheduled sampling for sequence prediction with recurrent Neural networks,

    S. Bengio, O. Vinyals, N. Jaitly, and N. Shazeer, “Scheduled sampling for sequence prediction with recurrent Neural networks,” inProceedings of the 29th International Conference on Neural Information Processing Systems - Volume 1, ser. NIPS’15, vol. 1, pp. 1171–1179