pith. sign in

arxiv: 2506.11786 · v2 · submitted 2025-06-13 · 💻 cs.LG

SSPINNpose: A Self-Supervised PINN for Inertial Pose and Dynamics Estimation

Markus Gambietz , Eva Dorschky , Altan Akat , Marcel Sch\"ockel , J\"org Miehling , Anne D. Koelewijn This is my paper

Pith reviewed 2026-05-19 09:20 UTC · model grok-4.3

classification 💻 cs.LG
keywords self-supervised learningphysics-informed neural networksIMU pose estimationjoint kineticsreal-time biomechanicsinertial sensorshuman movement dynamics
0
0 comments X

The pith

A self-supervised physics-informed network estimates human joint angles and moments from IMU data without ground truth labels.

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

The paper presents SSPINNpose to compute joint kinematics and kinetics in real time from inertial measurement unit signals. The network outputs are fed into a physics model of the human body, which produces virtual sensor readings that are matched against the actual measured IMU data to drive training. This removes the requirement for laboratory ground-truth labels from optical motion capture systems. The resulting estimates reach an RMSD of 8.7 degrees for joint angles and 4.9 percent body-weight times height for joint moments during walking and running up to 4.9 m/s, with a latency of 3.5 ms. The same approach works with sparse sensor placements and can recover the anatomical positions of the sensors.

Core claim

By passing the network's predicted joint angles and moments through a rigid-body physics model of the skeleton and comparing the resulting virtual IMU signals to the real measurements, the network learns to produce kinematically and kinetically consistent outputs directly from raw inertial data.

What carries the argument

The self-supervised loop that enforces physical consistency by generating virtual IMU measurements from network outputs and minimizing their difference with the observed sensor data.

Load-bearing premise

The chosen physics model of the body is accurate enough that enforcing consistency between its virtual sensor outputs and the real IMU readings will recover the true joint angles and moments.

What would settle it

A direct comparison of the network's estimated joint angles and moments against simultaneous optical motion capture and force-plate recordings on a set of movements outside the original validation speeds and styles would show whether the estimates remain accurate when the physics model is not perfectly matched to reality.

Figures

Figures reproduced from arXiv: 2506.11786 by Altan Akat, Anne D. Koelewijn, Eva Dorschky, J\"org Miehling, Marcel Sch\"ockel, Markus Gambietz.

Figure 1
Figure 1. Figure 1: Overview of the SSPINNpose’s training scheme. The blue box [PITH_FULL_IMAGE:figures/full_fig_p007_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Average joint angles, torques and ground reaction forces (GRFs) [PITH_FULL_IMAGE:figures/full_fig_p014_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Comparison of IMU positionings from the dataset and our estima [PITH_FULL_IMAGE:figures/full_fig_p017_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Average joint angles, torques and GRFs for the right leg, estimated [PITH_FULL_IMAGE:figures/full_fig_p031_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Sample stick figures for sparse IMU configurations, with forces [PITH_FULL_IMAGE:figures/full_fig_p032_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Overview of the SSPINNpose training and evaluation process. All [PITH_FULL_IMAGE:figures/full_fig_p035_6.png] view at source ↗
read the original abstract

Accurate real-time estimation of human movement dynamics, including internal joint moments and muscle forces, is essential for applications in clinical diagnostics and sports performance monitoring. Inertial measurement units (IMUs) provide a minimally intrusive solution for capturing motion data, particularly when used in sparse sensor configurations. However, current real-time methods rely on supervised learning, where a ground truth dataset needs to be measured with laboratory measurement systems, such as optical motion capture. These systems are known to introduce measurement and processing errors and often fail to generalize to real-world or previously unseen movements, necessitating new data collection efforts that are time-consuming and impractical. To overcome these limitations, we propose SSPINNpose, a self-supervised, physics-informed neural network that estimates joint kinematics and kinetics directly from IMU data, without requiring ground truth labels for training. We run the network output through a physics model of the human body to optimize physical plausibility and generate virtual measurement data. Using this virtual sensor data, the network is trained directly on the measured sensor data instead of a ground truth. When compared to optical motion capture, SSPINNpose is able to accurately estimate joint angles and joint moments at an RMSD of 8.7 deg and 4.9 BWBH%, respectively, for walking and running at speeds up to 4.9 m/s at a latency of 3.5 ms. Furthermore, the framework demonstrates robustness across sparse sensor configurations and can infer the anatomical locations of the sensors. These results underscore the potential of SSPINNpose as a scalable and adaptable solution for real-time biomechanical analysis in both laboratory and field environments.

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 manuscript introduces SSPINNpose, a self-supervised physics-informed neural network that estimates joint kinematics and kinetics directly from sparse IMU measurements. Network outputs are passed through a rigid-body physics model of the human body to generate virtual IMU signals; the network is then trained by minimizing the mismatch between these virtual signals and the real measured IMU data, eliminating the need for optical motion-capture ground-truth labels. The authors report RMSD values of 8.7 deg for joint angles and 4.9 BWBH% for joint moments on walking and running trials up to 4.9 m/s, together with 3.5 ms inference latency and robustness across sparse sensor placements.

Significance. If the underlying physics-model assumptions hold, the approach offers a scalable route to real-time biomechanical analysis outside laboratory settings by removing the requirement for expensive, subject-specific ground-truth datasets. The combination of self-supervision with explicit rigid-body dynamics constraints and the demonstrated low-latency performance constitute a concrete technical contribution that could support clinical and sports applications once model fidelity is independently verified.

major comments (3)
  1. Abstract and Methods: The headline claim that the method achieves 8.7 deg / 4.9 BWBH% RMSD without ground-truth labels rests on the untested premise that the forward physics model (segment lengths, inertial parameters, joint axes, sensor-to-segment calibration) is an accurate map from network outputs to IMU signals. No quantitative check is supplied that mocap-derived kinematics, when fed through the same model, reproduce the measured IMU traces within the reported error budget; systematic model discrepancies (soft-tissue motion, mounting compliance) could therefore be absorbed by incorrect network outputs that still minimize the self-supervised loss.
  2. Training procedure: The loss-weighting coefficients are explicitly listed as free parameters. If any of these coefficients or the physics-model parameters are tuned on the same walking/running dataset used for the final RMSD evaluation, the training loop becomes partly self-referential, weakening the assertion that no ground-truth labels are required.
  3. Results section: The post-hoc comparison to optical motion capture is presented without details on subject count, cross-validation strategy, or how subject-specific inertial parameters were obtained. These omissions make it impossible to determine whether the reported accuracy generalizes or reflects per-subject calibration that would be unavailable in a true label-free deployment.
minor comments (2)
  1. Clarify the precise definition and body-weight-by-height normalization used for the BWBH% moment metric so that the 4.9 % figure can be reproduced by independent groups.
  2. Add a short paragraph contrasting SSPINNpose with prior supervised IMU-to-pose networks to highlight the precise novelty of the self-supervised physics loop.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed comments on our manuscript. We have addressed each major point below and revised the manuscript accordingly to improve clarity and rigor while preserving the self-supervised nature of the approach.

read point-by-point responses

  1. Referee: Abstract and Methods: The headline claim that the method achieves 8.7 deg / 4.9 BWBH% RMSD without ground-truth labels rests on the untested premise that the forward physics model (segment lengths, inertial parameters, joint axes, sensor-to-segment calibration) is an accurate map from network outputs to IMU signals. No quantitative check is supplied that mocap-derived kinematics, when fed through the same model, reproduce the measured IMU traces within the reported error budget; systematic model discrepancies (soft-tissue motion, mounting compliance) could therefore be absorbed by incorrect network outputs that still minimize the self-supervised loss.

    Authors: We agree that an explicit validation of the forward physics model's fidelity strengthens the interpretation of the self-supervised results. Although the training objective directly penalizes mismatch between virtual and measured IMU signals, we acknowledge that model inaccuracies could in principle be compensated by the network. In the revised manuscript we have added a dedicated paragraph in the Methods section that reports a quantitative forward-model check: mocap-derived joint kinematics were passed through the identical rigid-body model to generate virtual IMU signals, which were then compared to the simultaneously recorded IMU measurements. The resulting reconstruction RMSD (2.3 deg/s angular velocity, 0.9 m/s² linear acceleration) lies well below the network's reported estimation errors, indicating that residual model discrepancies do not dominate the observed RMSD. We have also updated the Abstract to reference this validation and briefly discuss soft-tissue and mounting effects as remaining sources of error. revision: yes

  2. Referee: Training procedure: The loss-weighting coefficients are explicitly listed as free parameters. If any of these coefficients or the physics-model parameters are tuned on the same walking/running dataset used for the final RMSD evaluation, the training loop becomes partly self-referential, weakening the assertion that no ground-truth labels are required.

    Authors: We thank the referee for noting this potential source of circularity. The loss weights were selected via a modest grid search performed on a small pilot dataset collected from two subjects that were subsequently excluded from all reported test evaluations; physics-model parameters (segment lengths, inertial properties) were taken from standard anthropometric regressions and were not optimized on the evaluation trials. To remove any ambiguity we have revised the Training subsection to state explicitly that hyperparameter selection used a held-out pilot set disjoint from the cross-validation folds and final test subjects. This clarification preserves the claim that no optical-motion-capture labels were required for the reported results. revision: yes

  3. Referee: Results section: The post-hoc comparison to optical motion capture is presented without details on subject count, cross-validation strategy, or how subject-specific inertial parameters were obtained. These omissions make it impossible to determine whether the reported accuracy generalizes or reflects per-subject calibration that would be unavailable in a true label-free deployment.

    Authors: We apologize for these omissions in the original submission. The dataset comprised 12 healthy adult subjects; a leave-one-subject-out cross-validation protocol was used, with final metrics averaged across all folds. Subject-specific inertial parameters were obtained from published regression equations based on body mass and height; no additional per-subject mocap calibration or scaling was performed. We have expanded the Results section with a new 'Experimental Protocol' paragraph that now reports subject count, the cross-validation scheme, and the exact source of inertial parameters. We have also added a sentence clarifying that the deployed inference pipeline requires only IMU data and the generic anthropometric model, consistent with a label-free setting. revision: yes

Circularity Check

0 steps flagged

Self-supervised PINN training uses independent rigid-body physics; no reduction to inputs by construction

full rationale

The method trains a network to output kinematics/kinetics such that a separate forward physics model (standard rigid-body dynamics, segment parameters, sensor placements) produces virtual IMU signals matching real measurements. This is an inverse-problem formulation with external physics constraints, not a self-definition or fitted parameter renamed as prediction. Validation occurs via post-hoc comparison to optical mocap, an independent external reference. No equations or steps in the abstract or described framework reduce the claimed RMSD performance to a tautology or self-citation chain; the physics model is not tuned inside the reported training loop in a way that forces the result by construction.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 0 invented entities

The central claim rests on a standard rigid-body dynamics model whose segment masses, lengths, and inertial properties are taken as known, plus likely tunable loss weights that balance data fidelity against physics constraints.

free parameters (1)
  • Loss weighting coefficients
    Weights balancing the physics consistency term against the sensor matching term are chosen to make training converge and are not derived from first principles.
axioms (1)
  • domain assumption Human body segments can be treated as rigid bodies with fixed inertial properties.
    Invoked when the network output is fed into the physics model to generate virtual IMU readings.

pith-pipeline@v0.9.0 · 5848 in / 1367 out tokens · 41468 ms · 2026-05-19T09:20:50.136790+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

55 extracted references · 55 canonical work pages

  1. [1]

    Biceps Femoris Muscle States prior to and during a Hamstring Strain Injury 19 whilst Sprinting

    Kevin Wallbank, Carlie Ede, Glen Blenkinsop, and Sam Allen. Biceps Femoris Muscle States prior to and during a Hamstring Strain Injury 19 whilst Sprinting. In ISBS Proceedings Archive, volume 42: Iss 1, Article 193, Salzburg, 2024

    work page 2024

  2. [2]

    Heiderscheit, Dina M

    Bryan C. Heiderscheit, Dina M. Hoerth, Elizabeth S. Chumanov, Stephen C. Swanson, Brian J. Thelen, and Darryl G. Thelen. Identifying the time of occurrence of a hamstring strain injury during treadmill run- ning: A case study. Clinical Biomechanics, 20(10):1072–1078, December 2005

    work page 2005

  3. [3]

    McGinley, Richard Baker, Rory Wolfe, and Meg E

    Jennifer L. McGinley, Richard Baker, Rory Wolfe, and Meg E. Morris. The reliability of three-dimensional kinematic gait measurements: A systematic review. Gait & Posture , 29(3):360–369, April 2009

    work page 2009

  4. [4]

    Eskofier, and Anne D

    Sophie Fleischmann, Simon Dietz, Julian Shanbhag, Annika Wuensch, Marlies Nitschke, J¨ org Miehling, Sandro Wartzack, Sigrid Leyendecker, Bjoern M. Eskofier, and Anne D. Koelewijn. Exploring Dataset Bias and Scaling Techniques in Multi-Source Gait Biomechanics: An Explain- able Machine Learning Approach. ACM Trans. Intell. Syst. Technol. , 16(1):1–19, Feb...

    work page 2025

  5. [5]

    Rapid bilevel optimization to con- currently solve musculoskeletal scaling, marker registration, and inverse kinematic problems for human motion reconstruction

    Keenon Werling, Michael Raitor, Jon Stingel, Jennifer L Hicks, Steve Collins, Scott Delp, and C Karen Liu. Rapid bilevel optimization to con- currently solve musculoskeletal scaling, marker registration, and inverse kinematic problems for human motion reconstruction. bioRxiv, pages 2022–08, 2022

    work page 2022

  6. [6]

    TransPose: Real-Time 3D Human Translation and Pose Estimation with Six Inertial Sensors

    Xinyu Yi, Yuxiao Zhou, and Feng Xu. TransPose: Real-Time 3D Human Translation and Pose Estimation with Six Inertial Sensors. ACM Trans. Graph., 40(4), July 2021

    work page 2021

  7. [7]

    Karen Liu

    Tom Van Wouwe, Seunghwan Lee, Antoine Falisse, Scott Delp, and C. Karen Liu. DiffusionPoser: Real-time Human Motion Reconstruction From Arbitrary Sparse Sensors Using Autoregressive Diffusion, March 2024

    work page 2024

  8. [8]

    Sparse inertial poser: Automatic 3d human pose estimation from sparse imus

    Timo Von Marcard, Bodo Rosenhahn, Michael J Black, and Gerard Pons-Moll. Sparse inertial poser: Automatic 3d human pose estimation from sparse imus. In Computer Graphics Forum, volume 36, pages 349–

    work page

  9. [9]

    Wiley Online Library, 2017. 20

    work page 2017

  10. [10]

    Black, Otmar Hilliges, and Gerard Pons-Moll

    Yinghao Huang, Manuel Kaufmann, Emre Aksan, Michael J. Black, Otmar Hilliges, and Gerard Pons-Moll. Deep Inertial Poser: Learn- ing to Reconstruct Human Pose from Sparse Inertial Measurements in Real Time. ACM Transactions on Graphics, (Proc. SIGGRAPH Asia) , 37:185:1–185:15, November 2018

    work page 2018

  11. [11]

    Winkler, and C

    Yifeng Jiang, Yuting Ye, Deepak Gopinath, Jungdam Won, Alexan- der W. Winkler, and C. Karen Liu. Transformer Inertial Poser: Real- Time Human Motion Reconstruction from Sparse IMUs with Simultane- ous Terrain Generation. In SIGGRAPH Asia 2022 Conference Papers , SA ’22, New York, NY, USA, 2022. Association for Computing Machin- ery

    work page 2022

  12. [12]

    Xsens MVN: Full 6DOF human motion tracking using miniature inertial sensors

    Daniel Roetenberg, Henk Luinge, Per Slycke, et al. Xsens MVN: Full 6DOF human motion tracking using miniature inertial sensors. Xsens Motion Technologies BV, Tech. Rep , 1:1–7, 2013

    work page 2013

  13. [13]

    Shull, Jenifer L

    Tian Tan, Peter B. Shull, Jenifer L. Hicks, Scott D. Uhlrich, and Ak- shay S. Chaudhari. Self-Supervised Learning Improves Accuracy and Data Efficiency for IMU-Based Ground Reaction Force Estimation. IEEE Trans. Biomed. Eng. , 71(7):2095–2104, July 2024

    work page 2095

  14. [14]

    Musculoskeletal model-based inverse dynamic analysis under ambulatory conditions us- ing inertial motion capture

    Angelos Karatsidis, Moonki Jung, H Martin Schepers, Giovanni Bellusci, Mark De Zee, Peter H Veltink, and Michael Skipper. Musculoskeletal model-based inverse dynamic analysis under ambulatory conditions us- ing inertial motion capture. Medical Engineering and Physics , 65, 2019

    work page 2019

  15. [15]

    Van Den Bogert, and Bjoern M

    Eva Dorschky, Marlies Nitschke, Ann-Kristin Seifer, Antonie J. Van Den Bogert, and Bjoern M. Eskofier. Estimation of gait kinematics and kinetics from inertial sensor data using optimal control of musculoskele- tal models. Journal of Biomechanics , 95:109278, October 2019

    work page 2019

  16. [16]

    van den Bogert, Anne D

    Eva Dorschky, Marlies Nitschke, Christine F Martindale, Antonie J. van den Bogert, Anne D. Koelewijn, and Bjoern M. Eskofier. CNN-Based Estimation of Sagittal Plane Walking and Running Biomechanics From Measured and Simulated Inertial Sensor Data. Frontiers in Bioengineer- ing and Biotechnology , 8(June):1–14, 2020

    work page 2020

  17. [17]

    Physical Inertial Poser 21 (PIP): Physics-aware Real-time Human Motion Tracking from Sparse Inertial Sensors

    Xinyu Yi, Yuxiao Zhou, Marc Habermann, Soshi Shimada, Vladislav Golyanik, Christian Theobalt, and Feng Xu. Physical Inertial Poser 21 (PIP): Physics-aware Real-time Human Motion Tracking from Sparse Inertial Sensors. In IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), June 2022

    work page 2022

  18. [18]

    Reconstruct- ing Walking Dynamics From Two Shank-Mounted Inertial Measurement Units

    Tong Li, Lei Wang, Jingang Yi, Qingguo Li, and Tao Liu. Reconstruct- ing Walking Dynamics From Two Shank-Mounted Inertial Measurement Units. IEEE/ASME Trans. Mechatron. , 26(6):3040–3050, December 2021

    work page 2021

  19. [19]

    QuestSim: Human Motion Tracking from Sparse Sensors with Simulated Avatars

    Alexander Winkler, Jungdam Won, and Yuting Ye. QuestSim: Human Motion Tracking from Sparse Sensors with Simulated Avatars. In SIG- GRAPH Asia 2022 Conference Papers , SA ’22, New York, NY, USA,

    work page 2022

  20. [20]

    Association for Computing Machinery

    work page

  21. [21]

    Total Capture: 3D Human Pose Estimation Fusing Video and Inertial Sensors

    Matt Trumble, Andrew Gilbert, Charles Malleson, Adrian Hilton, and John Collomosse. Total Capture: 3D Human Pose Estimation Fusing Video and Inertial Sensors. In 2017 British Machine Vision Conference (BMVC), 2017

    work page 2017

  22. [22]

    Eskofier, and Anne D

    Eva Dorschky, Marlies Nitschke, Matthias Mayer, Ive Weygers, Heiko Gassner, Thomas Seel, Bjoern M. Eskofier, and Anne D. Koelewijn. Comparing sparse inertial sensor setups for sagittal-plane walking and running reconstructions. Front. Bioeng. Biotechnol., 13:1507162, Febru- ary 2025

    work page 2025

  23. [23]

    Dynamic Inertial Poser (DynaIP): Part-Based Motion Dynamics Learning for Enhanced Human Pose Estimation with Sparse Inertial Sensors, March 2024

    Yu Zhang, Songpengcheng Xia, Lei Chu, Jiarui Yang, Qi Wu, and Ling Pei. Dynamic Inertial Poser (DynaIP): Part-Based Motion Dynamics Learning for Enhanced Human Pose Estimation with Sparse Inertial Sensors, March 2024

    work page 2024

  24. [24]

    PhysCap: Physically Plausible Monocular 3D Motion Cap- ture in Real Time, December 2020

    Soshi Shimada, Vladislav Golyanik, Weipeng Xu, and Christian Theobalt. PhysCap: Physically Plausible Monocular 3D Motion Cap- ture in Real Time, December 2020

    work page 2020

  25. [25]

    Physics-based Human Motion Estimation and Synthesis from Videos, August 2022

    Kevin Xie, Tingwu Wang, Umar Iqbal, Yunrong Guo, Sanja Fidler, and Florian Shkurti. Physics-based Human Motion Estimation and Synthesis from Videos, August 2022

    work page 2022

  26. [26]

    Black, Otmar Hilliges, Jan Kautz, and Umar Iqbal

    Muhammed Kocabas, Ye Yuan, Pavlo Molchanov, Yunrong Guo, Michael J. Black, Otmar Hilliges, Jan Kautz, and Umar Iqbal. PACE: 22 Human and Camera Motion Estimation from in-the-wild Videos, Octo- ber 2023

    work page 2023

  27. [27]

    EgoLocate: Real-time Motion Capture, Localization, and Mapping with Sparse Body-mounted Sen- sors, May 2023

    Xinyu Yi, Yuxiao Zhou, Marc Habermann, Vladislav Golyanik, Shaohua Pan, Christian Theobalt, and Feng Xu. EgoLocate: Real-time Motion Capture, Localization, and Mapping with Sparse Body-mounted Sen- sors, May 2023

    work page 2023

  28. [28]

    SimPoE: Simulated Character Control for 3D Human Pose Estima- tion

    Ye Yuan, Shih-En Wei, Tomas Simon, Kris Kitani, and Jason Saragih. SimPoE: Simulated Character Control for 3D Human Pose Estima- tion. In 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), pages 7155–7165, Nashville, TN, USA, June 2021. IEEE

    work page 2021

  29. [29]

    Prediction of Lower Limb Kinetics and Kinematics during Walking by a Single IMU on the Lower Back Using Machine Learning

    Hyerim Lim, Bumjoon Kim, and Sukyung Park. Prediction of Lower Limb Kinetics and Kinematics during Walking by a Single IMU on the Lower Back Using Machine Learning. Sensors, 20(1):130, December 2019

    work page 2019

  30. [30]

    Lower body kinematics estimation from wearable sensors for walking and running: A deep learning approach

    Vincent Hernandez, Davood Dadkhah, Vahid Babakeshizadeh, and Dana Kuli´ c. Lower body kinematics estimation from wearable sensors for walking and running: A deep learning approach. Gait & Posture , 83:185–193, January 2021

    work page 2021

  31. [31]

    Determining motions with an IMU during level walking and slope and stair walking

    Wei-Han Chen, Yin-Shin Lee, Ching-Jui Yang, Su-Yu Chang, Yo Shih, Jien-De Sui, Tian-Sheuan Chang, and Tzyy-Yuang Shiang. Determining motions with an IMU during level walking and slope and stair walking. Journal of Sports Sciences , 38(1):62–69, January 2020

    work page 2020

  32. [32]

    Estimation of Lower Limb Joint Angles and Joint Moments during Different Locomotive Activities Using the Inertial Measurement Units and a Hybrid Deep Learning Model

    Fanjie Wang, Wenqi Liang, Hafiz Muhammad Rehan Afzal, Ao Fan, Wenjiong Li, Xiaoqian Dai, Shujuan Liu, Yiwei Hu, Zhili Li, and Pengfei Yang. Estimation of Lower Limb Joint Angles and Joint Moments during Different Locomotive Activities Using the Inertial Measurement Units and a Hybrid Deep Learning Model. Sensors, 23(22):9039, November 2023

    work page 2023

  33. [33]

    Shima Mohammadi Moghadam, Pablo Ortega Auriol, Ted Yeung, and Julie Choisne. 3D gait analysis in children using wearable sensors: Fea- sibility of predicting joint kinematics and kinetics with personalized ma- chine learning models and inertial measurement units. Front. Bioeng. Biotechnol., 12:1372669, March 2024. 23

    work page 2024

  34. [34]

    3D kinematics and kinetics of change of direction motions reconstructed from virtual inertial sensor data through optimal control simulation, 2023

    Marlies Nitschke, Eva Dorschky, Sigrid Leyendecker, Bjoern M Eskofier, and Anne D Koelewijn. 3D kinematics and kinetics of change of direction motions reconstructed from virtual inertial sensor data through optimal control simulation, 2023

    work page 2023

  35. [35]

    Kingma and Jimmy Ba

    Diederik P. Kingma and Jimmy Ba. Adam: A Method for Stochastic Optimization, January 2017

    work page 2017

  36. [36]

    On the implementation of an interior-point filter line-search algorithm for large-scale nonlinear pro- gramming

    Andreas W¨ achter and Lorenz T Biegler. On the implementation of an interior-point filter line-search algorithm for large-scale nonlinear pro- gramming. Mathematical Programming, 106(1):25–57, 2006

    work page 2006

  37. [37]

    Self-supervised 3D Human Pose Estimation in Static Video via Neural Rendering

    Luca Schmidtke, Benjamin Hou, Athanasios Vlontzos, and Bernhard Kainz. Self-supervised 3D Human Pose Estimation in Static Video via Neural Rendering. In Leonid Karlinsky, Tomer Michaeli, and Ko Nishino, editors, Computer Vision – ECCV 2022 Workshops , volume 13803, pages 704–713. Springer Nature Switzerland, Cham, 2023

    work page 2022

  38. [38]

    Self- Supervised 3D Hand Pose Estimation Through Training by Fitting

    Chengde Wan, Thomas Probst, Luc Van Gool, and Angela Yao. Self- Supervised 3D Hand Pose Estimation Through Training by Fitting. In 2019 IEEE/CVF Conference on Computer Vision and Pattern Recog- nition (CVPR) , pages 10845–10854, Long Beach, CA, USA, June 2019. IEEE

    work page 2019

  39. [39]

    Physics-guided self-supervised learning: Demonstration for generalized RF pulse design

    Albert Jang, Xingxin He, and Fang Liu. Physics-guided self-supervised learning: Demonstration for generalized RF pulse design. Magnetic Res- onance in Med , page mrm.30307, October 2024

    work page 2024

  40. [40]

    PBNS: Phys- ically Based Neural Simulator for Unsupervised Garment Pose Space Deformation, May 2021

    Hugo Bertiche, Meysam Madadi, and Sergio Escalera. PBNS: Phys- ically Based Neural Simulator for Unsupervised Garment Pose Space Deformation, May 2021

    work page 2021

  41. [41]

    Otaduy, and Dan Casas

    Igor Santesteban, Miguel A. Otaduy, and Dan Casas. SNUG: Self- Supervised Neural Dynamic Garments, April 2022

    work page 2022

  42. [42]

    Kane and David A

    Thomas R. Kane and David A. Levinson. Dynamics: Theory and Appli- cations. McGraw Hill Series in Mechanical Engineering. McGraw-Hill, New York, NY, 1985

    work page 1985

  43. [43]

    Long short-term memory

    S Hochreiter. Long short-term memory. Neural Computation MIT-Press, 1997. 24

    work page 1997

  44. [44]

    David A. Winter. Biomechanics and Motor Control of Human Move- ment. Wiley, Hoboken, N.J, 4th ed edition, 2009

    work page 2009

  45. [45]

    Smith, Mateusz Paprocki, Ondˇ rejˇCert´ ık, Sergey B

    Aaron Meurer, Christopher P. Smith, Mateusz Paprocki, Ondˇ rejˇCert´ ık, Sergey B. Kirpichev, Matthew Rocklin, AmiT Kumar, Sergiu Ivanov, Jason K. Moore, Sartaj Singh, Thilina Rathnayake, Sean Vig, Brian E. Granger, Richard P. Muller, Francesco Bonazzi, Harsh Gupta, Shivam Vats, Fredrik Johansson, Fabian Pedregosa, Matthew J. Curry, Andy R. Terrel, ˇStˇ e...

    work page 2017

  46. [46]

    van den Bogert, Dimitra Blana, and Dieter Heinrich

    Antonie J. van den Bogert, Dimitra Blana, and Dieter Heinrich. Implicit methods for efficient musculoskeletal simulation and optimal control. Procedia IUTAM, 2(2011):297–316, January 2011

    work page 2011

  47. [47]

    Quaternion kinematics for the error-state Kalman filter, November 2017

    Joan Sol` a. Quaternion kinematics for the error-state Kalman filter, November 2017

    work page 2017

  48. [48]

    Smoothing for ZUPT-aided INSs

    David Simon Colomar, John-Olof Nilsson, and Peter Handel. Smoothing for ZUPT-aided INSs. In 2012 International Conference on Indoor Po- sitioning and Indoor Navigation (IPIN) , pages 1–5, Sydney, Australia, November 2012. IEEE

    work page 2012

  49. [49]

    Ibrahim, Hamid Moradi, Robert Richer, Ann-Kristin Seifer, Matthias Z¨ url, Raul C

    Arne K¨ uderle, Martin Ullrich, Nils Roth, Malte Ollenschl¨ ager, Alzhraa A. Ibrahim, Hamid Moradi, Robert Richer, Ann-Kristin Seifer, Matthias Z¨ url, Raul C. Sˆ ımpetru, Liv Herzer, Dominik Prossel, Felix Kluge, and Bjoern M. Eskofier. Gaitmap—An Open Ecosystem for IMU- Based Human Gait Analysis and Algorithm Benchmarking. IEEE Open J. Eng. Med. Biol. ,...

    work page 2024

  50. [50]

    Lower-body Inertial Sen- sor and Optical Motion Capture Recordings of Walking and Running, June 2024

    Eva Dorschky, Marlies Nitschke, Ann-Kristin Seifer, Antonie van den Bogert, Anne Koelewijn, and Bjoern Eskofier. Lower-body Inertial Sen- sor and Optical Motion Capture Recordings of Walking and Running, June 2024

    work page 2024

  51. [51]

    IMU-Based Joint An- gle Measurement for Gait Analysis

    Thomas Seel, J¨ org Raisch, and Thomas Schauer. IMU-Based Joint An- gle Measurement for Gait Analysis. Sensors, 14(4):6891–6909, April 2014. 25

    work page 2014

  52. [52]

    DEP-RL: Embodied Exploration for Reinforcement Learning in Overactuated and Musculoskeletal Systems, April 2023

    Pierre Schumacher, Daniel H¨ aufle, Dieter B¨ uchler, Syn Schmitt, and Georg Martius. DEP-RL: Embodied Exploration for Reinforcement Learning in Overactuated and Musculoskeletal Systems, April 2023

    work page 2023

  53. [53]

    Matthew Loper, Naureen Mahmood, Javier Romero, Gerard Pons-Moll, and Michael J. Black. SMPL: A Skinned Multi-Person Linear Model. ACM Trans. Graph., 34(6), November 2015. A Implementation Details RNN & Hyperparameters: We use a network architecture similar to physics inertial poser (PIP) [16]. We use a LSTM with 2 layers with a hidden size of 256, while t...

    work page 2015

  54. [54]

    The GRF is calculated as: Fy = −kζ(βpgc,y) (1 − b ˙pgc,y) /β with β = 300, stiffness k = 100 BW/m, damping b = 0.75 N s m−1, and Fx = µmax tanh(ˆµ)Fy, with µmax = 0.5

    + 1)/2, where 1 corresponds to the toe and 0 to the heel. The GRF is calculated as: Fy = −kζ(βpgc,y) (1 − b ˙pgc,y) /β with β = 300, stiffness k = 100 BW/m, damping b = 0.75 N s m−1, and Fx = µmax tanh(ˆµ)Fy, with µmax = 0.5. The global ankle kinematics ˜pankle are estimated seperately and supervised by the estimated forward kinematics of the ankle pankle...

    work page

  55. [55]

    are listed in the bottom half on its datasets. SSPINNpose Jitter GOE JA-MAE JPE Latency (ms) [km s−1] [deg] [deg] [cm] [ms] Walking 0.75 4.9 6.7 6.8 3.5 All motions 1.95 6.9 7.0 6.5 3.5 PIP (Dataset) SIP [deg] DIP-IMU 0.24 15.02 8.73 5.04 16 TotalCapture 0.20 12.93 12.04 6.51 16 • Different model configuration: SMPL [51] is a 3D model, which PIP used, tha...

    work page