pith. sign in

arxiv: 2602.14092 · v2 · pith:HIXENO4Nnew · submitted 2026-02-15 · 📡 eess.SY · cs.RO· cs.SY

Simultaneous State Estimation and Online Model Learning in a Soft Robotic System

Jan-Hendrik Ewering , Max Bartholdt , Simon F. G. Ehlers , Niklas Wahlstr\"om , Thomas B. Sch\"on , Thomas Seel This is my paper

Pith reviewed 2026-05-21 13:04 UTC · model grok-4.3

classification 📡 eess.SY cs.ROcs.SY
keywords soft roboticsstate estimationonline model learningGaussian processparticle filterbending stiffnessconstant curvaturebase reactions
0
0 comments X

The pith

A particle filter learns a soft robot's bending stiffness as a Gaussian process while estimating its pose from base reactions alone.

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

The paper shows how a marginalized particle filter can track a soft robot's current shape while simultaneously learning a Gaussian process model of its bending stiffness. It starts from a simple constant-curvature kinematic model and updates both the pose estimate and the stiffness function using only sequential measurements of forces and torques at the robot base. This matters because soft robots rarely match any fixed model in advance, so the system must improve its internal predictions during normal operation. The learned stiffness model reduces error when the robot is simulated several steps ahead, unlike a simpler random-walk treatment of the same parameters.

Core claim

By embedding a Gaussian process prior over bending stiffness into a marginalized particle filter that uses nominal constant-curvature kinematics, the filter jointly tracks the robot configuration and refines the stiffness model from sequential base-reaction data. On a physical prototype this produces accurate pose estimates together with a stiffness function that reduces multi-step forward-prediction error relative to a random-walk baseline.

What carries the argument

Marginalized particle filter that couples nominal constant-curvature kinematic equations to a Gaussian process over bending stiffness parameters.

If this is right

  • Stiffness can be predicted for previously unseen configurations instead of merely tracked as a random variable.
  • Multi-step forward simulation error decreases once the learned Gaussian process is substituted for the random-walk model.
  • The same base-reaction measurements suffice for both accurate pose tracking and progressive model improvement.
  • Predictive controllers can operate on an internal model that continues to refine itself without requiring additional sensors or offline calibration.

Where Pith is reading between the lines

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

  • The same filter structure could be applied to other continuum or soft robots whose nominal kinematics are simple but whose material parameters change with load or temperature.
  • If base reactions are already sensed for low-level control, the method adds online adaptation at no extra hardware cost.
  • Long-term operation could allow the stiffness map to reveal slow changes such as material fatigue without any dedicated identification experiments.

Load-bearing premise

The nominal constant-curvature equations supply enough structure for the particle filter to separate pose estimation from learning the Gaussian process stiffness parameters using only base reaction measurements.

What would settle it

Running the identical robot experiment but replacing the learned Gaussian process with a random walk over stiffness values and finding that multi-step forward prediction error does not decrease would falsify the claim that the learned model improves overall quality.

Figures

Figures reproduced from arXiv: 2602.14092 by Jan-Hendrik Ewering, Max Bartholdt, Niklas Wahlstr\"om, Simon F. G. Ehlers, Thomas B. Sch\"on, Thomas Seel.

Figure 1
Figure 1. Figure 1: Nonlinear soft robotic system with the unknown states [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: True and estimated hidden states of the soft robot, which is driven [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Mean and variance of the learned nonlinear bending stiffness [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
read the original abstract

Operating complex real-world systems, such as soft robots, can benefit from precise predictive control schemes that require accurate state and model knowledge. This knowledge is typically not available in practical settings and must be inferred from noisy measurements. In particular, it is challenging to simultaneously estimate unknown states and learn a model online from sequentially arriving measurements. In this paper, we show how a recently proposed gray-box system identification tool enables the estimation of a soft robot's current pose while at the same time learning a bending stiffness model. For estimation and learning, we only need a nominal constant-curvature robot model and measurements of the robot's base reactions (e.g., base forces). The estimation scheme -- relying on a marginalized particle filter -- allows us to conveniently interface nominal constant-curvature equations with a Gaussian Process (GP) bending stiffness model to be learned. This, in contrast to estimation via a random walk over stiffness values, enables prediction of bending stiffness and improves overall model quality. We demonstrate, using a real-world soft robot, that the method learns a bending-stiffness model online while accurately estimating the robot's pose. Notably, reduced error in multi-step forward predictions indicates that the learned bending-stiffness GP improves overall model quality.

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

2 major / 2 minor

Summary. The manuscript presents a gray-box system identification method for soft robots that uses a marginalized particle filter to jointly estimate the robot's pose and learn a Gaussian Process model of bending stiffness online. The method interfaces a nominal constant-curvature kinematic model with the GP and relies solely on base reaction force/torque measurements. Real-robot experiments are reported to show accurate pose estimation and improved multi-step prediction performance over a random-walk stiffness baseline.

Significance. If the central claims hold, the work would be significant for online adaptive modeling in soft robotics, where accurate predictive models are crucial for control but difficult to obtain. The combination of particle filtering with nonparametric GP learning in a real-world setting is a notable contribution. The real-robot validation and focus on multi-step prediction quality are positive aspects that could influence practical applications in the field.

major comments (2)
  1. [§4 (Experiments)] §4 (Experiments): The abstract and results claim reduced error in multi-step forward predictions indicating improved model quality, but no specific quantitative metrics (e.g., RMSE values, percentage improvements), error bars, or statistical significance tests are provided in the summary of results. This makes it hard to evaluate the practical impact of the learned GP over the baseline.
  2. [Method section (MPF and CC interface)] Method section (MPF and CC interface): The central claim depends on the nominal constant-curvature equations providing sufficient structure for the marginalized particle filter to disentangle pose estimation from learning the GP bending stiffness parameters using only base reaction measurements. If the real deformation deviates from constant curvature, residuals might be misattributed to the GP in a non-predictive way. A sensitivity analysis or comparison with a more detailed finite-element model would strengthen this.
minor comments (2)
  1. [Abstract] Abstract: The abstract mentions 'improved overall model quality' but could specify the baseline more clearly (random walk over stiffness values).
  2. [Notation] Notation: Ensure consistent notation for the GP hyperparameters and the marginalized particle filter weights throughout the paper.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for their constructive feedback and positive evaluation of our manuscript. We address each of the major comments below, indicating where revisions will be made to strengthen the paper.

read point-by-point responses

  1. Referee: [§4 (Experiments)] §4 (Experiments): The abstract and results claim reduced error in multi-step forward predictions indicating improved model quality, but no specific quantitative metrics (e.g., RMSE values, percentage improvements), error bars, or statistical significance tests are provided in the summary of results. This makes it hard to evaluate the practical impact of the learned GP over the baseline.

    Authors: We agree that including explicit quantitative metrics would enhance the clarity and impact of our results. Although the manuscript includes figures showing the prediction errors, we will revise the text in Section 4 to report specific RMSE values for the multi-step predictions, percentage improvements over the baseline, error bars representing variability across trials, and statistical significance tests (e.g., paired t-tests) to quantify the improvement of the learned GP model. revision: yes

  2. Referee: [Method section (MPF and CC interface)] Method section (MPF and CC interface): The central claim depends on the nominal constant-curvature equations providing sufficient structure for the marginalized particle filter to disentangle pose estimation from learning the GP bending stiffness parameters using only base reaction measurements. If the real deformation deviates from constant curvature, residuals might be misattributed to the GP in a non-predictive way. A sensitivity analysis or comparison with a more detailed finite-element model would strengthen this.

    Authors: This is a valid concern regarding the modeling assumptions. The constant-curvature model provides the nominal kinematic structure, while the GP learns the bending stiffness as a function of the state to account for deviations. The fact that our approach yields improved multi-step prediction performance on the real robot indicates that the learned model generalizes beyond mere residual fitting. To address the referee's suggestion, we will add a sensitivity analysis in the revised manuscript by testing the method under perturbed curvature assumptions or with varying levels of model mismatch. A detailed comparison with finite-element models is outside the scope of this work but represents an interesting direction for future research. revision: partial

Circularity Check

0 steps flagged

No significant circularity; gray-box MPF interface remains independent of fitted outputs

full rationale

The derivation relies on a nominal constant-curvature kinematic model interfaced with a marginalized particle filter that jointly estimates pose and learns GP parameters for bending stiffness from base reaction measurements. The reported reduction in multi-step forward prediction error is presented as an empirical outcome of the learned GP versus a random-walk baseline, without evidence that this improvement is forced by construction from the identical data used in the fit. No self-definitional steps, fitted inputs renamed as predictions, or load-bearing self-citations appear in the abstract or described method. The central claim retains independent content from the filter's ability to marginalize and predict stiffness values.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Abstract-only view limits visibility; the approach implicitly relies on standard particle-filter assumptions and the adequacy of the constant-curvature nominal model as a structural prior for the GP.

axioms (1)
  • domain assumption Base reaction measurements contain sufficient information to jointly infer pose and bending stiffness parameters
    Invoked when stating that only base forces are needed for both estimation and learning

pith-pipeline@v0.9.0 · 5777 in / 1251 out tokens · 30091 ms · 2026-05-21T13:04:47.749446+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

35 extracted references · 35 canonical work pages · 1 internal anchor

  1. [1]

    Adaptive Model-Predictive Control of a Soft Continuum Robot Using a Physics-Informed Neural Network Based on Cosserat Rod Theory,

    J. Licher, M. Bartholdt, H. Krauss, T.-L. Habich, T. Seel, and M. Schappler, “Adaptive Model-Predictive Control of a Soft Continuum Robot Using a Physics-Informed Neural Network Based on Cosserat Rod Theory,”preprint. arXiv: 2508.12681, 2025

    work page arXiv 2025

  2. [2]

    A Non-Linear Model Predictive Task-Space Controller Satisfying Shape Constraints for Tendon-Driven Continuum Robots,

    M. Hachen, C. Shentu, S. Lilge, and J. Burgner-Kahrs, “A Non-Linear Model Predictive Task-Space Controller Satisfying Shape Constraints for Tendon-Driven Continuum Robots,”IEEE Robotics and Automa- tion Letters, vol. 10, no. 3, pp. 2438–2445, 2025

    work page 2025

  3. [3]

    Friction- adaptive stochastic nonlinear model predictive control for autonomous vehicles,

    S. Vaskov, R. Quirynen, M. Menner, and K. Berntorp, “Friction- adaptive stochastic nonlinear model predictive control for autonomous vehicles,”Vehicle System Dynamics, pp. 1–25, 2023

    work page 2023

  4. [4]

    S ¨arkk¨a and L

    S. S ¨arkk¨a and L. Svensson,Bayesian filtering and smoothing, 2nd ed. New York: Cambridge University Press, 2023

    work page 2023

  5. [5]

    Adaptive State Estimation with Constant-Curvature Dynamics Using Force-Torque Sensors with Application to a Soft Pneumatic Actuator,

    M. Mehl, M. Bartholdt, S. F. G. Ehlers, T. Seel, and M. Schappler, “Adaptive State Estimation with Constant-Curvature Dynamics Using Force-Torque Sensors with Application to a Soft Pneumatic Actuator,” inInt. Conference on Robotics and Autom., IEEE, 2024, pp. 14 939– 14 945

    work page 2024

  6. [6]

    A Sliding- Window Filter for Online Continuous-Time Continuum Robot State Estimation,

    S. Teetaert, S. Lilge, J. Burgner-Kahrs, and T. D. Barfoot, “A Sliding- Window Filter for Online Continuous-Time Continuum Robot State Estimation,”preprint. arXiv: 2510.26623, 2025

    work page internal anchor Pith review arXiv 2025

  7. [7]

    Model-Based Control of Soft Robots: A Survey of the State of the Art and Open Challenges,

    C. Della Santina, C. Duriez, and D. Rus, “Model-Based Control of Soft Robots: A Survey of the State of the Art and Open Challenges,” IEEE Control Systems, vol. 43, no. 3, pp. 30–65, 2023

    work page 2023

  8. [8]

    A Stochas- tic Framework for Continuous-Time State Estimation of Continuum Robots,

    S. Teetaert, S. Lilge, J. Burgner-Kahrs, and T. D. Barfoot, “A Stochas- tic Framework for Continuous-Time State Estimation of Continuum Robots,”preprint. arXiv: 2510.01381, 2025

    work page arXiv 2025

  9. [9]

    Bayesian Inference and Learning in Nonlinear Dynamical Systems: A Framework for Incorporating Explicit and Implicit Prior Knowledge,

    B. V olkmann, J.-H. Ewering, M. Meindl, S. F. G. Ehlers, M. A. M ¨uller, and T. Seel, “Bayesian Inference and Learning in Nonlinear Dynamical Systems: A Framework for Incorporating Explicit and Implicit Prior Knowledge,”preprint, arXiv: 2508.15345, 2025

    work page arXiv 2025

  10. [10]

    Nonlinear System Identification: Learning While Respecting Physical Models Using a Sequential Monte Carlo Method,

    A. Wigren, J. W ˚agberg, F. Lindsten, A. G. Wills, and T. B. Sch ¨on, “Nonlinear System Identification: Learning While Respecting Physical Models Using a Sequential Monte Carlo Method,”IEEE Control Systems, vol. 42, no. 1, pp. 75–102, 2022

    work page 2022

  11. [11]

    State-Space Inference and Learning with Gaussian Processes,

    R. Turner, M. Deisenroth, and C. Rasmussen, “State-Space Inference and Learning with Gaussian Processes,” inInt. Conf. on Artificial Intell. and Statistics, vol. 9, Sardinia, Italy: PMLR, 2010, pp. 868–875

    work page 2010

  12. [12]

    Bayesian Inference and Learning in Gaussian Process State-Space Models with Particle MCMC,

    R. Frigola, F. Lindsten, T. B. Sch ¨on, and C. E. Rasmussen, “Bayesian Inference and Learning in Gaussian Process State-Space Models with Particle MCMC,” inAdv. Neur. Inform. Proc. Syst., vol. 26, Curran, 2013

    work page 2013

  13. [13]

    Computationally Efficient Bayesian Learning of Gaussian Process State Space Models,

    A. Svensson, A. Solin, S. S ¨arkk¨a, and T. B. Sch ¨on, “Computationally Efficient Bayesian Learning of Gaussian Process State Space Models,” inInt. Conf. on AI and Statistics, vol. 51, Cadiz, Spain: PMLR, 2016, pp. 213–221

    work page 2016

  14. [14]

    A flexible state–space model for learn- ing nonlinear dynamical systems,

    A. Svensson and T. B. Sch ¨on, “A flexible state–space model for learn- ing nonlinear dynamical systems,”Automatica, vol. 80, pp. 189–199, 2017

    work page 2017

  15. [15]

    Learning Dynamics from Input-Output Data with Hamil- tonian Gaussian Processes,

    J.-H. Ewering, R. E. Herrmann, N. Wahlstr ¨om, T. B. Sch ¨on, and T. Seel, “Learning Dynamics from Input-Output Data with Hamil- tonian Gaussian Processes,” inLearning for Dynamics and Control Conference, 2026

    work page 2026

  16. [16]

    Online Bayesian inference and learning of Gaussian- process state–space models,

    K. Berntorp, “Online Bayesian inference and learning of Gaussian- process state–space models,”Automatica, vol. 129, p. 109 613, 2021

    work page 2021

  17. [17]

    Online Constrained Bayesian Inference and Learning of Gaussian-Process State-Space Models,

    K. Berntorp and M. Menner, “Online Constrained Bayesian Inference and Learning of Gaussian-Process State-Space Models,” inAmerican Control Conf., IEEE, 2022, pp. 940–945

    work page 2022

  18. [18]

    Efficient Online Inference and Learning in Partially Known Nonlinear State-Space Models by Learning Expressive Degrees of Freedom Of- fline,

    J.-H. Ewering, B. V olkmann, S. F. G. Ehlers, T. Seel, and M. Meindl, “Efficient Online Inference and Learning in Partially Known Nonlinear State-Space Models by Learning Expressive Degrees of Freedom Of- fline,” inConf. on Decision and Control, IEEE, 2024, pp. 4157–4164

    work page 2024

  19. [19]

    Online Joint State Inference and Learning of Partially Unknown State-Space Models,

    A. Kullberg, I. Skog, and G. Hendeby, “Online Joint State Inference and Learning of Partially Unknown State-Space Models,”IEEE Trans. Signal Process., vol. 69, pp. 4149–4161, 2021

    work page 2021

  20. [20]

    Approximating a Laplacian Prior for Joint State and Model Estimation within an UKF,

    R.-S. G ¨otte and J. Timmermann, “Approximating a Laplacian Prior for Joint State and Model Estimation within an UKF,”IFAC- PapersOnLine, vol. 56, no. 2, pp. 869–874, 2023

    work page 2023

  21. [21]

    Continuum robot state estimation using Gaussian process regression on SE(3),

    S. Lilge, T. D. Barfoot, and J. Burgner-Kahrs, “Continuum robot state estimation using Gaussian process regression on SE(3),”Int. Journal of Robotics Research, vol. 41, no. 13-14, pp. 1099–1120, 2022

    work page 2022

  22. [22]

    Real-time pose estimation and obstacle avoidance for multi-segment continuum manipulator in dynamic environments,

    A. Ataka et al., “Real-time pose estimation and obstacle avoidance for multi-segment continuum manipulator in dynamic environments,” in Int. Conf. on Intel. Robots and Systems, IEEE, 2016, pp. 2827–2832

    work page 2016

  23. [23]

    Non-linear System Identification and State Estimation in a Pneumatic Based Soft Continuum Robot,

    J. Y . Loo, K. C. Kong, C. P. Tan, and S. G. Nurzaman, “Non-linear System Identification and State Estimation in a Pneumatic Based Soft Continuum Robot,” inConf. on Control Techn. and Applications, IEEE, 2019, pp. 39–46

    work page 2019

  24. [24]

    Continuous Shape Estimation of Continuum Robots Using X-ray Images,

    E. J. Lobaton, J. Fu, L. G. Torres, and R. Alterovitz, “Continuous Shape Estimation of Continuum Robots Using X-ray Images,” inInt. Conf. on Robotics and Autom., IEEE, 2013, pp. 725–732

    work page 2013

  25. [25]

    Stochastic Adaptive Estimation in Polynomial Curvature Shape State Space for Continuum Robots,

    G. Zhang and L. Wang, “Stochastic Adaptive Estimation in Polynomial Curvature Shape State Space for Continuum Robots,”IEEE Transac- tions on Robotics, vol. 42, pp. 261–280, 2026

    work page 2026

  26. [26]

    State Estimation of Continuum Robots: A Nonlinear Constrained Moving Horizon Approach,

    H. Abdelaziz, A. Nada, H. Ishii, and H. El-Hussieny, “State Estimation of Continuum Robots: A Nonlinear Constrained Moving Horizon Approach,”preprint. arXiv: 2308.03931, 2023

    work page arXiv 2023

  27. [27]

    Probabilistic Modeling and Bayesian Filtering for Improved State Estimation for Soft Robots,

    D. Kim, M. Park, and Y .-L. Park, “Probabilistic Modeling and Bayesian Filtering for Improved State Estimation for Soft Robots,”IEEE Trans- actions on Robotics, vol. 37, no. 5, pp. 1728–1741, 2021

    work page 2021

  28. [28]

    Dynamic Modeling of Soft-Material Actuators Combining Constant Curvature Kinematics and Floating-Base Approach,

    M. Mehl, M. Bartholdt, and M. Schappler, “Dynamic Modeling of Soft-Material Actuators Combining Constant Curvature Kinematics and Floating-Base Approach,” inInt. Conf. on Soft Robotics, IEEE, 2022, pp. 1–8

    work page 2022

  29. [29]

    On an Improved State Parametrization for Soft Robots With Piecewise Constant Curvature and Its Use in Model Based Control,

    C. Della Santina, A. Bicchi, and D. Rus, “On an Improved State Parametrization for Soft Robots With Piecewise Constant Curvature and Its Use in Model Based Control,”IEEE Robotics and Automation Letters, vol. 5, no. 2, pp. 1001–1008, 2020

    work page 2020

  30. [30]

    C. E. Rasmussen and C. Williams,Gaussian Processes for Machine Learning. MIT Press, 2005

    work page 2005

  31. [31]

    Hilbert space methods for reduced-rank Gaussian process regression,

    A. Solin and S. S ¨arkk¨a, “Hilbert space methods for reduced-rank Gaussian process regression,”Statistics and Computing, vol. 30, no. 2, pp. 419–446, 2020

    work page 2020

  32. [32]

    Practical Hilbert space approximate Bayesian Gaussian processes for probabilistic programming,

    G. Riutort-Mayol, P.-C. B ¨urkner, M. R. Andersen, A. Solin, and A. Vehtari, “Practical Hilbert space approximate Bayesian Gaussian processes for probabilistic programming,”Statistics and Computing, vol. 33, no. 1, 2023

    work page 2023

  33. [33]

    Marginalized adaptive particle filtering for nonlinear models with unknown time-varying noise parameters,

    E. ¨Ozkan, V . ˇSm´ıdl, S. Saha, C. Lundquist, and F. Gustafsson, “Marginalized adaptive particle filtering for nonlinear models with unknown time-varying noise parameters,”Automatica, vol. 49, no. 6, pp. 1566–1575, 2013

    work page 2013

  34. [34]

    Conjugate Bayesian analysis of the Gaussian distribu- tion,

    K. P. Murphy, “Conjugate Bayesian analysis of the Gaussian distribu- tion,”Technical Report, 2007

    work page 2007

  35. [35]

    Ljung,System identification: Theory for the user, 2nd ed

    L. Ljung,System identification: Theory for the user, 2nd ed. N.J.: Prentice Hall, 1999

    work page 1999