pith. sign in

arxiv: 1906.10489 · v1 · pith:ZSOOKYQNnew · submitted 2019-06-25 · 📡 eess.SY · cs.RO· cs.SY

Keep soft robots soft -- a data-driven based trade-off between feed-forward and feedback control

Thomas Beckers , Sandra Hirche This is my paper

Pith reviewed 2026-05-25 16:36 UTC · model grok-4.3

classification 📡 eess.SY cs.ROcs.SY
keywords soft robotsGaussian Process regressionfeed-forward controlfeedback controltracking controldata-driven modelmodel uncertaintycompliance
0
0 comments X

The pith

A data-driven Gaussian Process model adapts feed-forward and feedback gains in soft robots to keep feedback low where model confidence is high.

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

Soft robots lose their safety advantage when high feedback gains are needed to overcome model uncertainties during tracking control. The paper uses Gaussian Process regression to learn a feed-forward compensator from data and then scales down feedback gains according to the model's local fidelity. This keeps the robot compliant in regions where the learned model is reliable while still handling uncertainties elsewhere. A reader would care because the method directly preserves the low-stiffness property that makes soft robots safe around humans and fragile objects.

Core claim

The central claim is that Gaussian Process regression produces a data-driven model for feed-forward compensation of unknown dynamics, and the model's fidelity is used to adapt the relative weighting of feed-forward and feedback terms so that low feedback gains can be used in regions of high model confidence.

What carries the argument

Gaussian Process regression whose uncertainty measure determines the balance between feed-forward compensation and feedback control gains.

If this is right

  • Low feedback gains are used only in regions where the data-driven model has high confidence.
  • The inherent compliance and safety of the soft robot are preserved by avoiding unnecessary stiffness.
  • Tracking accuracy improves through feed-forward compensation without raising overall feedback effort.
  • The same model fidelity signal can be used to switch between control modes in real time.

Where Pith is reading between the lines

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

  • The same uncertainty-driven gain scheduling could be tested on other underactuated or uncertain mechanical systems.
  • Online updating of the Gaussian Process during operation might further enlarge the low-feedback regions over time.
  • Safety certificates for the reduced-gain regions would require explicit bounds linking GP variance to closed-loop stability margins.

Load-bearing premise

The Gaussian Process uncertainty measure reliably indicates regions where reduced feedback will not destabilize the closed-loop system or violate safety constraints.

What would settle it

A closed-loop experiment on a soft robot in which feedback gains are lowered according to the GP uncertainty and instability or safety violation is observed.

Figures

Figures reproduced from arXiv: 1906.10489 by Sandra Hirche, Thomas Beckers.

Figure 1
Figure 1. Figure 1: Generation of the training data set for the GP. [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Small tracking error inside and outside the training data set. The [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: The robot remains soft in regions of high model confidence because [PITH_FULL_IMAGE:figures/full_fig_p002_3.png] view at source ↗
read the original abstract

Tracking control for soft robots is challenging due to uncertainties in the system model and environment. Using high feedback gains to overcome this issue results in an increasing stiffness that clearly destroys the inherent safety property of soft robots. However, accurate models for feed-forward control are often difficult to obtain. In this article, we employ Gaussian Process regression to obtain a data-driven model that is used for the feed-forward compensation of unknown dynamics. The model fidelity is used to adapt the feed-forward and feedback part allowing low feedback gains in regions of high model confidence.

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

1 major / 2 minor

Summary. The manuscript proposes a data-driven control method for soft robots that uses Gaussian Process regression to learn a model of unknown dynamics for feed-forward compensation. Model fidelity, quantified via the GP predictive variance, is then used to adaptively schedule the relative weighting between the feed-forward and feedback terms, permitting lower feedback gains (and thus lower stiffness) in regions of high model confidence while still achieving tracking performance.

Significance. If a rigorous stability argument can be supplied, the result would be significant for the field: it directly targets the safety-compliance trade-off that is central to soft-robotics applications by replacing conservative high-gain feedback with an uncertainty-aware blend of learned feed-forward and reduced-gain feedback. The approach is conceptually attractive because it leverages the natural uncertainty quantification of GPs rather than ad-hoc tuning.

major comments (1)
  1. [Abstract and adaptive control law] Abstract and § on the adaptive law: the central claim that GP predictive variance can be used to safely lower feedback gains rests on the unproven assumption that low variance implies a sufficiently tight bound on residual dynamics. No Lyapunov or small-gain argument is supplied that incorporates the scheduled gain reduction and the GP variance as an explicit uncertainty term; without such an argument the closed-loop stability claim is unsupported.
minor comments (2)
  1. Notation for the GP kernel and the exact functional form of the fidelity measure used to schedule the gains should be stated explicitly rather than left at the level of 'model fidelity'.
  2. The experimental section would benefit from a direct comparison of closed-loop eigenvalues or gain margins when the scheduler is active versus when a fixed high-gain controller is used.

Simulated Author's Rebuttal

1 responses · 0 unresolved

We thank the referee for the detailed and constructive review. We agree that the stability claim requires a rigorous argument that explicitly incorporates the GP-variance-based gain scheduling, and we will supply such an analysis in the revision.

read point-by-point responses

  1. Referee: [Abstract and adaptive control law] Abstract and § on the adaptive law: the central claim that GP predictive variance can be used to safely lower feedback gains rests on the unproven assumption that low variance implies a sufficiently tight bound on residual dynamics. No Lyapunov or small-gain argument is supplied that incorporates the scheduled gain reduction and the GP variance as an explicit uncertainty term; without such an argument the closed-loop stability claim is unsupported.

    Authors: We acknowledge that the manuscript currently relies on the heuristic that low GP predictive variance corresponds to small residual dynamics without a formal proof. The adaptive law reduces feedback gains proportionally to the variance, but no Lyapunov or small-gain theorem is stated that treats the variance as a state-dependent uncertainty bound. In the revised version we will add a stability section that (i) recalls the standard GP error bound under the assumption of bounded RKHS norm, (ii) defines a composite Lyapunov function whose derivative is shown negative definite when the scheduled gain remains above a variance-dependent threshold, and (iii) proves that the resulting closed-loop trajectories remain uniformly ultimately bounded. This will directly address the referee’s concern and strengthen the safety claim. revision: yes

Circularity Check

0 steps flagged

No significant circularity; GP trained on external data with minor self-citation not load-bearing

full rationale

The derivation relies on Gaussian Process regression fitted to external measurement data for the feed-forward model, with model fidelity (predictive variance) used to schedule feed-forward/feedback gains. No equation or claim reduces by construction to a self-defined quantity or renames a fitted input as a prediction. Self-citations, if present, do not carry the central stability or performance argument; the approach remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only; no explicit free parameters, axioms, or invented entities are stated.

pith-pipeline@v0.9.0 · 5616 in / 927 out tokens · 23164 ms · 2026-05-25T16:36:35.842120+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

15 extracted references · 15 canonical work pages · 2 internal anchors

  1. [1]

    Amiri Moghadam and K

    A.A. Amiri Moghadam and K. et al. Torabi. Control- oriented modeling of a polymeric soft robot. Soft Robotics, 3(2), 2016. URL https://doi.org/10.1089/soro. 2016.0002

    work page doi:10.1089/soro 2016

  2. [2]

    Beckers, D

    T. Beckers, D. Kulic, and S. Hirche. Stable Gaussian process based tracking control of Euler-Lagrange sys- tems. Automatica, 103:390–397, 2019. URL https: //doi.org/10.1016/j.automatica.2019.01.023

    work page doi:10.1016/j.automatica.2019.01.023 2019

  3. [3]

    Pattern recognition and machine learning , volume 4

    Christopher M Bishop et al. Pattern recognition and machine learning , volume 4. Springer New York, 2006. URL https://www.springer.com/de/book/ 9780387310732

    work page 2006

  4. [4]

    Modeling and Control of Soft Robots Using the Koopman Operator and Model Predictive Control

    Daniel Bruder, Brent Gillespie, C David Remy, and Ram Vasudevan. Modeling and control of soft robots using the koopman operator and model predictive control. arXiv preprint arXiv:1902.02827, 2019. URL https://arxiv.org/ abs/1902.02827

    work page internal anchor Pith review Pith/arXiv arXiv 1902

  5. [5]

    Controlling soft robots: balancing feedback and feedforward elements

    Cosimo Della Santina, Matteo Bianchi, Giorgio Grioli, Franco Angelini, Manuel Catalano, Manolo Garabini, and Antonio Bicchi. Controlling soft robots: balancing feedback and feedforward elements. IEEE Robotics & Automation Magazine , 24(3):75–83, 2017. URL https: //doi.org/10.1109/MRA.2016.2636360

    work page doi:10.1109/mra.2016.2636360 2017

  6. [6]

    Dynamic control of soft robots interacting with the environment

    Cosimo Della Santina, Robert K Katzschmann, Antonio Biechi, and Daniela Rus. Dynamic control of soft robots interacting with the environment. In 2018 IEEE International Conference on Soft Robotics (RoboSoft) , pages 46–53. IEEE, 2018. URL https://doi.org/10.1109/ ROBOSOFT.2018.8404895

    work page arXiv 2018

  7. [7]

    Dy- namic modeling of bellows-actuated continuum robots using the euler–lagrange formalism

    Valentin Falkenhahn, Tobias Mahl, Alexander Hilde- brandt, R ¨udiger Neumann, and Oliver Sawodny. Dy- namic modeling of bellows-actuated continuum robots using the euler–lagrange formalism. IEEE Transactions on Robotics , 31(6):1483–1496, 2015. URL https://doi. org/10.1109/TRO.2015.2496826

    work page doi:10.1109/tro.2015.2496826 2015

  8. [8]

    Falanga, S

    Ge Fang, Xiaomei Wang, Kui Wang, Kit-Hang Lee, Justin DL Ho, Hing-Choi Fu, Denny Kin Chung Fu, and Ka-Wai Kwok. Vision-based online learning kinematic control for soft robots using local gaussian process regression. IEEE Robotics and Automation Letters , 4 (2):1194–1201, 2019. URL https://doi.org/10.1109/LRA. 2019.2893691

    work page doi:10.1109/lra 2019

  9. [9]

    SOF A: A Multi- Model Framework for Interactive Physical Simulation , pages 283–321

    Franc ¸ois Faure, Christian Duriez, Herv ´e Delingette, J´er´emie Allard, Benjamin Gilles, St´ephanie Marchesseau, Hugo Talbot, Hadrien Courtecuisse, Guillaume Bous- quet, Igor Peterlik, and St ´ephane Cotin. SOF A: A Multi- Model Framework for Interactive Physical Simulation , pages 283–321. Springer Berlin Heidelberg, 2012. URL https://doi.org/10.1007/8...

    work page doi:10.1007/8415 2012

  10. [10]

    Learning nonlinear dynamic models of soft robots for model predictive control with neural networks

    Morgan T Gillespie, Charles M Best, Eric C Townsend, David Wingate, and Marc D Killpack. Learning nonlinear dynamic models of soft robots for model predictive control with neural networks. In 2018 IEEE International Conference on Soft Robotics (RoboSoft) , pages 39–45. IEEE, 2018. URL https://doi.org/10.1109/ROBOSOFT. 2018.8404894

    work page doi:10.1109/robosoft 2018

  11. [11]

    Remarks on multi-output gaussian process regression

    Haitao Liu, Jianfei Cai, and Yew-Soon Ong. Remarks on multi-output gaussian process regression. Knowledge- Based Systems, 144:102–121, 2018. URL https://doi.org/ 10.1016/j.knosys.2017.12.034

    work page doi:10.1016/j.knosys.2017.12.034 2018

  12. [12]

    On the Modelling of Soft-robots as Quasi-Continuum Lagrangian Dynamical Systems with Well-posed Input Matrix

    Ernesto Olgu ´ın-D´ıaz, Christian A Trejo-Ramos, Vicente Parra-Vega, and David Navarro-Alarc ´on. On the mod- elling of soft-robots as quasi-continuum lagrangian dy- namical systems with well-posed input matrix. arXiv preprint arXiv:1812.02738, 2018. URL https://arxiv.org/ abs/1812.02738

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  13. [13]

    Gaussian processes for machine learning , volume 1

    Carl Edward Rasmussen and Christopher KI Williams. Gaussian processes for machine learning , volume 1. MIT press Cambridge, 2006. URL http://www. gaussianprocess.org/gpml/chapters/

    work page 2006

  14. [14]

    Hy- brid analytical and data-driven modeling for feed-forward robot control

    Ren ´e Reinhart, Zeeshan Shareef, and Jochen Steil. Hy- brid analytical and data-driven modeling for feed-forward robot control. Sensors, 17(2):311, 2017. URL https: //doi.org/10.3390/s17020311

    work page doi:10.3390/s17020311 2017

  15. [15]

    Robot modeling and control, volume 3

    Mark W Spong, Seth Hutchinson, and Mathukumalli Vidyasagar. Robot modeling and control, volume 3. wiley New York, 2006. URL https://www.wiley.com/en-us/ Robot+Modeling+and+Control-p-9780471649908

    work page 2006