Pith. sign in

REVIEW 5 major objections 4 minor 48 references

Rapid Mismatch Estimation via Neural Network Informed Variational Inference

T0 review · 5 major / 4 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read Rapid Mismatch Estimation claims a robot can detect and compensate sudden changes in a load's mass and center of mass in about 400 ms, using only joint sensing, while keeping the impedance controller passive.

desk verdict A useful, honestly-reported estimation framework for payload mismatch in impedance control, but the passivity claim is weaker than advertised; worth refereeing with revisions, not a desk reject. read the letter →

arxiv 2508.21007 v1 pith:XBFSQV32 submitted 2025-08-28 cs.RO

Mateusz Jaszczuk , Nadia Figueroa This is my paper · ORCID
classification cs.RO
keywords modelmismatchestimationvariationalinferencepassiveimpedancecontrolpayloadidentificationcenter-of-masshuman-robotinteractionproprioceptivefeedbackneuralnetworkprior
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper introduces Rapid Mismatch Estimation (RME), a framework that lets a torque-controlled robot detect when an unknown object changes the mass and center of mass at its end-effector, and compensate for the change in about 400 milliseconds using only its own joint feedback, with no force-torque sensor on the wrist. The central claim is that the compensation arrives fast and accurately enough that the impedance controller's passivity and convergence guarantees survive the disturbance, so the robot stays compliant and safe even while a human loads or unloads its end-effector. The method couples a neural network, trained in simulation, which proposes an initial guess, with a variational inference solver that refines the guess into a posterior distribution over mass and center-of-mass parameters, quantifying its own uncertainty. If the claim holds, collaborative robots could handle unknown and changing payloads without added sensing hardware or retuning, preserving the soft, energy-based behavior that makes physical human-robot interaction work.

What carries the argument

The load-bearing object is the mismatch model itself: the unknown payload is a point mass m at center of mass r_CoM, producing the joint torque τ_mm = J(q)ᵀ[F_m; r_CoM × F_m] with F_m = [0,0,mg] (equation 4). That four-parameter model enters the Gaussian likelihood p(D|θ) = N(f_ID(q,θ), diag(σ²_likelihood)) of equation (10), turning estimation into Bayesian inference. The inference is two-stage: a convolution-plus-attention network f_NN, trained on simulated pseudo-wrench sequences, sets the mean of the Gaussian prior, and mean-field variational inference then minimizes the KL divergence by maximizing the ELBO via the reparameterization trick, returning a posterior whose mean plugs into the

What would settle it

Attach a payload of known, non-negligible rotational inertia (say a dumbbell) and drive the arm through fast, high-acceleration motion while RME runs. Because the likelihood contains only the gravitational wrench, the predicted behavior is that the inferred (m, r_CoM) drift from known ground truth as acceleration grows, surfacing as a growing residual wrench and a shifted equilibrium. A cleaner check of the observability claim: hold a centered load at a fixed pose and record the variational posterior variance of r_z — it should stay large and the estimate should hug the prior mean rather than

Watch

Extended reading notes

Core claim

The paper's central claim is that the gravitational wrench signature of an unknown end-effector load — treated as a point mass m at a center of mass r_CoM — can be read out of a roughly 200 ms window of proprioceptive torque measurements, and that a variational posterior over θ = {m, rx, ry, rz}, seeded by a simulation-trained neural network, converges to the true values fast enough to feed a compensation term τ̂c = τc − J(q)ᵀ[F̂m; r̂CoM × F̂m] into the nominal passive impedance controller. Proposition 4.1 states that even with imperfect estimates the closed loop stays passive with respect to the port (F_ext + ΔF_mm, ẋ): residual mismatch cannot inject energy, it can only reshape the attract

Load-bearing premise

The load is assumed to be exactly a point mass at the end-effector, felt only through gravity: the mismatch torque in equation (4) is [0,0,mg] acting at a center of mass. If the object has appreciable rotational inertia or the robot accelerates rapidly, the model is misspecified and the inferred mass and center of mass are biased — the paper concedes this in its Limitations section.

Editorial extensions

If this is right

  • A torque-controlled robot can absorb abrupt payload changes during a running task — a human attaching a basket or dropping heavy items into it — and correct itself in about 400 ms, without pausing, re-identifying offline, or adding a wrist force-torque sensor.
  • The estimator is controller-agnostic: it can be stacked on any passive impedance controller or passivity-preserving learned policy, because it injects only the estimated mismatch wrench instead of altering the nominal control law.
  • Even imperfect estimates are safe: by Proposition 4.1 the closed loop remains passive with respect to the port (F_ext + ΔF_mm, ẋ), so residual mismatch shifts or reshapes the attractor but cannot inject energy and destabilize the interaction.
  • Because each estimate completes in roughly 400 ms, the framework can run sequentially and self-correct a biased earlier estimate when the load changes again.
  • When a load's center of mass lies near the end-effector z-axis, r_z is weakly observable and the estimate stays near the prior, yet mass remains accurate and the robot still converges; the paper attributes this to the nonlinear dependence of inverse dynamics on θ.

Reading between the lines

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

  • The z-axis blind spot looks like an identifiability fact, not a tuning flaw: at fixed pose the gravitational wrench depends on r_z far more weakly than on m, so no longer data window can separate them. A testable extension — one the paper flags as future work — is to exploit the posterior's uncertainty to command brief orientation probes that restore CoM observability.
  • Because the likelihood contains only the gravitational term, estimates should drift progressively as acceleration grows. Extending the parameter set to a full inertia tensor (mass, CoM, inertia) would widen validity at the cost of some of the 400 ms speed; the paper's point-mass concession makes this the natural next step.
  • The passivity argument treats residual mismatch as an exogenous wrench; if the detection heuristic mistakes a sustained human push for a load — a failure the paper concedes can happen — compensation would partially cancel the human's force. A detector that consults the posterior's uncertainty or the full wrench profile would close that safety gap.
  • The successful sim-to-real transfer of the network prior suggests the bottleneck is the statistical mismatch model, not the network: any residual physics that can be written into the likelihood, such as joint friction or actuator dynamics, could in principle be estimated by the same NN-seeded variational inference pipeline.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

5 major / 4 minor

Summary. The paper proposes Rapid Mismatch Estimation (RME), a probabilistic framework that estimates end-effector dynamic mismatches—mass and center-of-mass—online from proprioceptive torque measurements, without external force/torque sensors. A neural network trained in simulation supplies the mean of a Gaussian prior, and variational inference refines the estimate in roughly 400 ms. The estimated mismatch is used to augment a passivity-based impedance controller (CPIC). The authors report static and dynamic experiments on a 7-DoF Franka arm, including a sequential human-robot interaction scenario with a basket, and claim that the RME-augmented controller preserves closed-loop passivity and task convergence under imperfect estimates. The central theoretical vehicle is Proposition 4.1, whose proof is given in Appendix H.

Significance. The practical contribution is genuine: RME demonstrates online mass estimation using proprioception alone on a real manipulator, with a useful NN-initialized VI formulation, 60 static trials, a sequential pHRI demonstration, and ablations showing the NN prior improves CoM MSE. If the theoretical guarantees were tightened, this would be a valuable systems contribution to adaptive impedance control. However, the advertised passivity guarantee is currently proven only under an unstated conservative-field assumption and for a combined port, and the CoM estimation is unreliable in two of six static test conditions. The paper is therefore of interest, but the claims are stronger than the evidence and proof support.

major comments (5)
  1. [Appendix H, Eq. (35)] The passivity proof inserts 'we assume to be conservative f(x)=-∇V(x)' immediately before Eq. (35). This assumption is not stated in Section 3 or in Proposition 4.1, and it is not implied by the stated Lyapunov condition ∇V(x)^T f(x) ≤ 0; a stable limit-cycle DS, such as the one evaluated in §5.3, satisfies that inequality without being a gradient flow. Without the added assumption, the term λ1(f(x)+∇V(x)) does not cancel, so the inequality ˙S ≤ ˙x^T(F_ext+ΔF_mm) does not follow. The proposition must either be restricted to conservative DS with the assumption stated explicitly, or proved for the general stable-DS class it claims.
  2. [Proposition 4.1 / Eq. (36)] The proof establishes passivity with respect to the input-output port (F_ext+ΔF_mm, ˙x), as stated in Prop. 4.1 and Eq. (36). This is not passivity with respect to the environment port (F_ext, ˙x), which is the property invoked in the abstract ('passively respond to contact') and in Section 4.3. Since ΔF_mm is an internal estimation error, it cannot be treated as an exogenous environment input. Unless the authors show that ΔF_mm is bounded and its energy injection is dissipated—e.g., through a bounded-error or dissipativity argument—the claim that RME preserves closed-loop passivity during imperfect estimation is stronger than the proof supports.
  3. [Table 2, §5.2] CoM estimation fails for two of the six applied-mismatch conditions: for true m=1.1 kg and 1.29 kg, the inferred rz is 0.00±0.04 and 0.00±0.06 m, respectively, versus true rz=0.13 m, while all lower-mass conditions recover approximately 0.11–0.15 m. The text attributes this to observability, but the stated goal in §2 is to estimate θ={m,rx,ry,rz} online, and Table 2 shows that the central CoM claim is unsupported in one-third of the static conditions. In addition, the smallest mass (0.300 kg) is overestimated by 27% (mean 0.380 kg). The §5.2 statement of 'accurate estimation of mismatch parameters θ' should be qualified accordingly.
  4. [Appendix H, Eqs. (22)–(24)] The statement after Eq. (22) that the constrained QP is 'always feasible' is not established and is false in general for inequality-constrained QPs: the E-CBF constraint Aτ_c ≥ ν may be infeasible for a given state and gain vector, and no feasibility condition for h, K, or state is provided. The subsequent KKT analysis assumes an active set exists. To use the 'passive when feasible' argument, the authors need either to prove feasibility of (8) under the stated assumptions or to characterize the infeasible set and show that the controller behaves safely there.
  5. [§5.3 and §7 Limitations] The dynamic limit-cycle experiment operates in the regime the authors later admit is outside the model's validity. The Limitations state that under 'rapid accelerations along the global z-axis ... the mismatch prediction might be biased' and that the point-mass assumption 'holds when the manipulator is not rapidly accelerating.' The stable-limit-cycle experiment is a dynamic task with such accelerations, so the §5.3 claim that RME 'rapidly estimates mismatch parameters' in the dynamic setting is not supported by the model-valid regime. The authors should provide quantitative accuracy results for the limit-cycle condition or explicitly report the expected bias as a limitation of the dynamic evaluation.
minor comments (4)
  1. [Section 4.1, Figure 2] 'mean polling' should be 'mean pooling' (also in the Figure 2 caption).
  2. [Table 2] Prediction formatting is inconsistent (e.g., '0.380 ±0.018' vs '-0.04±0.03'). Use consistent spacing and parentheses for standard deviations.
  3. [Figure 14] The y-axis tick labels render as '10 2' and '10 1'; use proper superscripts for readability.
  4. [Abstract / §5.2 / Appendix F] The paper reports '~400 ms' estimation, while §5.2 gives an average model estimation time of 226 ms and Appendix F adds a 200 ms data-collection window. Please clarify whether the 400 ms figure includes detection, data collection, and inference, so the timing claim is unambiguous.

Circularity Check

0 steps flagged · score 0.0 of 10

No circular derivation: RME is a fitted estimator and the passivity proof, though scoped by an unstated conservative-field assumption, is not a reduction to its inputs.

full rationale

None of the paper's load-bearing claims reduces to its own inputs by construction. The RME estimate θ is obtained by minimizing the VI loss (16) with likelihood (10); this is a standard maximum-likelihood/regularized fit of θ to measured proprioceptive torques, not a first-principles derivation. The NN prior fNN(D) is trained in simulation (Section G) and used only to initialize the prior mean; the physical-robot experiments in Section 5 are external validation, so the estimation claim is not forced by the training data. The compensation law (5) uses the same point-mass gravitational model as the likelihood, which makes the estimator self-consistent, but the parameters are not defined as the NN output; they are optimized against the torque data. The passivity claim in Proposition 4.1 is a separate mathematical statement; Appendix H derives Sdot ≤ ẋ^T(Fext+ΔFmm) from the closed-loop equations. The insertion 'for this analysis we assume to be conservative f(x) = -∇V(x)' immediately before (35) is a scope limitation: it is stronger than the Lyapunov condition ∇V^T f ≤ 0 stated in (6), and it is not satisfied by the limit-cycle policy in Section 5.3. That is a correctness/assumption gap in the proof, not a circular reduction. Self-citations [32], [35], [46] supply the nominal controller and background; the augmented passivity proof is re-derived rather than imported wholesale, and the estimation contribution is independent of those citations. Therefore no circular step can be exhibited from the text.

Assumptions & free parameters 6 free parameters · 6 assumptions · 0 invented entities

The central estimation claim rests on a known inverse-dynamics model, a point-mass mismatch model, and a learned prior; none are derived from first principles. The passivity proof additionally relies on the nominal controller's passivity and an extra conservative-field assumption.

free parameters (6)
  • VI prior standard deviation σ_prior = [0.5, 0.02, 0.02, 0.05]
    Empirically tuned in Section 5; used in (10) to build p(θ) and shapes posterior uncertainty.
  • Likelihood noise σ²_likelihood = per-joint noise amplitude from D
    Set from measured torque noise in each dataset (Section 4); scales the likelihood in (10).
  • Mismatch detection thresholds = activation 0.7-1.1 Nm, stabilization 0.2 Nm, interval 230 ms
    Found empirically from torque profile analysis (Appendix A); gates whether RME runs at all.
  • VI learning rate η = 0.025
    Adam hyperparameter for ELBO optimization (17).
  • Damping coefficient λ for pseudo-wrench computation = not reported
    Used in (11) to compute Ŵext from τext; value absent, hindering reproduction.
  • Data collection interval = 200 ms, M=20
    Chosen from ablation in Appendix F as a speed-accuracy trade-off.
assumptions (6)
  • domain assumption Nominal rigid-body dynamics (1) are known and satisfy skew-symmetry (2)
    Required for passivity arguments in Section 2 and Appendix H; standard idealization for revolute-joint manipulators.
  • ad hoc to paper Model mismatch is represented as a point mass at the end-effector with gravitational force Fm=[0,0,mg]
    Equation (4); fails for significant object inertia or rapid acceleration, as the paper concedes in Limitations.
  • domain assumption The nominal passive interaction controller (7)-(8) is passive wrt (Fext, ẋ)
    Taken from prior work [13,32]; needed for the claim that compensation preserves passivity.
  • ad hoc to paper The dynamical system f(x) is conservative, f(x)=-∇V(x)
    Assumed inside the proof of Proposition 4.1 (Appendix H, after eq. 34); not stated as a general controller requirement.
  • ad hoc to paper The QP safety filter (8) is always feasible
    Appendix H (eq. 19) asserts 'always feasible' without proof; a linear inequality in τc may be infeasible.
  • domain assumption External joint torque measurements τext accurately reflect the mismatch signal
    The likelihood (10) treats τext as the true mismatch torque; the paper notes joint friction and actuator dynamics can bias it.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Rapid Mismatch Estimation via Neural Network Informed Variational Inference." pith.science (2026). https://pith.science/paper/XBFSQV32

@misc{pith2026250821007,
  author       = {Pith},
  title        = {Pith review of: Rapid Mismatch Estimation via Neural Network Informed Variational Inference},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XBFSQV32}},
  note         = {Machine review of arXiv:2508.21007}
}
abstract

With robots increasingly operating in human-centric environments, ensuring soft and safe physical interactions, whether with humans, surroundings, or other machines, is essential. While compliant hardware can facilitate such interactions, this work focuses on impedance controllers that allow torque-controlled robots to safely and passively respond to contact while accurately executing tasks. From inverse dynamics to quadratic programming-based controllers, the effectiveness of these methods relies on accurate dynamics models of the robot and the object it manipulates. Any model mismatch results in task failures and unsafe behaviors. Thus, we introduce Rapid Mismatch Estimation (RME), an adaptive, controller-agnostic, probabilistic framework that estimates end-effector dynamics mismatches online, without relying on external force-torque sensors. From the robot's proprioceptive feedback, a Neural Network Model Mismatch Estimator generates a prior for a Variational Inference solver, which rapidly converges to the unknown parameters while quantifying uncertainty. With a real 7-DoF manipulator driven by a state-of-the-art passive impedance controller, RME adapts to sudden changes in mass and center of mass at the end-effector in $\sim400$ ms, in static and dynamic settings. We demonstrate RME in a collaborative scenario where a human attaches an unknown basket to the robot's end-effector and dynamically adds/removes heavy items, showcasing fast and safe adaptation to changing dynamics during physical interaction without any external sensory system.

Figures

Figures reproduced from arXiv: 2508.21007 by the authors.

Figure 1
Figure 1. Rapid Mismatch Estimation (RME) Framework. The gray block denotes our novel contribution. Note that RME is controller agnostic and can be implemented with any impedance controller. In this work, we choose the constrained passive interaction controller [32]. with task-space gravity vector Gx(q) ∈ R 6 , damping matrix D(x) ∈ R 6×6 , and de￾sired velocity from f(x). By aligning D(x) to the direction of f(x) we can gene… view at source ↗
Figure 2
Figure 2. RME Neural Network Architecture. In the network, we input a sequence of pseudo￾wrenches of dimension RM×64, apply a convolution layer, positional embedding, and multi-head attention. Further, we mean pool over the attention scores and apply a sequential Multilayer Percep￾tron, which performs a final regression to mismatch parameters θ. a NN fNN(·) as the mean for the Gaussian Distribution, as described in Section 4.… view at source ↗
Figure 3
Figure 3. Parity plot for mass predictions over 60 independent experiments. Results show con￾sistent and stable mass estimation using RME [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] view at source ↗
Figures from the paper (9 more)
Figure 5
Figure 5. Figure 5: Manipulator tracking a DS with a stable limit cycle in the y-z plane subject to mismatch in the dynamics model. Tracking a Stable Limit Cycle To test the model in the dy￾namic scenario, we designed a Dynamical System with a sta￾ble limit cycle in the y-z plane, where t…
Figure 6
Figure 6. Figure 6: Manipulator adaptation to sequential changes in the dynamics model, while subject to [PITH_FULL_IMAGE:figures/full_fig_p008_6.png]
Figure 7
Figure 7. Figure 7: Example of mismatch detection algorithm execution. The algorithm monitors [PITH_FULL_IMAGE:figures/full_fig_p014_7.png]
Figure 8
Figure 8. Figure 8: Comparison of parity plots for RME estimation of mismatch parameters [PITH_FULL_IMAGE:figures/full_fig_p014_8.png]
Figure 9
Figure 9. Figure 9: Manipulator adaptation to sequential changes in the dynamics model while interacting with [PITH_FULL_IMAGE:figures/full_fig_p015_9.png]
Figure 10
Figure 10. Figure 10: Manipulator intercepting basket of unknown mass (1200 grams) from a user and placing it [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]
Figure 11
Figure 11. Figure 11: End-effectors and unknown objects used in the RME evaluation. The left image shows [PITH_FULL_IMAGE:figures/full_fig_p016_11.png]
Figure 12
Figure 12. Figure 12: Comparison of manipulators’ behavior subject to unknown mismatch in dynamics model [PITH_FULL_IMAGE:figures/full_fig_p017_12.png]
Figure 13
Figure 13. Figure 13: Computer simulations of the manipulator’s dynamics subject to an unknown mass at [PITH_FULL_IMAGE:figures/full_fig_p018_13.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

48 extracted references · 35 canonical work pages

  1. [1]

    Stramigioli

    S. Stramigioli. Energy-aware robotics. In M. K. Camlibel, A. A. Julius, R. Pasumarthy, and J. M. Scherpen, editors, Mathematical Control Theory I, pages 37–50, Cham, 2015. Springer International Publishing. ISBN 978-3-319-20988-3

    work page 2015

  2. [2]

    A. Billard. On the mechanical, cognitive and sociable facets of human compliance and their robotic counterparts. Robotics and Autonomous Systems, 88:157–164, 2017. ISSN 0921-8890. doi:https://doi.org/10.1016/j.robot.2016.08.030. URL https://www.sciencedirect.com/ science/article/pii/S0921889016306303

    work page doi:10.1016/j.robot.2016.08.030 2017

  3. [3]

    Sharif Razavian, B

    R. Sharif Razavian, B. Ghannadi, and J. McPhee. On the relationship between mus- cle synergies and redundant degrees of freedom in musculoskeletal systems. Fron- tiers in Computational Neuroscience , V olume 13 - 2019, 2019. ISSN 1662-

    work page 2019

  4. [4]

    G. L. Gottlieb. Muscle compliance: implications for the control of movement. Exerc Sport Sci Rev, 24:1–34, 1996

    work page 1996

  5. [5]

    J. Colgate. Robust control of dynamically interacting systems. International Journal of Con- trol, 48(1):65–88, July 1988. ISSN 0020-7179. doi:10.1080/00207178808906161

    work page doi:10.1080/00207178808906161 1988

  6. [6]

    N. Hogan. Impedance control: An approach to manipulation. In 1984 American Control Conference, pages 304–313, 1984. doi:10.23919/ACC.1984.4788393. 9

    arXiv 1984

  7. [7]

    Stramigioli, C

    S. Stramigioli, C. Melchiorri, and S. Andreotti. A passivity-based control scheme for robotic grasping and manipulation. In Proceedings of the 38th IEEE Conference on Decision and Control (Cat. No.99CH36304), volume 3, pages 2951–2956 vol.3, 1999. doi:10.1109/CDC. 1999.831385

    arXiv 1999

  8. [8]

    Li and R

    P. Li and R. Horowitz. Passive velocity field control of mechanical manipulators. IEEE Trans- actions on Robotics and Automation, 15(4):751–763, 1999. doi:10.1109/70.782030

    work page doi:10.1109/70.782030 1999

Show all 48 references
  1. [9]

    Kishi, Z

    Y . Kishi, Z. W. Luo, F. Asano, and S. Hosoe. Passive impedance control with time-varying impedance center. In Proceedings 2003 IEEE International Symposium on Computational Intelligence in Robotics and Automation. Computational Intelligence in Robotics and Automa- tion for th...

    2003 arXiv

  2. [10]

    Duindam, S

    V . Duindam, S. Stramigioli, and J. Scherpen. Passive compensation of nonlinear robot dynamics. IEEE Transactions on Robotics and Automation , 20(3):480–488, 2004. doi: 10.1109/TRA.2004.824693

    2004

  3. [11]

    Albu-Sch ¨affer, C

    A. Albu-Sch ¨affer, C. Ott, and G. Hirzinger. A unified passivity-based control framework for position, torque and impedance control of flexible joint robots. The International Journal of Robotics Research, 26(1):23–39, 2007. doi:10.1177/0278364907073776

    2007 doi

  4. [12]

    Ferraguti, C

    F. Ferraguti, C. Secchi, and C. Fantuzzi. A tank-based approach to impedance control with variable stiffness. In 2013 IEEE International Conference on Robotics and Automation, pages 4948–4953, 2013. doi:10.1109/ICRA.2013.6631284

    2013

  5. [13]

    Kronander and A

    K. Kronander and A. Billard. Passive interaction control with dynamical systems. IEEE Robotics and Automation Letters, 1(1):106–113, 2015

    2015

  6. [14]

    Ferraguti, N

    F. Ferraguti, N. Preda, A. Manurung, M. Bonf`e, O. Lambercy, R. Gassert, R. Muradore, P. Fior- ini, and C. Secchi. An energy tank-based interactive control architecture for autonomous and teleoperated robotic surgery. IEEE Transactions on Robotics , 31(5):1073–1088, 2015. doi:...

    2015

  7. [15]

    G. A. Folkertsma and S. Stramigioli. Energy in robotics.Foundations and Trends® in Robotics, 6(3):140–210, 2017. ISSN 1935-8253. doi:10.1561/2300000038. URL http://dx.doi. org/10.1561/2300000038

    2017 doi

  8. [16]

    Dietrich, X

    A. Dietrich, X. Wu, K. Bussmann, C. Ott, A. Albu-Sch ¨affer, and S. Stramigioli. Passive hierarchical impedance control via energy tanks. IEEE Robotics and Automation Letters , 2 (2):522–529, 2017. doi:10.1109/LRA.2016.2645504

    2017

  9. [17]

    F. J. Abu-Dakka and M. Saveriano. Variable impedance control and learning—a review. Fron- tiers in Robotics and AI, 7, 2020. ISSN 2296-9144. doi:10.3389/frobt.2020.590681

    2020

  10. [18]

    B. L. Riemann and S. M. Lephart. The sensorimotor system, part II: The role of proprioception in motor control and functional joint stability.Journal of Athletic Training, 37(1):80–84, 2002

    2002

  11. [19]

    K. Oh, W. Z. Rymer, and J. Choi. The speed of adaptation is dependent on the load type during target reaching by intact human subjects. Experimental Brain Research, 239(10):3091–3104, Oct. 2021. ISSN 1432-1106

    2021

  12. [20]

    Y . Sun, W. L. Ubellacker, W.-L. Ma, X. Zhang, C. Wang, N. V . Csomay-Shanklin, M. Tomizuka, K. Sreenath, and A. D. Ames. Online learning of unknown dynamics for model- based controllers in legged locomotion. IEEE Robotics and Automation Letters , 6(4):8442– 8449, 2021. doi:10...

    2021

  13. [21]

    Khadivar, K

    F. Khadivar, K. Chatzilygeroudis, and A. Billard. Self-correcting quadratic programming- based robot control. IEEE Transactions on Systems, Man, and Cybernetics: Systems , 53(8): 5236–5247, 2023. doi:10.1109/TSMC.2023.3262954. 10

    2023

  14. [22]

    Nguyen-Tuong and J

    D. Nguyen-Tuong and J. Peters. Using model knowledge for learning inverse dynamics. In 2010 IEEE International Conference on Robotics and Automation , pages 2677–2682, 2010. doi:10.1109/ROBOT.2010.5509858

    2010

  15. [23]

    Koryakovskiy, M

    I. Koryakovskiy, M. Kudruss, H. Vallery, R. Babu ˇska, and W. Caarls. Model-plant mismatch compensation using reinforcement learning. IEEE Robotics and Automation Letters , 3(3): 2471–2477, 2018. doi:10.1109/LRA.2018.2800106

    2018

  16. [24]

    F. Gao, C. Yu, Y . Wang, and Y . Wu. Neural internal model control: Learning a robust control policy via predictive error feedback.IEEE Robotics and Automation Letters, 10(7):6848–6855,

  17. [25]

    Srour, A

    A. Srour, A. Franchi, P. R. Giordano, and M. Cognetti. Experimental validation of sensitivity- aware trajectory planning for a redundant robotic manipulator under payload uncertainty.IEEE Robotics and Automation Letters, 10(2):1561–1568, 2025. doi:10.1109/LRA.2024.3519857

    2025

  18. [26]

    Zhang, Z

    B. Zhang, Z. Zhou, and R. Vasudevan. Provably-safe, online system identification. InRobotics: Science and Systems, 2025

    2025

  19. [27]

    S. Jin, Y . Mo, and W. Yuan. Learning to double guess: An active perception approach for estimating the center of mass of arbitrary objects, 2025. URL https://arxiv.org/abs/ 2502.02663

    2025 arXiv

  20. [28]

    Colom ´e, D

    A. Colom ´e, D. Pardo, G. Aleny `a, and C. Torras. External force estimation during compliant robot manipulation. In 2013 IEEE International Conference on Robotics and Automation , pages 3535–3540, 2013. doi:10.1109/ICRA.2013.6631072

    2013

  21. [29]

    M. W. Spong, S. Hutchinson, and M. Vidyasagar. Robot Modeling and Control. Wiley, 2020

    2020

  22. [30]

    Dietrich, X

    A. Dietrich, X. Wu, K. Bussmann, M. Harder, M. Iskandar, J. Englsberger, C. Ott, and A. Albu-Sch ¨affer. Practical consequences of inertia shaping for interaction and tracking in robot control. Control Engineering Practice , 114:104875, 2021. ISSN 0967-0661. doi: https://doi.o...

    2021

  23. [31]

    Kronander and A

    K. Kronander and A. Billard. Stability considerations for variable impedance control. IEEE Transactions on Robotics, 32(5):1298–1305, 2016. doi:10.1109/TRO.2016.2593492

    2016

  24. [33]

    Sacerdoti, F

    D. Sacerdoti, F. Benzi, and C. Secchi. A reinforcement learning-based control strategy for robust interaction of robotic systems with uncertain environments. In 2024 IEEE In- ternational Conference on Robotics and Automation (ICRA) , pages 5788–5794, 2024. doi: 10.1109/ICRA571...

    2024

  25. [34]

    Zanella, F

    R. Zanella, F. Califano, C. Secchi, and S. Stramigioli. Learning passive policies. Springer Pro- ceedings in Advanced Robotics, 32 SPAR:338 – 343, 2024. doi:10.1007/978-3-031-76424-0

    2024 doi

  26. [35]

    Billard, S

    A. Billard, S. S. Mirrazavi Salehian, and N. Figueroa. Learning for Adap- tive and Reactive Robot Control: A Dynamical Systems Approach . MIT Press, Cambridge, USA, 2022. URL https://mitpress.mit.edu/books/ learning-adaptive-and-reactive-robot-control

    2022

  27. [36]

    Y . Hu, F. J. Abu-Dakka, F. Chen, X. Luo, Z. Li, A. Knoll, and W. Ding. Fusion dynamical sys- tems with machine learning in imitation learning: A comprehensive overview.Information Fu- sion, 108:102379, 2024. ISSN 1566-2535. doi:https://doi.org/10.1016/j.inffus.2024.102379. 11

    2024

  28. [37]

    Nguyen and K

    Q. Nguyen and K. Sreenath. Exponential control barrier functions for enforcing high relative- degree safety-critical constraints. In 2016 American Control Conference (ACC) , pages 322–

    2016

  29. [38]

    D. M. Blei, A. Kucukelbir, and J. D. McAuliffe. Variational inference: A review for statisti- cians. Journal of the American Statistical Association , 112(518):859–877, Apr. 2017. ISSN 1537-274X. doi:10.1080/01621459.2017.1285773

    2017

  30. [39]

    D. P. Kingma and M. Welling. Auto-encoding variational bayes, 2022. URLhttps://arxiv. org/abs/1312.6114

    2022 arXiv

  31. [40]

    Lecun, L

    Y . Lecun, L. Bottou, Y . Bengio, and P. Haffner. Gradient-based learning applied to document recognition. Proceedings of the IEEE, 86(11):2278–2324, 1998. doi:10.1109/5.726791

    1998 doi

  32. [41]

    Vaswani, N

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, L. Kaiser, and I. Polo- sukhin. Attention is all you need, 2023. URL https://arxiv.org/abs/1706.03762

    2023 arXiv

  33. [42]

    Srivastava, G

    N. Srivastava, G. Hinton, A. Krizhevsky, I. Sutskever, and R. Salakhutdinov. Dropout: A sim- ple way to prevent neural networks from overfitting. Journal of Machine Learning Research, 15(56):1929–1958, 2014. URL http://jmlr.org/papers/v15/srivastava14a.html

    1929

  34. [43]

    Kucukelbir, D

    A. Kucukelbir, D. Tran, R. Ranganath, A. Gelman, and D. M. Blei. Automatic differentiation variational inference, 2016. URL https://arxiv.org/abs/1603.00788

    2016 arXiv

  35. [44]

    D. P. Kingma and J. Ba. Adam: A method for stochastic optimization, 2017. URL https: //arxiv.org/abs/1412.6980

    2017 arXiv

  36. [45]

    Mattingley and S

    J. Mattingley and S. Boyd. Cvxgen: a code generator for embedded convex optimization. Optimization and Engineering , 13(1):1–27, 2012. doi:10.1007/s11081-011-9176-9. URL https://doi.org/10.1007/s11081-011-9176-9

    2012 doi

  37. [46]

    S. S. M. Salehian, N. Figueroa, and A. Billard. A unified framework for coordinated multi-arm motion planning. The International Journal of Robotics Research , 37(10):1205–1232, 2018. doi:10.1177/0278364918765952. URL https://doi.org/10.1177/0278364918765952. 12 Appendix A Mis...

    2018 doi

  38. [49]

    Further, it evaluates stabilization condition by computing the mean ∥τext∥2 2 over short time windows within the stabilization interval and compares it with the mean ∥τext∥2 2 at the end of the interval, allowing detection of both adding and removing objects from the end effec...

  39. [2025]

    doi:10.1109/LRA.2025.3573169

    2025

  40. [5188]

    URL https://www.frontiersin.org/journals/ computational-neuroscience/articles/10.3389/fncom.2019.00023

    doi:10.3389/fncom.2019.00023. URL https://www.frontiersin.org/journals/ computational-neuroscience/articles/10.3389/fncom.2019.00023

    2019

Pith tools

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