Pith. sign in

REVIEW 4 major objections 7 minor 43 references

Newtonian and Lagrangian Neural Networks: A Comparison Towards Efficient Inverse Dynamics Identification

T0 review · 4 major / 7 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read The paper claims that for inverse dynamics learned from estimated motor torques, Newtonian networks outperform Lagrangian networks because Lagrangian networks cannot explicitly represent dissipative torques.

desk verdict Useful real-robot benchmark with an overstated causal claim; worth reviewing after the abstract is fixed. read the letter →

arxiv 2506.17994 v1 pith:Q36MVNGN submitted 2025-06-22 cs.RO cs.LG

Minh Trinh , Andreas René Geist , Josefine Monnet , Stefan Vilceanu , Sebastian Trimpe , Christian Brecher This is my paper
classification cs.ROcs.LG
keywords inversedynamicsmotortorqueestimationLagrangianneuralnetworksNewtonianfrictionmodelingphysics-informedmachinelearningindustrialrobotsrecursiveNewton-Euleralgorithm
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

Industrial robots are usually controlled with predicted motor torques, but in most real systems only the motor current, and therefore an estimate of motor torque, is available rather than the joint torque itself. This paper compares two physics-informed model families on a six-axis MABI MAX 100 robot in that setting: Newtonian networks, which add a learned residual to the recursive Newton-Euler algorithm, and Lagrangian networks, which put neural networks inside the Euler-Lagrange equations. The central claim is that when the target is estimated motor torque, Lagrangian networks are less effective because they do not explicitly model dissipative torques such as friction, whereas the Newtonian residual structure absorbs them. On the collected data, the RNEA+MLP model significantly outperforms all other models, especially on the first four joints where friction forces are large. The practical message is a model-selection rule: with a nominal rigid-body model and estimated motor torques, prefer a Newtonian residual model with a learned friction term.

What carries the argument

The comparison turns on the additive decomposition of motor torque into rigid-body dynamics, motor inertia, and dissipative torque, written as $(\tau_u)_i = (\tau_{\mathrm{RBD}})_i/\psi_{m,i} + I_{M,i}\psi_{m,i}\ddot q_i + \tau_{D,i}(\dot q_i)$. Newtonian networks keep the analytical recursive Newton-Euler model fixed and add a neural network as an explicit residual that can absorb friction and other dissipative effects. Lagrangian networks instead replace the Lagrangian or the inertia matrix with a neural network and compute torques by differentiation, so dissipative torques must be folded into a conservative potential-like structure. The missing explicit dissipative term is the structural reason the paper gives for the performance gap.

What would settle it

Insert a joint-torque sensor into the first or second axis of the same robot and run the identical comparison; if Lagrangian networks match RNEA+MLP on directly measured joint torques while RNEA+MLP remains better on motor-torque estimates, the missing dissipative term is confirmed as the cause of the gap.

Watch

Extended reading notes

Core claim

The paper establishes a conditional separation between the two physics-informed modeling families. When joint torques are directly measured, Lagrangian networks can learn friction. When only motor torques estimated from currents are available, the target becomes the difference between joint torque and dissipative torque, and Lagrangian networks must represent dissipation through derivatives of a scalar Lagrangian, which they do poorly. The authors report that the Newtonian network (RNEA+MLP) significantly outperforms all other models, including DeLaN, LNN, and LNN plus an added MLP, with the largest margins on the first four joints, which carry the largest friction forces. Adding an MLP to a Lagrangian network improves its accuracy but does not close the gap with the Newtonian model.

Load-bearing premise

The result rests on the manufacturer-provided rigid-body model being accurate enough that the only substantial residual is friction; if the nominal inertia values or the high-gear-ratio simplification are wrong, the residual network must compensate for large structured errors, and the Newtonian advantage could disappear or reverse.

Editorial extensions

If this is right

  • When only motor-torque estimates are available, practitioners with a nominal rigid-body model should build an RNEA residual network with a learned friction term rather than a pure Lagrangian network.
  • Adding an extra MLP for dissipation makes Lagrangian networks usable but does not make them competitive with the Newtonian model on the tested robot.
  • For joints with small friction, here axes five and six, the gap between Newtonian and Lagrangian networks narrows, which supports friction as the decisive factor.
  • The conclusion is contingent on the measured variable: the paper leaves open that directly measured joint torques would make Lagrangian networks competitive.

Reading between the lines

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

  • The paper tests one robot, so a natural extension is a second platform with load-dependent friction or elastic joints, where the residual network in the Newtonian model might need a structured friction term instead of a generic MLP.
  • Because the stated reason is that dissipative torques are not gradients of a scalar Lagrangian, the same limitation should apply to other conservative-form networks such as Hamiltonian networks, although the paper does not test them.
  • A cheap check of the mechanism would be to train the residual network on the friction estimate alone using only velocity as input, since the paper's friction model depends only on $\dot q_i$; if that matches RNEA+MLP, friction is indeed the dominant residual.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

4 major / 7 minor

Summary. The paper compares physics-informed inverse dynamics models of the Lagrangian and Newtonian type on data from a MABI MAX 100 industrial robot, focusing on the practically relevant setting in which only motor-torque estimates (derived from motor currents) are available rather than direct joint-torque measurements. The models compared are DeLaN, LNN, LNN augmented with an MLP, an RNEA-based Newtonian model with a learned residual (RNEA+MLP), and two black-box baselines (MLP, LSTM). The central claim is that in this motor-torque setting, Lagrangian networks are less effective than Newtonian networks because they do not explicitly model dissipative torques, with RNEA+MLP reported as significantly outperforming all other models.

Significance. If the central claim were fully supported, the paper would provide practically useful guidance for inverse dynamics identification in industrial robots: use a nominal rigid-body model plus a learned residual when the output is estimated motor torque. The paper also makes a valuable conceptual distinction between joint torque and motor torque, and it shows that vanilla Lagrangian formulations cannot represent dissipative forces. Strengths include the use of a real industrial robot, a physically grounded RNEA baseline built from manufacturer data, and comparisons against standard non-physics baselines. The main weakness is that the causal attribution in the abstract is not isolated by the experiment, and the empirical evidence rests on a single dataset with no uncertainty quantification. The paper is therefore a useful step but requires a revision of both the experiment design and the strength of the claims.

major comments (4)
  1. [Abstract; Sec. 3.1-3.2; Sec. 4.4, Eq. (10)] The experiment confounds two factors: RNEA+MLP benefits from a fixed manufacturer-provided nominal rigid-body model (Eq. (10) with fixed θ_RBD), while DeLaN and LNN must learn inertia and gravity from data and, in their vanilla forms, cannot represent dissipative torques. The inclusion of LNN+MLP, which does add an explicit dissipative MLP, does not resolve the confound because it still underperforms RNEA+MLP on joints 1-4 in Table 1. The remaining gap could therefore be due to the nominal model, to optimization and initialization difficulty (as acknowledged in Sec. 3.1 and the conclusion), or to a combination of factors, rather than specifically to the absence of friction terms. To make the causal claim in the abstract stand, the authors should add an ablation that holds the nominal model fixed across architectures: for example, use the same nominal inertia and gravity terms from Eq. (13) inside a Lagrangian network and add an MLP only for the residual/dissipative component. Without such an ablation, the reported comparison is consistent with the abstract's mechanism but does not establish it.
  2. [Sec. 4.4, Table 1] The statement that "RNEA+MLP significantly outperforms all other models" is not supported by the reported numbers. On joint 6, RNEA+MLP (0.1654) is worse than LNN+MLP (0.1487) and LSTM (0.1375), and on joint 5 the RMSE differences among all models are below 0.004. The claim is also not backed by confidence intervals, repeated training seeds, or statistical significance tests. The authors should either restrict the claim to joints 1-4, where the pattern is consistent, or provide uncertainty estimates and tests that justify a global superiority statement.
  3. [Sec. 4.2, Eq. (13)] The Newtonian baseline depends on assumptions that are not independently validated. Specifically, Eq. (13) assumes that motor inertia couples only to the same joint due to high gear ratios, and the model uses manufacturer-provided kinematics and inertia parameters that are never checked against a joint-torque sensor. Because the MLP residual in RNEA+MLP can absorb errors in the nominal rigid-body model, the observed advantage could stem from the accuracy of this nominal model rather than from explicit dissipative-torque modeling. A sensitivity analysis that perturbs the manufacturer parameters, or a version in which θ_RBD is estimated from data, would strengthen the interpretation.
  4. [Abstract; Sec. 1; Sec. 5, Conclusion] The abstract states an unconditional causal claim: "Lagrangian networks prove less effective compared to Newtonian networks as they do not explicitly model dissipative torques." However, the introduction already restricts the conclusion to the case where "an estimate of the frictionless dynamics is given," and the experiment does not isolate the dissipative-torque mechanism. The abstract and Sec. 4.4 should be rephrased to state what was actually shown: on the tested robot, an RNEA residual model with a learned MLP had lower RMSE on the first four joints, while the causal role of explicit dissipation modeling remains a hypothesis that requires further ablations.
minor comments (7)
  1. [Sec. 4.3] The LNN+MLP model is not fully defined; the manuscript should specify where the MLP enters the LNN output (e.g., as an additive residual force) and what inputs it receives.
  2. [Sec. 4.3] The hyperparameter search is described only as "extensive" and was run in Weights & Biases; listing the search ranges and the number of trials per model would improve reproducibility.
  3. [Sec. 4.1] The low-pass filter cutoff frequencies are reported as identified in preliminary tests, but no details of the identification procedure or a sensitivity analysis of the conclusions to these values are provided.
  4. [Fig. 3] The stacked contribution plot in Figure 3 has no legend or axis labels in the manuscript text; adding a legend and y-axis units (Nm) would make the figure interpretable.
  5. [Table 1] The table reports RMSE values without units or a statement of whether the values are in Nm; adding the units and a brief note on the number of test samples would be helpful.
  6. [Sec. 5, Conclusion] The final sentence is grammatically incomplete: "an evaluation of the computational efficiencies of the models using different types of robots crucial" should read "...is crucial."
  7. [General] No data or code availability statement is provided; making the dataset and model implementations available (or at least stating their availability) would substantially strengthen the reproducibility of the empirical comparison.

Circularity Check

0 steps flagged · score 1.0 of 10

Empirical benchmark with no derivation chain to reduce; minor tuning and self-citation signals are not load-bearing circularity.

full rationale

The paper makes no first-principles derivation; its central claim is an empirical benchmark result on a MABI MAX 100 robot. Models in Table 1 (RNEA+MLP, DeLaN, LNN, LNN+MLP, LSTM, MLP) are trained on a training split and evaluated on held-out test data. The Newtonian baseline uses manufacturer-provided kinematics and inertia plus least-squares fitted friction coefficients (Eq. 13), but the prediction target is measured motor torque tau_u = psi_m*K_m*I_m, which is not defined by or equal to any fitted parameter. Thus the reported test RMSE is an independent performance measure rather than a fitted value renamed as a prediction. Friction coefficients and filter cutoffs are tuned on data, but this is standard hyperparameter/parameter fitting, not circularity. The paper's self-citations (Geist and Trimpe 2020, 2021; Rath et al. 2022) are survey and method references, and the central conclusion does not rest on an imported uniqueness theorem or ansatz. The abstract's causal phrasing, that Lagrangian networks are less effective because they do not explicitly model dissipative torques, is broader than what the experiment isolates: the comparison varies both the architecture and the availability of a strong nominal rigid-body model, and LNN+MLP adds an explicit dissipative network yet still underperforms on joints 1-4. That is a confound and an interpretation risk, but not a circular reduction. No equation in the paper is equivalent to its own input by construction, and the empirical comparison is self-contained against external baselines; hence the circularity score is low.

Assumptions & free parameters 3 free parameters · 5 assumptions · 0 invented entities

The central comparison rests on the accuracy of the nominal rigid-body model, the current-to-torque conversion, and the assumption that only velocity-dependent friction is missing. The paper states most of these assumptions clearly, but they are not independently validated against a joint-torque sensor.

free parameters (3)
  • joint friction coefficients theta_D,i = not reported (six values)
    Estimated by least squares on the training data in Section 4.2 and used in the RNEA+LQ baseline. They determine the friction residual used to compare models.
  • low-pass filter cutoff frequencies f_pass = [1.2, 0.3, 0.3, 0.7, 0.5, 0.7] Hz
    Selected in extensive preliminary tests in Section 4.1. They preprocess the motor torque estimates and could affect model comparisons.
  • trajectory Fourier coefficients a0, ak, bk = not reported
    Chosen manually in Section 4.1 to generate informative excitation trajectories. They shape the data distribution and are not fully specified in the paper.
assumptions (5)
  • domain assumption The recursive Newton-Euler algorithm with manufacturer-provided rigid-body parameters describes the frictionless component of motor torque accurately.
    Section 4.2, Eq. (13). The residual MLP is only asked to learn friction, so errors in the nominal model move into the MLP and change the comparison.
  • domain assumption Motor torque is obtained from motor current via tau_u = psi_m K_m I_m, and motor inertia affects only the corresponding joint because gear ratios are high.
    Section 4.2, Eq. (4) and Eq. (13), citing Khalil (2019). If gear-ratio coupling is non-negligible, the RNEA baseline is biased.
  • domain assumption Dissipative torques cannot be represented by the tested LNN and DeLaN forms without an explicit friction term or additional network.
    Section 3.3 and Figure 1. The paper gives experimental evidence and cites analytical friction potentials in Layton (2012) and Xiao et al. (2024).
  • domain assumption Friction on each joint depends only on that joint's velocity.
    Section 4.2: tau_D,i(q_dot_i) = -theta_D,i q_dot_i. The authors note that real friction also changes with normal force.
  • domain assumption Non-causal low-pass filtering of torque estimates does not remove relevant inverse-dynamics signals.
    Section 4.1. Cutoff frequencies are tuned in preliminary tests, and no independent validation of the filtering effect is provided.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Newtonian and Lagrangian Neural Networks: A Comparison Towards Efficient Inverse Dynamics Identification." pith.science (2026). https://pith.science/paper/Q36MVNGN

@misc{pith2026250617994,
  author       = {Pith},
  title        = {Pith review of: Newtonian and Lagrangian Neural Networks: A Comparison Towards Efficient Inverse Dynamics Identification},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Q36MVNGN}},
  note         = {Machine review of arXiv:2506.17994}
}
read the original abstract

Accurate inverse dynamics models are essential tools for controlling industrial robots. Recent research combines neural network regression with inverse dynamics formulations of the Newton-Euler and the Euler-Lagrange equations of motion, resulting in so-called Newtonian neural networks and Lagrangian neural networks, respectively. These physics-informed models seek to identify unknowns in the analytical equations from data. Despite their potential, current literature lacks guidance on choosing between Lagrangian and Newtonian networks. In this study, we show that when motor torques are estimated instead of directly measuring joint torques, Lagrangian networks prove less effective compared to Newtonian networks as they do not explicitly model dissipative torques. The performance of these models is compared to neural network regression on data of a MABI MAX 100 industrial robot.

Figures

Figures reproduced from arXiv: 2506.17994 by the authors.

Figure 1
Figure 1. The additive structure of inverse dynamics is visualized [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The inverse dynamics of a MABI MAX 100 shall be learned using physics-informed machine learning. In Newtonian [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. The motor torque τu is obtained as the sum of various analytical functions underlying the robot’s dynamics. 4.2 Deriving a model baseline via the RNEA To set up the Newtonian networks (10), we first derive the joint torques (1) via the RNEA using the kinematics and inertia parameters provided by the manufacturer. The resulting joint torque vector τRBD does not contain torques due to motor inertia. For the sake of si… view at source ↗
Figures from the paper (1 more)
Figure 5
Figure 5. Figure 5: Dissipative torque estimates over a test data trajectory [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

43 extracted references · 35 canonical work pages

  1. [1]

    , " * write output.state after.block = add.period write newline

    ENTRY address author booktitle chapter doi edition editor eid howpublished institution journal key month note number organization pages publisher school series title type url volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1 'mid.sent...

  2. [2]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize " " * FUNCT...

  3. [3]

    , " * write output.state after.block = add.period write newline

    ENTRY address author booktitle chapter edition editor howpublished institution journal key month note number organization pages publisher school series title type url volume year label extra.label sort.label short.list INTEGERS output.state before.all mid.sentence after.sentence after.block FUNCTION init.state.consts #0 'before.all := #1 'mid.sentence := ...

  4. [4]

    write newline

    " write newline "" before.all 'output.state := FUNCTION n.dashify 't := "" t empty not t #1 #1 substring "-" = t #1 #2 substring "--" = not "--" * t #2 global.max substring 't := t #1 #1 substring "-" = "-" * t #2 global.max substring 't := while if t #1 #1 substring * t #2 global.max substring 't := if while FUNCTION word.in bbl.in capitalize ":" * " " *...

  5. [5]

    An, C.H., Atkeson, C.G., and Hollerbach, J.M. (1985). Estimation of inertial parameters of rigid body links of manipulators. In 1985 24th IEEE Conference on Decision and Control, 990--995. IEEE

    work page 1985

  6. [6]

    Biewald, L. (2025). Experiment tracking with weights and biases. ://www.wandb.com/. Software available from wandb.com

    work page 2025

  7. [7]

    Cheng, C.A., Huang, H.P., Hsu, H.K., Lai, W.Z., and Cheng, C.C. (2015). Learning the inverse dynamics of robotic manipulators in structured reproducing kernel hilbert space. IEEE transactions on cybernetics, 46(7), 1691--1703

    work page 2015

  8. [8]

    Cranmer, M., Greydanus, S., Hoyer, S., Battaglia, P., Spergel, D., and Ho, S. (2020). Lagrangian neural networks. arXiv preprint arXiv:2003.04630

    arXiv 2020

Show all 43 references
  1. [9]

    Cuomo, S., Di Cola, V.S., Giampaolo, F., Rozza, G., Raissi, M., and Piccialli, F. (2022). Scientific machine learning through physics--informed neural networks: Where we are and what’s next. Journal of Scientific Computing, 92(3), 88

    2022

  2. [10]

    De La Cruz, J.S., Kuli \'c , D., and Owen, W. (2011). Online incremental learning of inverse dynamics incorporating prior knowledge. In Autonomous and Intelligent Systems: Second International Conference, AIS 2011, 167--176. Springer

    2011

  3. [11]

    Djeumou, F., Neary, C., Goubault, E., Putot, S., and Topcu, U. (2022). Neural networks with physics-informed architectures and constraints for dynamical systems modeling. In Learning for Dynamics and Control Conference, 263--277. PMLR

    2022

  4. [12]

    Featherstone, R. (2014). Rigid body dynamics algorithms

    2014

  5. [13]

    Finzi, M., Wang, K.A., and Wilson, A.G. (2020). Simplifying hamiltonian and lagrangian neural networks via explicit constraints. volume 33, 13880--13889. Curran Associates, Inc

    2020

  6. [14]

    and Trimpe, S

    Geist, A.R. and Trimpe, S. (2020). Learning constrained dynamics with gauss principle adhering gaussian processes. In Proceedings of the 2nd Conference on Learning for Dynamics and Control, volume 120 of Proceedings of Machine Learning Research (PMLR), 225--234. PMLR

    2020

  7. [15]

    and Trimpe, S

    Geist, A.R. and Trimpe, S. (2021). Structured learning of rigid-body dynamics: A survey and unified view from a robotics perspective. GAMM-Mitteilungen, 44(2), e202100009

    2021

  8. [16]

    Giacomuzzo, G., Dalla Libera, A., Carli, R., et al. (2023). Embedding the physics in black-box inverse dynamics identification: a comparison between gaussian processes and neural networks. IFAC-PapersOnLine, 56(2), 1584--1590

    2023

  9. [17]

    Greydanus, S., Dzamba, M., and Yosinski, J. (2019). Hamiltonian neural networks. In Advances in Neural Information Processing Systems, 15379--15389

    2019

  10. [18]

    Guo, X., Zhang, L., and Han, K. (2018). Dynamic parameter identification of robot manipulators based on the optimal excitation trajectory. 2018 IEEE International Conference on Mechatronics and Automation (ICMA), 2145--2150

    2018

  11. [19]

    Gupta, J.K., Menda, K., Manchester, Z., and Kochenderfer, M. (2020). Structured mechanical models for robot learning and control. In Learning for Dynamics and Control, 328--337

    2020

  12. [20]

    Hwangbo, J., Lee, J., Dosovitskiy, A., Bellicoso, D., Tsounis, V., and Koltun, V.e.a. (2019). Learning agile and dynamic motor skills for legged robots. Science Robotics, 4(26)

    2019

  13. [21]

    and Gans, N

    Jin, J. and Gans, N. (2015). Parameter identification for industrial robots with a fast and robust trajectory design approach. Robotics and Computer-Integrated Manufacturing, 31

    2015

  14. [22]

    Khalil, W. (2019). Modeling and Control of Manipulators - Part II: Dynamics and Control . Lecture

    2019

  15. [23]

    and Ba, J

    Kingma, D.P. and Ba, J. (2014). Adam: A method for stochastic optimization. arXiv preprint arXiv:1412.6980

    2014 arXiv

  16. [24]

    and Pucker, N

    Lang, C.B. and Pucker, N. (2016). Mathematische Methoden in Der Physik. Springer, Berlin, Germany, 3 edition

    2016

  17. [25]

    Layton, R.A. (2012). Principles of analytical system dynamics. Springer Science & Business Media

    2012

  18. [26]

    (2019 a )

    Lutter, M., Ritter, C., and Peters, J. (2019 a ). Deep lagrangian networks: Using physics as model prior for deep learning. In International Conference on Learning Representations (ICLR 2019). OpenReview. net

    2019

  19. [27]

    (2019 b )

    Lutter, M., Listmann, K., and Peters, J. (2019 b ). Deep lagrangian networks for end-to-end learning of energy-based control for under-actuated systems. In 2019 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 7718--7725. IEEE

    2019

  20. [28]

    and Peters, J

    Lutter, M. and Peters, J. (2023). Combining physics and deep learning to learn continuous-time dynamics models. The International Journal of Robotics Research, 42(3), 83--107

    2023

  21. [29]

    Lutter, M., Ritter, C., and Peters, J. (2018). Deep lagrangian networks: Using physics as model prior for deep learning. In International Conference on Learning Representations

    2018

  22. [30]

    Lutter, M., Silberbauer, J., Watson, J., and Peters, J. (2020). A differentiable newton euler algorithm for multi-body model learning. arXiv preprint arXiv:2010.09802

    2020 arXiv

  23. [31]

    Lutter, M., Silberbauer, J., Watson, J., and Peters, J. (2021). Differentiable physics models for real-world offline model-based reinforcement learning. In 2021 IEEE International Conference on Robotics and Automation (ICRA)

    2021

  24. [32]

    and Peters, J

    Nguyen-Tuong, D. and Peters, J. (2010). Using model knowledge for learning inverse dynamics. In 2010 IEEE International Conference on Robotics and Automation, 2677--2682

    2010

  25. [33]

    Nguyen-Tuong, D., Peters, J., Seeger, M., and Sch \"o lkopf, B. (2008). Learning inverse dynamics: a comparison. In European symposium on artificial neural networks, CONF

    2008

  26. [34]

    u r die Planung und Umsetzung wirtschaftlicher Roboterl\

    Pott, A. and Dietz, T. (2019). Industrielle Robotersysteme: Entscheiderwissen f\" u r die Planung und Umsetzung wirtschaftlicher Roboterl\" o sungen . Springer

    2019

  27. [35]

    Rath, L., Geist, A.R., and Trimpe, S. (2022). Using physics knowledge for learning rigid-body forward dynamics with gaussian process force priors. In Proceedings of the 5th Conference on Robot Learning, volume 164 of Proceedings of Machine Learning Research, 101--111. PMLR

    2022

  28. [36]

    Ross, D., Nemitz, M.P., and Stokes, A.A. (2016). Controlling and simulating soft robotic systems: Insights from a thermodynamic perspective. Soft Robotics, 3(4), 170--176

    2016

  29. [37]

    Siciliano, B., Sciavicco, L., Villani, L., and Oriolo, G. (2009). Modelling, planning and control. Advanced Textbooks in Control and Signal Processing. Springer,

    2009

  30. [38]

    Sutanto, G., Wang, A., Lin, Y., Mukadam, M., Sukhatme, G., Rai, A., and Meier, F. (2020). Encoding physical constraints in differentiable newton-euler algorithm. In Proceedings of the 2nd Conference on Learning for Dynamics and Control, volume 120 of Proceedings of Machine Lea...

    2020

  31. [39]

    Traversaro, S., Brossette, S., Escande, A., and Nori, F. (2016). Identification of fully physical consistent inertial parameters using optimization on manifolds. In 2016 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 5446--5451. IEEE

    2016

  32. [40]

    and Sch \"o lkopf, B

    Von Luxburg, U. and Sch \"o lkopf, B. (2011). Statistical learning theory: Models, concepts, and results. In Handbook of the History of Logic, volume 10, 651--706. Elsevier

    2011

  33. [41]

    Xiao, S., Zhang, J., and Tang, Y. (2024). Generalized lagrangian neural networks. arXiv preprint arXiv:2401.03728

    2024 arXiv

  34. [42]

    Yilmaz, N., Wu, J.Y., Kazanzides, P., and Tumerdem, U. (2020). Neural network based inverse dynamics identification and external force estimation on the da vinci research kit. In 2020 IEEE International Conference on Robotics and Automation (ICRA), 1387--1393. IEEE

    2020

  35. [43]

    and Belda, K

    Z \'a da, V. and Belda, K. (2017). Application of hamiltonian mechanics to control design for industrial robotic manipulators. In 2017 22nd International Conference on Methods and Models in Automation and Robotics (MMAR), 390--395

    2017

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.