Pith. sign in

REVIEW 4 major objections 6 minor 35 references

Learning-Based Data-Assisted Port-Hamiltonian Control for Free-Floating Space Manipulators

T0 review · 4 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read A virtual-port split lets a learning module improve control of a free-floating space manipulator while a Lyapunov proof guarantees the learning cannot destabilize the primary dynamics.

desk verdict A plausible hybrid pH control architecture with a solid case study, but the headline stability guarantee rests on an empirical error bound and needs either a proof or a more honest claim. read the letter →

arxiv 2509.09563 v1 pith:Y37WAETX submitted 2025-09-11 eess.SY cs.SY

Mostafa Eslami , Maryam Babazadeh This is my paper
classification eess.SYcs.SY
keywords data-assistedcontrolport-Hamiltonianfree-floatingspacemanipulatorslidingmodedeeplearningLyapunovstabilitytime-scaleseparationcollisionavoidance
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

Space robots must control a manipulator whose base floats freely; any arm motion pushes the base, so the dynamics are coupled, path-dependent, and partly uncertain. This paper proposes a control architecture that separates the well-understood, energy-conserving part of the dynamics from everything uncertain—actuation, dissipation, disturbances—using a new physical observable called the virtual port. The known part is stabilized by a sliding-mode law; the uncertain part is handled by a fast decentralized integrator and an online neural network that learns the port-to-input mapping. A Lyapunov argument shows that despite bounded learning errors and delays, the overall closed loop stays stable, and a simulation of a planar two-link free-floating manipulator reports a 77.5% reduction in orientation tracking error compared with a purely model-based controller.

What carries the argument

The virtual port variable Π (specialized to τ) is the observable that mediates all power exchange between the conservative dynamics and the environment, defined by the identity ˙p + ∂H/∂q = τ. This identity splits the system into a left-hand side whose structure is exactly known and a right-hand side that is a static map from desired port force to actuator inputs, dissipation, and disturbances. The control stack is a sliding-mode law on the left, a decentralized integrator on the right, and a physics-informed neural network that learns the right-hand side matrices, using softplus activation to keep dissipation positive definite. The time-scale separation condition σ_RHS > T_d λ_LHS, expresse

What would settle it

Run the same controller in a simulation where the learning network is deliberately hampered (e.g., slow optimizer or adversarial features) so that the observed port error ∥e_τ∥ exceeds χ while the LHS is operating; if the Lyapunov derivative ˙V becomes positive and the sliding surface norm grows, the claim of learning-safe stability under bounded errors is falsified. Alternatively, directly compute the time-scale separation using the actual linearized model eigenvalues in Eqs. (36)-(37) and check whether the sign inconsistency in the Jacobian term changes the decay-rate inequality.

Watch

Extended reading notes

Core claim

The central claim is that rewriting the port-Hamiltonian equations as ˙p + ∂H/∂q = τ makes it possible to confine learning to the right-hand side only. The left-hand side is the fully known conservative dynamics of the reduced-order free-floating manipulator, derived through the Generalized Jacobian Matrix; the right-hand side is an algebraic relation τ = −D̃(q,p) ˙q + B̃(q,p)u + d(t). A sliding-mode controller stabilizes the left side with a Lyapunov proof, while a high-gain decentralized integral controller plus a neural network estimates B̃ and D̃ on the right side. A strict time-scale separation—the right-side error decays much faster than the left-side sliding dynamics—ensures that even

Load-bearing premise

The stability proof assumes the RHS tracking error e_τ stays below a known bound χ; the paper obtains χ from a moving average of simulation samples rather than a derived worst-case guarantee, and the strict time-scale separation that keeps e_τ small is established only for the linearized RHS model.

Editorial extensions

If this is right

  • The same architecture transfers to any robotic system whose conservative dynamics are known: the learning module only has to approximate a low-dimensional algebraic map, so data requirements drop sharply.
  • Because the integrator gains are updated online from the learned matrices, the controller remains stable even when the plant changes, as long as the updated gains satisfy the time-scale separation inequality.
  • The null-space potential for collision avoidance can be added to the sliding-mode reference velocity without changing the primary orientation objective, so secondary safety tasks do not sacrifice tracking.
  • The paper's Lyapunov bound (eq. 48) gives an explicit, checkable condition for selecting the RHS integrator gain once a bound on learning error is known.

Reading between the lines

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

  • If the virtual-port decomposition is as generic as claimed, it offers a template for other momentum-dominated robots (underwater vehicles, legged robots, free-flying drones): find a conserved quantity that produces a known left-hand side, and relegate all interaction effects to the right-hand side for learning.
  • The paper's χ bound is taken from a moving average of simulation samples; replacing it with a worst-case guaranteed bound (e.g., from Lipschitz constants of the learned matrices and the integrator's L1 gain) would turn the stability theorem into a fully a priori guarantee.
  • A testable prediction: if a high-frequency disturbance is correlated with the control input, the learned matrices B̂ and D̂ should absorb its average effect, and the tracking error should remain small because the integral loop compensates the residual; injecting such a disturbance in a fresh simulation would confirm or refute this.
  • Because the stability proof depends on strict time-scale separation, the method may lose its safety margin in higher-dimensional arms where actuator bandwidths approach the LHS dynamics; a 7-DOF case study with realistic actuator limits would reveal whether condition (48) remains practical.
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 / 6 minor

Summary. The paper proposes a learning-based data-assisted control architecture within the port-Hamiltonian (pH) framework. It splits the pH dynamics into a conservative left-hand side (LHS) and an algebraic right-hand side (RHS) that collects dissipation, actuation, and external disturbances through a 'virtual port' variable. For a free-floating space manipulator, the authors derive a reduced-order pH model using the generalized Jacobian matrix, design a sliding-mode controller for end-effector orientation tracking on the LHS, and add an inner decentralized integral loop plus a physics-informed DNN on the RHS to learn the uncertain mapping. Stability is argued via a Lyapunov analysis assuming a bounded RHS tracking error, and simulations on a 2-DOF planar space manipulator show improved tracking and disturbance rejection compared to a purely model-based controller.

Significance. If the stability claims can be made fully rigorous, the architecture is a useful template for safely integrating learning into passivity-based robotic control: it preserves the Hamiltonian structure, confines learning to a low-dimensional algebraic mapping, and provides a transparent separation of model-based and data-driven components. The LHS sliding-mode Lyapunov argument is standard and, given a certified bound on the RHS error, the proof in Eqs. (42)-(44) is sound. The application to free-floating space manipulators is relevant and the simulation comparison is encouraging. However, the current manuscript does not establish a certified bound on the RHS error, and a sign inconsistency in the RHS linearization undermines the formal gain-design conditions. These issues must be resolved before the central claims can be accepted.

major comments (4)
  1. [Section 3.4, Eqs. (40)-(44), Fig. 2b] The stability proof is conditional on ∥eτ∥ ≤ χ, but χ is not a derived worst-case bound; it is introduced as a 'moving average of 1k samples' in Fig. 2b. The RHS analysis in Eqs. (36)-(38) is linearized and local, and does not account for DNN parameter-update transients, unmodeled high-frequency disturbances d_HF, or parameter drift. Consequently the abstract's claims of 'globally stable' and 'learning cannot destabilize the primary dynamics' are not supported by the formal analysis. Please either provide a certified χ (e.g., from known uncertainty bounds and integral gains using an input-to-state-stability estimate) or reframe the contribution as conditional stability with empirical validation.
  2. [Section 3.3, Eq. (36)] There is a sign inconsistency in the linearized RHS error. Since eτ = τc^r - (B(x)u - D(x)qdot + d), the Taylor expansion gives δeτ = δτc^r - δe - δd - (B0 δu - D0 δqdot + ∂(B(x)u0 - D(x)qdot0)/∂x δx). With J0 defined as ∂(D(x)qdot0 - B(x)u0)/∂x, the Jacobian term inside the parentheses must be -J0 δx, not +J0 δx. The sign error propagates into Eq. (39) and the uncertainty factor β, so the DIC/gain-design inequalities built on this expression are not trustworthy. Correct the derivation and re-evaluate the gain conditions.
  3. [Section 3.3, after Eq. (33)] No explicit nonlinear RHS control law is given. The text defines only a perturbative controller δu(s) = K(s) δeτ(s); it never states how the actual input u(t) is assembled from the learned matrices B(x), D(x), the integral correction, and the requested port torque τc^r. The update rules for the integrator gains ki and for c are also not written down. This makes the algorithm non-reproducible and leaves it unclear whether the DNN output directly affects the control signal. Please provide the complete RHS control law and the online parameter-update rules.
  4. [Section 2.2, Eq. (3)] The 'new physical observable' Π is defined as exactly the residual of the pH equation: Π = -R∇H + g u. This is an algebraic identity that holds for any pH system, not an independently measured or newly discovered physical quantity. The decomposition is still a potentially useful control-oriented reformulation, but the paper should be careful not to overstate it as a new physical observable. This does not invalidate the control architecture, but the framing should be adjusted.
minor comments (6)
  1. [Section 3.2.2, Eqs. (29)-(31)] The eigenvalue notation is unclear: the over/under bars are lost in typesetting. Please define λ_min and λ_max explicitly for M and Kd, and check the direction of the inequalities in Eq. (30).
  2. [Section 3.3, Eq. (45)] The quantities λ_LHS^s and T_d appear without definition. Please define the LHS sliding-mode decay rate and the time-scale-separation factor before using them in the design inequality.
  3. [Section 3.3, Eq. (38)] The summation index p conflicts with the generalized momentum p used elsewhere. Use a different index, e.g., k or r.
  4. [Section 4, simulation setup] Please state explicitly whether the discontinuous term -χ s/∥s∥ in Eq. (40) or a smooth approximation (e.g., tanh(s/ε)) was used in the simulations. The parameter ε = 0.2 is listed but no corresponding equation is given.
  5. [Figure 3] The comparison label 'DAC OFF' is described as 'Model-Based'. Please clarify what the model-based baseline is: the same LHS sliding-mode controller without the RHS learning and integral loop, or a different law?
  6. [References] References [12] and [14] are self-citations to arXiv/preprint material. If the journal requires archival sources, please replace or supplement them with peer-reviewed versions.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the DAC-pH split is a definitional decomposition, not a derived prediction, and the empirical χ is a conditional-certificate gap rather than a circular inference.

full rationale

The port variable Π in Eq. (3) is defined as the rearranged RHS of the pH equation, so the statement that Π collects all non-conservative effects is true by definition rather than by discovery. However, the paper does not derive a falsifiable physical prediction from this definition; the definition is the intended modeling decomposition, and the downstream control design and simulations stand on their own. The Lyapunov analysis in Eqs. (40)–(44) is a valid conditional argument: if ∥eτ∥ ≤ χ, then V̇ ≤ −sᵀK_d s. The robustness bound χ is described as 'empirically calculated' and appears in Fig. 2b as a moving average of 1k samples, so the claim of 'global stability with explicit bounds' is weaker than stated—it is a certificate that must be checked online rather than an a priori guarantee. That is a rigor/correctness concern, not a circular dependency, because the proof does not reduce to the data used to set χ. Self-citations ([12], [13], [14], [24], [35]) provide terminology, standard sliding-mode design, and DIC conditions; none imports an unverified conclusion specific to this plant or makes the main result equivalent to its inputs. The simulation identification of B and D is falsifiable against the hidden true matrices in Eqs. (49)–(50).

Assumptions & free parameters 5 free parameters · 6 assumptions · 1 invented entities

The central claim rests on the pH decomposition being physically meaningful, a bounded learning error that is only simulated, a linearized RHS model for gain selection, and a set of hand-chosen simulation and controller parameters. The virtual port is an algebraic construct, not an independently observable quantity.

free parameters (5)
  • Robustness bound chi = moving average of observed ||e_tau|| over 1000 samples
    Section 3.4 requires ||e_tau|| <= chi for Lyapunov stability, but chi is taken from simulation data rather than an a priori error bound.
  • Controller gains Kd, Lambda, epsilon, eta, Td = Kd=0.5I, Lambda=I, epsilon=0.2, eta=0.1, Td=2
    Hand-selected in Section 4; the stability inequalities require these to satisfy the time-scale separation condition.
  • Integral gain parameter c and k_i = updated online via Eq. (48) with an estimate of beta
    The gains are adjusted based on the learned model and the estimated multiplicative uncertainty beta.
  • DNN hyperparameters = two hidden layers, 16 and 8 neurons, ADAM beta1=0.9, beta2=0.999
    Chosen to match the simulated uncertainty complexity; no systematic tuning or sensitivity analysis is reported.
  • Simulation uncertainty coefficients in Delta D and Delta B = 0.02, 0.05, 0.4, 0.4, -0.05, 0.1, -0.1, 0.05, 0.15, etc.
    These define the ground truth for the simulation; they are ad hoc choices, not measured from a real system.
assumptions (6)
  • standard math Standard port-Hamiltonian state-space model with canonical interconnection matrix J(x), Eq. (1)-(2)
    The entire decomposition starts from this representation and the canonical mechanical J.
  • domain assumption Conservation of linear and angular momentum for the free-floating system, Eq. (8), and invertibility of H_q0, Eq. (10)
    The reduced-order model and GJM rely on zero initial momentum and non-singular H_q0.
  • domain assumption The LHS is exactly known; only the RHS is uncertain
    Section 3.3 states 'this work assumes it is known', so parametric uncertainty in the LHS is out of scope.
  • domain assumption All external interactions are modeled as disturbances; interactions imposing new kinematic constraints are out of scope
    Section 2 explicitly excludes interactions that would alter the LHS Hamiltonian.
  • ad hoc to paper Bounded learning error ||e_tau|| <= chi and strict time-scale separation sigma_RHS > Td lambda_LHS
    These are the central assumptions in Section 3.4; no proof is given that the DNN-in-the-loop actually maintains these bounds.
  • ad hoc to paper The RHS error dynamics can be linearized around a nominal trajectory and remain valid for gain design
    Eq. (36)-(37) build the integral controller and stability conditions on a first-order Taylor expansion.
invented entities (1)
  • Virtual port Pi
    purpose: A single interface through which all non-conservative effects enter the conservative Hamiltonian dynamics
    Defined in Eq. (3) as the RHS residual of the pH equation; there is no independent measurement or falsifiable prediction outside the model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning-Based Data-Assisted Port-Hamiltonian Control for Free-Floating Space Manipulators." pith.science (2026). https://pith.science/paper/Y37WAETX

@misc{pith2026250909563,
  author       = {Pith},
  title        = {Pith review of: Learning-Based Data-Assisted Port-Hamiltonian Control for Free-Floating Space Manipulators},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/Y37WAETX}},
  note         = {Machine review of arXiv:2509.09563}
}
read the original abstract

A generic data-assisted control architecture within the port-Hamiltonian framework is proposed, introducing a physically meaningful observable that links conservative dynamics to all actuation, dissipation, and disturbance channels. A robust, model-based controller combined with a high-gain decentralized integrator establishes large robustness margins and strict time-scale separation, ensuring that subsequent learning cannot destabilize the primary dynamics. Learning, selected for its generalizability, is then applied to capture complex, unmodeled effects, despite inherent delay and transient error during adaptation. Formal Lyapunov analysis with explicit stability bounds guarantees convergence under bounded learning errors. The structured design confines learning to the simplest part of the dynamics, enhancing data efficiency while preserving physical interpretability. The approach is generic, with a free-floating space manipulator orientation control task, including integrated null-space collision avoidance, serving as a case study to demonstrate robust tracking performance and applicability to broader robotic domains.

Figures

Figures reproduced from arXiv: 2509.09563 by the authors.

Figure 1
Figure 1. DNN Performance in identification of uncertain RHS [PITH_FULL_IMAGE:figures/full_fig_p012_1.png] view at source ↗
Figure 2
Figure 2. Detailed DNN Performance, RHS controller gain var [PITH_FULL_IMAGE:figures/full_fig_p013_2.png] view at source ↗
Figure 3
Figure 3. DAC Performance comparison with Model-Based Cont [PITH_FULL_IMAGE:figures/full_fig_p013_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

35 extracted references · 2 linked inside Pith

  1. [14]

    On the generaliza tion of data-assisted control in Hamiltonian systems (dac-ph)

    Mostafa Eslami and Maryam Babazadeh. On the generaliza tion of data-assisted control in Hamiltonian systems (dac-ph). arXiv preprint arXiv:2506.07079 , 2025

    arXiv 2025

  2. [1]

    Tutorial review on space manipulators for s pace debris mitigation

    Alex Ellery. Tutorial review on space manipulators for s pace debris mitigation. Robotics, 8(2):34, 2019. 13 Learning-based Data-Assisted Port-Hamiltonian Control f or Free-Floating Space Manipulators

    2019

  3. [2]

    Robotic manipulation and capture in space: A survey

    Evangelos Papadopoulos, Farhad Aghili, Ou Ma, and Rober to Lampariello. Robotic manipulation and capture in space: A survey. Frontiers in Robotics and AI, 8:686723, 2021

    2021

  4. [3]

    Spacecraft-mounted robotics

    Panagiotis Tsiotras, Matthew King-Smith, and Lorenzo T icozzi. Spacecraft-mounted robotics. Annual Review of Control, Robotics, and Autonomous Systems , 6(1):335–362, 2023

    2023

  5. [4]

    A comprehensive surve y of space robotic manipulators for on-orbit servicing

    Mohammad Alizadeh and Zheng H Zhu. A comprehensive surve y of space robotic manipulators for on-orbit servicing. Frontiers in Robotics and AI, 11:1470950, 2024

    2024

  6. [5]

    Hamiltonian systems: an introduct ory survey

    Arjan V an Der Schaft. Hamiltonian systems: an introduct ory survey. In Proceedings of the international congress of mathematicians, volume 3, pages 1339–1365. Marta Sanz-Sole, Javier Soria, Juan Luis V erona, Joan V erdura, Madrid, Spain, 2006

    2006

  7. [6]

    Modeling and control of complex physical systems: the Hamiltonian approach

    Vincent Duindam, Alessandro Macchelli, Stefano Strami gioli, and Herman Bruyninckx. Modeling and control of complex physical systems: the Hamiltonian approach . Springer Science & Business Media, 2009

    2009

  8. [7]

    Port-Hamiltonian neural ODE networks on Lie groups for robot dynamics learning and control

    Thai Duong, Abdullah Altawaitan, Jason Stanley, and Nik olay Atanasov. Port-Hamiltonian neural ODE networks on Lie groups for robot dynamics learning and control. IEEE Transactions on Robotics, 40:3695–3715, 2024

    2024

Show all 35 references
  1. [8]

    Nageshrao, and Gabriel A

    Olivier Sprangers, Robert Babuška, Subramanya P . Nageshrao, and Gabriel A. D. Lopes. Reinforcement learning for Hamiltonian systems. IEEE Transactions on Cybernetics, 45(5):1017–1027, 2015

    2015

  2. [9]

    Nageshrao, G.A.D

    S.P . Nageshrao, G.A.D. Lopes, D. Jeltsema, and R. Babušk a. Passivity-based reinforcement learning control of a 2-dof manipulator arm. Mechatronics, 24(8):1001–1007, 2014

    2014

  3. [10]

    Control of interactive robotic interfaces: A Hamilto- nian approach

    Cristian Secchi, Cesare Fantuzzi, and Stefano Stramig ioli. Control of interactive robotic interfaces: A Hamilto- nian approach. Springer, 2007

    2007

  4. [11]

    Hamiltonian based control for human-r obot team interaction

    Martin Angerer, Selma Musi ´c, and Sandra Hirche. Hamiltonian based control for human-r obot team interaction. In 2017 IEEE International Conference on Robotics and Automat ion (ICRA), pages 2292–2299. IEEE, 2017

    2017

  5. [12]

    Data-assisted co ntrol: A framework development by exploiting nasa generic transport platform

    Mostafa Eslami and Afshin Banazadeh. Data-assisted co ntrol: A framework development by exploiting nasa generic transport platform. International Journal of Robust and Nonlinear Control , 34(3):1898–1920, 2024

    1920

  6. [13]

    Sequential data- assisted control in flight

    Mostafa Eslami and Afshin Banazadeh. Sequential data- assisted control in flight. arXiv, 2023

    2023

  7. [15]

    Dynamic Singularities in Free-floating Space Manipulators , pages 77–100

    Evangelos Papadopoulos and Steven Dubowsky. Dynamic Singularities in Free-floating Space Manipulators , pages 77–100. Springer US, Boston, MA, 1993

    1993

  8. [16]

    Dynamic control of a space robot system with no thrust jets controlled base

    Y angsheng Xu and Heung-Y eung Shum. Dynamic control of a space robot system with no thrust jets controlled base. Technical report, 1991

    1991

  9. [17]

    Nonlinear control of space manipulators with model uncertainty

    Manoj Mittal, C-H Chuang, and Jer-Nan Juang. Nonlinear control of space manipulators with model uncertainty. In Guidance, Navigation, and Control Conference , page 3654, 1994

    1994

  10. [18]

    Nonholonomic behavior in fr ee-floating space manipulators and its utilization

    Evangelos G Papadopoulos. Nonholonomic behavior in fr ee-floating space manipulators and its utilization. In Nonholonomic Motion Planning , pages 423–445. Springer, 1993

    1993

  11. [19]

    Robot modeling and control, volume 3

    Mark W Spong, Seth Hutchinson, Mathukumalli Vidyasaga r, et al. Robot modeling and control, volume 3. Wiley New Y ork, 2006

    2006

  12. [20]

    On cartesian motions with singularities avoidance for free-floating space robots

    Kostas Nanos and Evangelos Papadopoulos. On cartesian motions with singularities avoidance for free-floating space robots. In 2012 IEEE International Conference on Robotics and Automat ion, pages 5398–5403. IEEE, 2012

    2012

  13. [21]

    Hybrid analytical and data-driven modeling for feed-forward robot control

    René Felix Reinhart, Zeeshan Shareef, and Jochen Jakob Steil. Hybrid analytical and data-driven modeling for feed-forward robot control. Sensors, 17(2), 2017

    2017

  14. [22]

    Umetani and K

    Y . Umetani and K. Y oshida. Resolved motion rate controlof space manipulators with generalized jacobian matrix. IEEE Transactions on Robotics and Automation , 5(3):303–314, 1989

    1989

  15. [23]

    Adaptive control of robot manipulators wi th uncertain kinematics and dynamics

    Hanlei Wang. Adaptive control of robot manipulators wi th uncertain kinematics and dynamics. IEEE Transac- tions on Automatic Control, 62(2):948–954, 2016

    2016

  16. [24]

    Integrity of lti time -delay systems

    Mostafa Eslami and Amin Nobakhti. Integrity of lti time -delay systems. IEEE Transactions on Automatic Control, 61(2):562–567, 2016

    2016

  17. [25]

    Physics- informed machine learning

    George Em Karniadakis, Ioannis G Kevrekidis, Lu Lu, Par is Perdikaris, Sifan Wang, and Liu Y ang. Physics- informed machine learning. Nature Reviews Physics, 3(6):422–440, 2021. 14 Learning-based Data-Assisted Port-Hamiltonian Control f or Free-Floating Space Manipulators

    2021

  18. [26]

    Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involvi ng nonlinear partial differential equations

    Maziar Raissi, Paris Perdikaris, and George E Karniada kis. Physics-informed neural networks: A deep learning framework for solving forward and inverse problems involvi ng nonlinear partial differential equations. Journal of Computational physics , 378:686–707, 2019

    2019

  19. [27]

    Enforcing analytic constraints in neural networks emulating physical systems

    Tom Beucler, Michael Pritchard, Stephan Rasp, Jordan O tt, Pierre Baldi, and Pierre Gentine. Enforcing analytic constraints in neural networks emulating physical systems . Physical review letters, 126(9):098302, 2021

    2021

  20. [28]

    Deep lagrangian networks: Using physics as model prior for deep learning

    Michael Lutter, Christian Ritter, and Jan Peters. Deep lagrangian networks: Using physics as model prior for deep learning. arXiv preprint arXiv:1907.04490 , 2019

    1907 arXiv

  21. [29]

    Ha miltonian neural networks

    Samuel Greydanus, Misko Dzamba, and Jason Y osinski. Ha miltonian neural networks. Advances in neural information processing systems, 32, 2019

    2019

  22. [30]

    Physics-informed neural networks for inverse problems in nano-optics and metamaterials

    Y uyao Chen, Lu Lu, George Em Karniadakis, and Luca Dal Negro. Physics-informed neural networks for inverse problems in nano-optics and metamaterials. Optics express, 28(8):11618–11633, 2020

    2020

  23. [31]

    James Blaise and Michael C. F. Bazzocchi. Space manipul ator collision avoidance using a deep reinforcement learning control. Aerospace, 10(9), 2023

    2023

  24. [32]

    Slotine and Weiping Li

    Jean-Jacques E. Slotine and Weiping Li. On the adaptive control of robot manipulators. The International Journal of Robotics Research, 6(3):49–59, 1987

    1987

  25. [33]

    O. Khatib. Real-time obstacle avoidance for manipulat ors and mobile robots. In Proceedings. 1985 IEEE Inter- national Conference on Robotics and Automation , volume 2, pages 500–505, 1985

    1985

  26. [34]

    System identification

    Lennart Ljung. System identification. In Signal analysis and prediction , pages 163–173. Springer, 1998

    1998

  27. [35]

    Ro bust modeling, sliding-mode controller, and simu- lation of an underactuated rov under parametric uncertaint ies and disturbances

    Mostafa Eslami, Cheng Siong Chin, and Amin Nobakhti. Ro bust modeling, sliding-mode controller, and simu- lation of an underactuated rov under parametric uncertaint ies and disturbances. Journal of Marine Science and Application, 18(2):213–227, 2019. 15

    2019

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.