pith. sign in

arxiv: 2606.05840 · v1 · pith:V6LQRYULnew · submitted 2026-06-04 · 📡 eess.SY · cs.RO· cs.SY

Amortized Nonlinear Model Predictive Control

Francesco Pillitteri , Alberto Bemporad This is my paper

Pith reviewed 2026-06-28 00:12 UTC · model grok-4.3

classification 📡 eess.SY cs.ROcs.SY
keywords nonlinear model predictive controlquadratic programming approximationinput-affine systemsneural network residual correctiondifferentiable optimizationamortized controlrobotic arm tracking
0
0 comments X

The pith

For input-affine nonlinear systems the optimal first control move of nonlinear MPC can be approximated by solving a state-dependent quadratic program whose cost parameters are output by a neural network.

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

The paper establishes that the repeated solution of a full nonlinear program in model predictive control can be replaced, for input-affine systems, by the solution of a quadratic program whose cost matrices and vectors are produced from the current state and reference. A single neural network is trained offline on data from an exact NLP solver; it learns only the residual corrections to a simple analytic baseline, and the resulting QP is solved exactly by a differentiable interior-point layer. The first control action therefore satisfies constraints by construction. On a three-link robotic arm the method delivers tracking performance close to the original NLP while running orders of magnitude faster.

Core claim

For the broad class of input-affine nonlinear systems the optimal control move obtained from a constrained nonlinear program can be closely approximated by a state-dependent quadratic program whose cost parameters depend on the current state and reference; these parameters are generated by a single-network residual-corrector architecture that starts from an analytic baseline and learns only the corrections needed to match the NLP solution, trained with a hybrid supervised and KKT-residual loss and solved via a differentiable interior-point layer.

What carries the argument

The single-network residual-corrector architecture that supplies state-dependent cost parameters to a quadratic program solved by a differentiable interior-point layer.

If this is right

  • The first control action is guaranteed to satisfy the original constraints because the quadratic program is solved exactly.
  • Computation time drops by orders of magnitude relative to repeated nonlinear-program solves.
  • Comparable Cartesian end-effector tracking performance is obtained on the three-link robotic arm example.
  • The approximation applies to any input-affine nonlinear system for which an NLP solver can generate training data.

Where Pith is reading between the lines

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

  • The same residual-corrector idea could be tried on systems that are not strictly input-affine by first embedding them in a larger input-affine representation.
  • If plant parameters drift, the network would require periodic retraining on new NLP data rather than online adaptation.
  • The differentiable QP layer already present could be used to propagate gradients through longer horizons or to co-optimize other controller parameters.

Load-bearing premise

A neural network trained offline on NLP solutions can learn corrections to an analytic baseline such that the first move of the resulting quadratic program matches the nonlinear program solution closely enough for the intended applications.

What would settle it

On the three-link planar robotic arm, the end-effector tracking error or constraint violation rate produced by the learned quadratic program exceeds that of the full nonlinear program by more than a small margin.

Figures

Figures reproduced from arXiv: 2606.05840 by Alberto Bemporad, Francesco Pillitteri.

Figure 1
Figure 1. Figure 1: Residual-corrector architecture for amortized input-affine MPC. An analytic baseline [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: End-effector trajectories in the Cartesian workspace for the amortized QP (solid blue), RTI (dashed orange), IPOPT (dashed green), and direct MLP [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Per-step solve time. The amortized policy (blue) is [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
read the original abstract

Nonlinear Model Predictive Control requires solving a constrained nonlinear program (NLP) in real-time at every sampling instant, a computational bottleneck that limits deployment on resource-constrained hardware or at high sampling rates. We address this challenge for the broad class of input-affine nonlinear systems to show that the optimal control move can be approximated by a state-dependent quadratic program (QP) whose cost parameters depend on the current state and reference. We propose a single-network residual-corrector architecture: a state-dependent analytic baseline provides initial QP parameters, and the network learns only the corrections needed to match the full NLP solution; the QP is solved by a differentiable interior-point layer, guaranteeing constraint satisfaction for the first control action. The network is trained offline on data generated by an NLP solver using a hybrid loss that combines supervised imitation and KKT-residual penalties. We validate the approach on a three-link planar robotic arm with Cartesian end-effector tracking, demonstrating orders-of-magnitude speedup over the NLP solver while maintaining comparable tracking performance.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper claims that for the broad class of input-affine nonlinear systems, the first optimal control action from an NMPC problem can be recovered by solving a state-dependent QP whose cost parameters are produced by a single residual-corrector neural network (analytic baseline plus learned corrections). The network is trained offline on NLP-generated data using a hybrid supervised imitation plus KKT-residual loss; the QP is solved via a differentiable interior-point layer to enforce feasibility. Validation on a three-link planar arm Cartesian-tracking task is reported to yield orders-of-magnitude speedup over direct NLP solution while preserving comparable tracking performance.

Significance. If the residual-corrector architecture generalizes, the method would offer a practical route to real-time NMPC on embedded hardware by amortizing the nonlinear solve into a fast QP whose parameters are state-dependent. The combination of an analytic baseline, differentiable optimization layer for constraint satisfaction, and hybrid loss is a constructive strength that could be reused in other control settings.

major comments (2)
  1. [Abstract] Abstract and validation results: the central claim that the approach works for the 'broad class of input-affine nonlinear systems' is load-bearing yet supported only by a single 3-link planar arm Cartesian-tracking example. No additional plants (underactuated, non-minimum-phase, or higher-dimensional input-affine systems) are tested, so there is no evidence that the learned correction map remains effective when the underlying NLP KKT conditions change structure.
  2. [Abstract] Abstract: the statements of 'orders-of-magnitude speedup' and 'comparable tracking performance' are presented without any quantitative metrics, error tables, or timing comparisons in the provided description, leaving the empirical support for the QP approximation unquantified.
minor comments (1)
  1. The description of the differentiable interior-point layer would benefit from an explicit citation or one-sentence recap of its constraint-handling properties to improve self-contained readability.

Simulated Author's Rebuttal

2 responses · 0 unresolved

Thank you for the constructive feedback. We address each major comment point-by-point below, with planned revisions where appropriate.

read point-by-point responses

  1. Referee: [Abstract] Abstract and validation results: the central claim that the approach works for the 'broad class of input-affine nonlinear systems' is load-bearing yet supported only by a single 3-link planar arm Cartesian-tracking example. No additional plants (underactuated, non-minimum-phase, or higher-dimensional input-affine systems) are tested, so there is no evidence that the learned correction map remains effective when the underlying NLP KKT conditions change structure.

    Authors: The residual-corrector architecture and hybrid loss are derived directly from the KKT stationarity conditions that arise for any input-affine nonlinear system; the QP structure follows from the affine dependence of the dynamics on the control. The three-link arm is a standard, fully coupled nonlinear benchmark within this class. We agree that additional plants would provide stronger empirical support. We will therefore revise the abstract to qualify the claim as applying to input-affine systems and add a dedicated paragraph in the discussion section on structural assumptions and potential limitations for other system classes. revision: partial

  2. Referee: [Abstract] Abstract: the statements of 'orders-of-magnitude speedup' and 'comparable tracking performance' are presented without any quantitative metrics, error tables, or timing comparisons in the provided description, leaving the empirical support for the QP approximation unquantified.

    Authors: The abstract is a high-level summary; the full manuscript reports concrete timing benchmarks (NLP vs. QP solve times) and tracking-error tables in the experimental evaluation. To make the abstract self-contained, we will insert specific quantitative statements such as 'achieving >100x speedup with Cartesian tracking error within 2% of the NLP baseline'. revision: yes

Circularity Check

0 steps flagged

No circularity: approximation learned from external NLP data via standard hybrid loss

full rationale

The paper's central claim is an empirical approximation of NLP solutions for input-affine systems via a residual-corrector network trained offline on solver-generated data. The analytic baseline plus learned corrections, differentiable IP layer for QP solving, and hybrid supervised+KKT loss are standard architectural and training choices that do not reduce the output to the inputs by definition or self-citation. No load-bearing step equates a 'prediction' to a fitted parameter or renames a known result. The single 3-link arm validation is an empirical demonstration, not a derivation that collapses into its own assumptions.

Axiom & Free-Parameter Ledger

1 free parameters · 2 axioms · 1 invented entities

The claim rests on the domain assumption that input-affine nonlinear systems admit a useful QP approximation whose parameters can be learned by a residual network, plus the practical assumption that offline NLP data plus hybrid loss suffice for training.

free parameters (1)
  • neural network weights
    Parameters of the residual-corrector network are fitted during offline training on NLP solver data.
axioms (2)
  • domain assumption Systems are input-affine nonlinear
    The QP approximation and network architecture are derived specifically for this class.
  • domain assumption Differentiable interior-point layer guarantees first-action constraint satisfaction
    Invoked to ensure feasibility of the approximated control move.
invented entities (1)
  • residual-corrector network no independent evidence
    purpose: Learns only corrections to an analytic baseline for QP cost parameters
    New architecture introduced to combine analytic speed with learned accuracy.

pith-pipeline@v0.9.1-grok · 5700 in / 1248 out tokens · 29654 ms · 2026-06-28T00:12:29.299280+00:00 · methodology

discussion (0)

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

Reference graph

Works this paper leans on

20 extracted references · 3 canonical work pages · 1 internal anchor

  1. [1]

    J. B. Rawlings, D. Q. Mayne, and M. Diehl,Model Predictive Control: Theory, Computation, and Design, 2nd ed. Nob Hill Publishing, 2017

    2017

  2. [2]

    qpoases: A parametric active-set algorithm for quadratic program- ming,

    H. J. Ferreau, C. Kirches, A. Potschka, H. G. Bock, and M. Diehl, “qpoases: A parametric active-set algorithm for quadratic program- ming,”Mathematical Programming Computation, vol. 6, no. 4, pp. 327–363, 2014. TABLE II CLOSED-LOOP COST AND CONSTRAINT SATISFACTION(100 SCENARIOS).J cl RATIOS RELATIVE TOIPOPT. Metric Amort. RTI IPOPT MLP MeanJ cl 3642.5 427...

    2014

  3. [3]

    ODYS QP Solver,

    G. Cimini, A. Bemporad, and D. Bernardini, “ODYS QP Solver,” ODYS S.r.l. (https://odys.it/qp), Sep. 2017

    2017

  4. [4]

    Hpipm: a high-performance quadratic programming framework for model predictive control,

    G. Frison and M. Diehl, “Hpipm: a high-performance quadratic programming framework for model predictive control,”IFAC- PapersOnLine, vol. 53, no. 2, pp. 6563–6569, 2020

    2020

  5. [5]

    The ex- plicit linear quadratic regulator for constrained systems,

    A. Bemporad, M. Morari, V . Dua, and E. N. Pistikopoulos, “The ex- plicit linear quadratic regulator for constrained systems,”Automatica, vol. 38, no. 1, pp. 3–20, 2002

    2002

  6. [6]

    A real-time iteration scheme for nonlinear optimization in optimal feedback control,

    M. Diehl, H. G. Bock, and J. P. Schl ¨oder, “A real-time iteration scheme for nonlinear optimization in optimal feedback control,”SIAM Journal on Control and Optimization, vol. 43, no. 5, pp. 1714–1736, 2005

    2005

  7. [7]

    From linear to nonlinear MPC: Bridging the gap via the real-time iteration,

    S. Gros, M. Zanon, R. Quirynen, A. Bemporad, and M. Diehl, “From linear to nonlinear MPC: Bridging the gap via the real-time iteration,” International Journal of Control, vol. 93, no. 1, pp. 62–80, 2020

    2020

  8. [8]

    Tutorial on amortized optimization,

    B. Amos, “Tutorial on amortized optimization,”Foundations and Trends in Machine Learning, vol. 16, no. 5, pp. 592–732, 2023

    2023

  9. [9]

    Learning Lyapunov terminal costs from data for complexity reduction in nonlinear model predictive control,

    S. Abdufattokhov, M. Zanon, and A. Bemporad, “Learning Lyapunov terminal costs from data for complexity reduction in nonlinear model predictive control,”Int. Journal of Robust and Nonlinear Control, vol. 34, no. 13, pp. 8676–8691, 2024

    2024

  10. [10]

    Learning Parametric Convex Functions,

    M. Schaller, A. Bemporad, and S. Boyd, “Learning Parametric Convex Functions,” 2025, http://arxiv.org/abs/2506.04183

    work page arXiv 2025

  11. [11]

    A predictive safety filter for learning-based control of constrained nonlinear dynamical systems,

    K. P. Wabersich and M. N. Zeilinger, “A predictive safety filter for learning-based control of constrained nonlinear dynamical systems,” Automatica, vol. 129, p. 109597, 2021

    2021

  12. [12]

    Learning to warm-start fixed-point optimization algorithms,

    R. Sambharya, G. Hall, B. Amos, and B. Stellato, “Learning to warm-start fixed-point optimization algorithms,”J. Mach. Learn. Res., vol. 25, no. 1, Jan. 2024

    2024

  13. [13]

    OptNet: Differentiable optimization as a layer in neural networks,

    B. Amos and J. Z. Kolter, “OptNet: Differentiable optimization as a layer in neural networks,” inProc. International Conference on Machine Learning (ICML), 2017, pp. 136–145

    2017

  14. [14]

    On the implementation of an interior- point filter line-search algorithm for large-scale nonlinear program- ming,

    A. W ¨achter and L. T. Biegler, “On the implementation of an interior- point filter line-search algorithm for large-scale nonlinear program- ming,”Mathematical programming, vol. 106, no. 1, pp. 25–57, 2006

    2006

  15. [15]

    A special Newton-type optimization method,

    A. Fischer, “A special Newton-type optimization method,”Optimiza- tion, vol. 24, no. 3–4, pp. 269–284, 1992

    1992

  16. [16]

    Self-supervised learning of iterative solvers for constrained optimization,

    L. L ¨uken and S. Lucia, “Self-supervised learning of iterative solvers for constrained optimization,” 2025. [Online]. Available: https://arxiv.org/abs/2409.08066

    work page arXiv 2025

  17. [17]

    On the differentiability of the primal- dual interior-point method,

    K. Tracy and Z. Manchester, “On the differentiability of the primal- dual interior-point method,” 2024

    2024

  18. [18]

    Adam: A Method for Stochastic Optimization

    D. P. Kingma and J. Ba, “Adam: A method for stochastic optimiza- tion,”arXiv preprint 1412.6980, 2014

    work page internal anchor Pith review Pith/arXiv arXiv 2014

  19. [19]

    JAX: composable transformations of Python+NumPy programs,

    J. Bradbury, R. Frostig, P. Hawkins, M. J. Johnson, C. Leary, D. Maclaurin, G. Necula, A. Paszke, J. VanderPlas, S. Wanderman- Milne, and Q. Zhang, “JAX: composable transformations of Python+NumPy programs,” 2018. [Online]. Available: http://github. com/jax-ml/jax

    2018

  20. [20]

    acados – a modular open-source framework for fast embedded optimal control,

    R. Verschueren, G. Frison, D. Kouzoupis, J. Frey, N. v. Duijkeren, A. Zanelli, B. Novoselnik, T. Albin, R. Quirynen, and M. Diehl, “acados – a modular open-source framework for fast embedded optimal control,”Mathematical Programming Computation, vol. 14, no. 1, pp. 147–183, 2022

    2022