Pith. sign in

REVIEW 4 major objections 6 minor 32 references

Neural Co-state Regulator: A Data-Driven Paradigm for Real-time Optimal Control with Input Constraints

T0 review · 4 major / 6 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read The paper claims that a neural network trained without expert trajectories or labels can predict Pontryagin co-states and, combined with a small quadratic program, deliver constrained optimal feedback control that matches or outperforms…

desk verdict The NCR's core assumption that co-state optimality is constraint-independent is false, so the paper's main claim does not hold; the unsupervised idea is worth exploring but needs major revision. read the letter →

arxiv 2507.12259 v1 pith:M6HATDDT submitted 2025-07-16 eess.SY cs.SY

classification eess.SYcs.SY
keywords neuralco-stateregulatorPontryagin'sminimumprincipleinputconstraintsquadraticprogrammodelpredictivecontrolunsupervisedlearningreal-timeoptimalunicycle
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 is trying to establish that a simple feedforward neural network, trained without any expert solver or labeled optimal trajectories, can act as a real-time optimal feedback controller for constrained nonlinear systems. The network predicts the co-state trajectory prescribed by Pontryagin's Minimum Principle for a given state, and a quadratic program enforces the input constraints; the authors call the combination a neural co-state regulator (NCR). On a unicycle model, they report that the NCR matches or beats a nonlinear MPC solver in convergence error and produces smoother input trajectories, including for initial states outside the training distribution, while running about two orders of magnitude faster. The significance, if true, is that the expensive nonlinear programming step of MPC can be replaced by one neural-network forward pass plus a tiny QP, without relying on imperfect expert demonstrations.

What carries the argument

The central object is the neural co-state regulator (NCR): a feedforward co-state network (CoNN) that maps a state $z_k$ to a predicted co-state trajectory $\hat{\Lambda}_k$ of length $n$, trained with the loss $L_{\text{stage}} + L_{\text{terminal}} + \beta \|\hat{\Lambda}_k\|_{1,1}$, where $\beta$ is a scalar hyperparameter. At deployment, the first predicted co-state $\hat{\lambda}_k$ feeds the quadratic program of Eq. (17), which enforces the input constraint $u_k \in \mathcal{U}$. During training the unconstrained PMP relation $u_k = -\tfrac{1}{2}R^{-1}g(z_k)^\top \hat{\lambda}_k$ supplies the controls used to evaluate the stage cost, so the network learns the dynamics-informed co-state without ever solving the two-point boundary value problem explicitly. The same QP step converts the predicted co-state into a feasible constrained control at runtime.

What would settle it

Run the NCR on a control-affine system whose optimal input saturates for a nontrivial time interval; compute the true constrained-optimal co-state from a numerical two-point boundary value problem with active constraints, and compare it with the NCR's prediction during the saturated arc. If they diverge on that arc, or if the closed-loop cost is worse than nonlinear MPC's, the paper's key premise fails.

Watch

Extended reading notes

Core claim

The paper's central claim is that the optimal co-state trajectory for a control-affine quadratic-regulation problem can be learned directly from the system model and cost, without ground-truth co-states or expert solutions. A co-state neural network (CoNN) is trained with a PMP-informed loss that combines a stage cost, a terminal cost, and a regularizer on the co-state trajectory; controls during training come from the unconstrained PMP stationarity condition $u = -\tfrac{1}{2}R^{-1}g(z)^\top \lambda$. At deployment, the first predicted co-state vector is passed to a quadratic program that enforces the input constraint and returns the control input. The authors show on the unicycle model that the resulting feedback law generalizes to unseen initial conditions and nonzero references and, in the tested cases, gives comparable or better convergence error and smoother inputs than nonlinear MPC, with roughly two orders of magnitude less computation per step.

Load-bearing premise

The method rests on the premise that the optimal co-state trajectory is the same whether or not the input limits are active, so a network trained on unconstrained controls can be paired with a constrained quadratic program; if saturating constraints change the co-state, the learned trajectory may not be the constrained-optimal one.

Editorial extensions

If this is right

  • Online computation per step becomes a fixed-size network forward pass plus a small quadratic program, independent of the prediction horizon.
  • Prediction horizons can be lengthened to improve closed-loop performance without the exponential growth in solve time seen with nonlinear MPC.
  • Training no longer requires optimal trajectories from expert solvers, so the method can be applied where a good TPBVP or MPC solution is expensive or unavailable.
  • The trained regulator can be deployed on unseen initial conditions and nonzero setpoint tracking without retraining, at least for the unicycle example.
  • In cases where the initial state is far from the reference, the NCR reports smoother control inputs and lower cumulative absolute state error than MPC.

Reading between the lines

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

  • An immediate test of the method's reach is to apply it to a system with long saturation arcs or state constraints; if active constraints alter the co-state, the unconstrained-trained NCR may need an online correction step.
  • The minimum-norm co-state regularization selects one of possibly many co-state solutions, so the weight $\beta$ may act as a design parameter rather than a neutral regularizer.
  • The separation of a learned co-state from a convex projection step could extend beyond box input constraints to any constraint set that keeps the quadratic program tractable.
  • Because the network is trained from model knowledge alone, the NCR could serve as a warm start for online MPC, making the solver faster while retaining its guarantees.
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 neural co-state regulator (NCR): a feedforward network that maps a state to a predicted co-state trajectory, trained in an unsupervised manner with a loss composed of a stage cost, a terminal cost, and an L1 regularizer on the co-state. During deployment, the first predicted co-state is used to solve a small QP that imposes input constraints. The method is tested on a unicycle model and compared with nonlinear MPC, with claims of comparable or better convergence, smoother inputs, faster computation, and generalization to unseen initial states and nonzero references.

Significance. If the central claim were valid, the NCR would be a practically attractive real-time constrained optimal feedback policy that avoids online nonlinear programming. The paper has a clearly stated problem, a simple architecture, and a concrete experimental comparison, and the reported two-orders-of-magnitude speedup is potentially valuable. However, the theoretical foundation of the method is not established: the training loss does not enforce the PMP necessary conditions, the constraint-independence assumption for co-states is generally false, and the experiments do not compare against a true constrained optimal solution. No code, data, or machine-checked derivations are provided, which limits reproducibility. The paper's own conclusion admits that the regularized co-state loss requires better design, underscoring that the core mechanism is currently ad hoc.

major comments (4)
  1. [§IV.B.2, Eqs. (16)-(17)] The load-bearing assumption that 'the optimality of the co-state is independent of the input constraints' is not a consequence of PMP. In the co-state equation (11), the term involving ∇_z^T(g(z)u)λ depends on the actual control input u(t). When an input constraint becomes active, the optimal u(t), the state trajectory, and hence the adjoint trajectory all change. Training with the unconstrained control from Eq. (13) and then solving the QP in Eq. (17) with the resulting λ minimizes the Hamiltonian with the wrong adjoint variable for the constrained OCP (8). This invalidates the abstract claim that the NCR produces constrained-optimal PMP solutions.
  2. [§IV.B.2 and §VI] The loss function does not enforce the co-state dynamics (11) or the transversality condition λ(t_f) = ∇_z φ(z(t_f)). Moreover, the regularizer β||Λ||_{1,1} actively drives the terminal co-state toward zero. For the experiments, φ(z) = z^T S z with S = 50Q = diag(500,500,500), so transversality requires λ(t_f) = 2S z(t_f), which is nonzero whenever the final state error is nonzero. The training objective is therefore in tension with PMP optimality, and the paper's Section VI itself concedes that the regularized co-state loss needs better design. The network is not shown to produce PMP co-states.
  3. [§IV.B.2 and Algorithm 1] Because the control input used in the loss is computed from the co-state prediction through Eq. (13), the stage loss in Eq. (16) is simply the original quadratic control cost written in terms of the network output. The NN therefore directly minimizes the control cost plus an ad hoc L1 penalty; this is a reparameterized policy optimization rather than an indirect PMP method. The claim that the NCR 'learns optimal co-state trajectories' is a restatement of the optimization objective, not a verification of PMP optimality.
  4. [§V.A-C and Table I] The experimental comparison does not establish optimality or superiority over MPC. The MPC baseline is itself an approximate finite-horizon NLP solution, the reported final states are not converged to the origin (e.g., Case A gives y_f,NCR = 0.06 and y_f,MPC = 0.14), and no comparison is made against a known constrained TPBVP solution or a tight-tolerance direct solver. The generalization claim rests on three hand-picked initial conditions and one nonzero reference, with no error bars or statistical evaluation. These experiments support only a claim of reasonable feedback performance, not the headline claim of outperforming an optimal controller.
minor comments (6)
  1. [§IV.B.1] The training data are described as '1000 evenly sampled states' and later as a uniform 10×10×10 grid; the earlier phrase is ambiguous and should be aligned with the actual sampling procedure.
  2. [§II.B] The co-state terminal condition λ(t_f) = ∇_z φ(z(t_f)) is stated without an equation number; adding one would help cross-referencing.
  3. [Algorithm 1, line 6] The step 'Obtain [u_k,...,u_{k+n-1}] based on PMP' is ambiguous during training because input constraints are not enforced; the text should specify that Eq. (13) is used.
  4. [Fig. 6] The right panel axis is labeled 'natural log of time' but the units and the logarithm base are not specified, and the left panel does not indicate which curve corresponds to which case.
  5. [Table I] The use of italics for 'better performance' is inconsistent with the text's discussion of Cases B and C state MSD, where MPC has a lower value but is judged less favorable due to abrupt θ changes; the criterion should be clarified.
  6. [Throughout] There are minor typographical issues, including 'pontryagin's' in Section II and 'close-loop' in Section V.A.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: the NCR is direct cost minimization benchmarked against an external MPC solver; the constraint-independence premise is a correctness risk, not a circular step.

full rationale

The claimed derivation chain is not circular. The CoNN is trained by minimizing Lstage + Lterminal with controls generated from the predicted co-state via Eq. (13), which is direct policy/cost optimization rather than supervised fitting to target co-states or to MPC outputs. The added L1 regularizer even prevents the trained network from being an exact minimizer of the stated OCP, so the reported performance is not forced by construction. The constrained deployment QP (17) is a post-processing step whose input is a learned latent co-state, and the paper's optimality claim for constrained cases rests on the stated but unproved premise that 'the optimality of the co-state is independent of the input constraints.' That premise is a genuine correctness risk—constrained PMP co-states generally depend on active constraints—and the training loss never enforces the co-state dynamics (11) or the terminal transversality condition. However, an unsupported or false premise is not circularity unless the conclusion is already assumed in the premise. The comparison to an external, suboptimal MPC solver is empirically falsifiable, and no equation in the paper makes the MPC-outperforming trajectories equal by construction to the training data or to a fitted output. The only self-citations ([25] in Section I.A and the conclusion) introduce the prior supervised CoNN concept and are not load-bearing for the new unsupervised loss or the QP deployment. Therefore the paper has no significant circularity; its weaknesses are matters of correctness and validation, not of circular derivation.

Assumptions & free parameters 1 free parameters · 4 assumptions · 0 invented entities

The central claim rests on PMP (standard), but also on two ad hoc additions: the constraint-independence assumption and the L1 co-state penalty. Both are stated without proof and can change the solution. The only hand-chosen numeric parameter is the regularization weight beta=0.1.

free parameters (1)
  • co-state regularization weight beta = 0.1
    Hand-chosen in Section V; scales the L1 penalty on co-states, modifying the OCP objective.
assumptions (4)
  • standard math Pontryagin's Minimum Principle provides necessary optimality conditions for the OCP.
    Used to derive the Hamiltonian, co-state dynamics, and the unconstrained control law (Section II.B, Section III).
  • ad hoc to paper The optimal co-state trajectory is independent of control input constraints.
    Stated in Section IV.B.2 without proof; generally false for constrained OCPs and load-bearing for the training scheme.
  • ad hoc to paper The L1 co-state regularization drives the network toward a 'norm-optimal' co-state without harming optimality.
    Introduced in Section IV.B.2 as a heuristic (beta * ||Lambda||_{1,1}); no justification or ablation.
  • domain assumption The system dynamics and cost matrices are known exactly and the training set is representative.
    The training assumes access to the unicycle model and the chosen Q, R, S matrices (Section V).

how reviews work

0 comments
Cite this review

Pith. "Pith review of Neural Co-state Regulator: A Data-Driven Paradigm for Real-time Optimal Control with Input Constraints." pith.science (2026). https://pith.science/paper/M6HATDDT

@misc{pith2026250712259,
  author       = {Pith},
  title        = {Pith review of: Neural Co-state Regulator: A Data-Driven Paradigm for Real-time Optimal Control with Input Constraints},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/M6HATDDT}},
  note         = {Machine review of arXiv:2507.12259}
}
read the original abstract

We propose a novel unsupervised learning framework for solving nonlinear optimal control problems (OCPs) with input constraints in real-time. In this framework, a neural network (NN) learns to predict the optimal co-state trajectory that minimizes the control Hamiltonian for a given system, at any system's state, based on the Pontryagin's Minimum Principle (PMP). Specifically, the NN is trained to find the norm-optimal co-state solution that simultaneously satisfies the nonlinear system dynamics and minimizes a quadratic regulation cost. The control input is then extracted from the predicted optimal co-state trajectory by solving a quadratic program (QP) to satisfy input constraints and optimality conditions. We coin the term neural co-state regulator (NCR) to describe the combination of the co-state NN and control input QP solver. To demonstrate the effectiveness of the NCR, we compare its feedback control performance with that of an expert nonlinear model predictive control (MPC) solver on a unicycle model. Because the NCR's training does not rely on expert nonlinear control solvers which are often suboptimal, the NCR is able to produce solutions that outperform the nonlinear MPC solver in terms of convergence error and input trajectory smoothness even for system conditions that are outside its original training domain. At the same time, the NCR offers two orders of magnitude less computational time than the nonlinear MPC.

Figures

Figures reproduced from arXiv: 2507.12259 by the authors.

Figure 1
Figure 1. NCR training procedures and CoNN architecture. Given the state [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Neural co-state regulator validation block diagram. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Comparison of MPC and NCR solutions for initial [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Comparison of MPC and NCR solutions for unseen [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Comparison of MPC and NCR solutions for unseen [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

32 extracted references · 26 canonical work pages

  1. [1]

    Applied and computational optimal control,

    K. L. Teo, B. Li, C. Yu, V . Rehbock et al., “Applied and computational optimal control,” Optimization and Its Applications , 2021

  2. [2]

    A survey of computational complexity results in systems and control,

    D. Peaucelle and D. Henrion, “A survey of computational complexity results in systems and control,” Automatica, vol. 46, no. 7, pp. 1067– 1084, 2010

  3. [3]

    J. T. Betts, Practical Methods for Optimal Control and Estimation Using Nonlinear Programming . Philadelphia, PA: SIAM, 2010

  4. [4]

    Gr ¨une and J

    L. Gr ¨une and J. Pannek, Nonlinear Model Predictive Control: Theory and Algorithms, 1st ed., ser. Communications and Control Engineer- ing. New York: Springer, 2011

  5. [5]

    Review on model predictive control: An engineering perspective,

    M. Schwenzer, M. Ay, T. Bergs, and D. Abel, “Review on model predictive control: An engineering perspective,” Journal of Control, Automation and Electrical Systems , vol. 32, no. 5, pp. 1214–1232, 2021

  6. [6]

    Optimal power-split of hybrid energy storage system using pontryagin’s minimum principle and deep reinforcement learning approach for electric vehicle application,

    P. Nambisan and M. Khanra, “Optimal power-split of hybrid energy storage system using pontryagin’s minimum principle and deep reinforcement learning approach for electric vehicle application,” Engineering Applications of Artificial Intelligence , vol. 135, p. 108769, September 2024

  7. [7]

    D. E. Kirk, Optimal Control Theory: An Introduction , reprint edi- tion ed. Mineola, New York: Dover Publications, 2004, ch. 5.3, pp. 227–239, chapter 5.3: Pontryagin’s Minimum Principle and State Inequality Constraints

  8. [8]

    A survey of numerical methods for optimal control,

    A. V . Rao, “A survey of numerical methods for optimal control,” Advances in the astronautical Sciences , vol. 135, no. 1, pp. 497–528, 2009

Show all 32 references
  1. [9]

    A penalty function approach to constrained pontryagin- based nonlinear model predictive control,

    M. Pagone, M. Boggio, C. Novara, A. Proskurnikov, and G. C. Calafiore, “A penalty function approach to constrained pontryagin- based nonlinear model predictive control,” in Proceedings of the 61st IEEE Conference on Decision and Control (CDC) . IEEE, 2022, pp. 3705–3710

  2. [10]

    Sequential convex programming for non-linear stochastic optimal control,

    R. Bonalli, T. Lew, and M. Pavone, “Sequential convex programming for non-linear stochastic optimal control,” ESAIM: Control, Optimisation and Calculus of Variations , vol. 28, p. 64, 2022

  3. [11]

    Nonlinear model predictive control with aggregated constraints,

    M. de Freitas Virgilio Pereira, I. V . Kolmanovsky, and C. E. S. Cesnik, “Nonlinear model predictive control with aggregated constraints,” Automatica, vol. 132, p. 109746, 2021, brief paper

  4. [12]

    Optimal control operator perspective and a neural adaptive spectral method,

    M. Feng, Z. Chen, Y . Huang, Y . Liu, and J. Yan, “Optimal control operator perspective and a neural adaptive spectral method,” arXiv preprint arXiv:2412.12469, 2024, license: CC BY-NC-ND 4.0

  5. [13]

    R. S. Sutton, A. G. Barto et al., Reinforcement learning: An introduc- tion. MIT press Cambridge, 1998, vol. 1, no. 1

  6. [14]

    Continuous control with deep reinforce- ment learning,

    T. P. Lillicrap, J. J. Hunt, A. Pritzel, N. Heess, T. Erez, Y . Tassa, D. Silver, and D. Wierstra, “Continuous control with deep reinforce- ment learning,” arXiv preprint arXiv:1509.02971 , 2015

  7. [15]

    Proximal policy optimization algorithms,

    J. Schulman, F. Wolski, P. Dhariwal, A. Radford, and O. Klimov, “Proximal policy optimization algorithms,” arXiv preprint arXiv:1707.06347, 2017

  8. [16]

    Quan- tifying generalization in reinforcement learning,

    K. Cobbe, O. Klimov, C. Hesse, T. Kim, and J. Schulman, “Quan- tifying generalization in reinforcement learning,” in International conference on machine learning . PMLR, 2019, pp. 1282–1289

  9. [17]

    Tedrake, Underactuated Robotics, 2023

    R. Tedrake, Underactuated Robotics, 2023

  10. [18]

    Imitation learning with neural network-based model predictive control,

    Y . Chen, M. Chen, and M. Tomizuka, “Imitation learning with neural network-based model predictive control,” arXiv preprint arXiv:2001.02533, 2020

  11. [19]

    End to end learning for self-driving cars,

    M. Bojarski, D. D. Testa, D. Dworakowski, B. Firner, B. Flepp, P. Goyal, L. D. Jackel, M. Monfort, U. Muller, J. Zhang, X. Zhang, J. Zhao, and K. Zieba, “End to end learning for self-driving cars,” https://arxiv.org/abs/1604.07316, 2016, nVIDIA Technical Report

  12. [20]

    A reduction of imitation learning and structured prediction to no-regret online learning,

    S. Ross, G. Gordon, and D. Bagnell, “A reduction of imitation learning and structured prediction to no-regret online learning,” in Proceedings of the fourteenth international conference on artificial intelligence and statistics. JMLR Workshop and Conference Proceedings, 2011, ...

  13. [21]

    Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware,

    T. Z. Zhao, V . Kumar, S. Levine, and C. Finn, “Learning Fine-Grained Bimanual Manipulation with Low-Cost Hardware,” in Proceedings of Robotics: Science and Systems , Daegu, Republic of Korea, July 2023

  14. [22]

    Optimal control problem via neural networks,

    S. Effati and M. Pakdaman, “Optimal control problem via neural networks,” Neural Computing and Applications , vol. 23, no. 7-8, pp. 2093–2100, 2013

  15. [23]

    Pontryagin neural networks with functional interpolation for optimal intercept problems,

    A. D’Ambrosio, E. Schiassi, F. Curti, and R. Furfaro, “Pontryagin neural networks with functional interpolation for optimal intercept problems,” Mathematics, vol. 9, no. 9, 2021

  16. [24]

    A machine learning enhanced algorithm for the optimal landing problem,

    Y . Zang, J. Long, X. Zhang, W. Hu, J. Han et al., “A machine learning enhanced algorithm for the optimal landing problem,” in Mathematical and Scientific Machine Learning . PMLR, 2022, pp. 319–334

  17. [25]

    Co-state neural network for real- time nonlinear optimal control with input constraints,

    L. Lian and U. Inyang-Udoh, “Co-state neural network for real- time nonlinear optimal control with input constraints,” arXiv preprint arXiv:2503.00529, 2025

  18. [26]

    L. T. Biegler, Nonlinear programming: concepts, algorithms, and applications to chemical processes . SIAM, 2010

  19. [27]

    H. B. Keller, Numerical solution of two point boundary value prob- lems. SIAM, 1976

  20. [28]

    Use of orthogonal collocation method in optimal control problems,

    S. Oh and R. Luus, “Use of orthogonal collocation method in optimal control problems,” International Journal of Control, vol. 26, no. 5, pp. 657–673, 1977

  21. [29]

    Q-learning and pontryagin’s minimum prin- ciple,

    P. Mehta and S. Meyn, “Q-learning and pontryagin’s minimum prin- ciple,” in Proceedings of the 48th IEEE Conference on Decision and Control (CDC). IEEE, 2009, pp. 3598–3605

  22. [30]

    Human-level control through deep reinforcement learning,

    V . Mnih, K. Kavukcuoglu, D. Silver, A. A. Rusu, J. Veness, M. G. Bellemare, A. Graves, M. Riedmiller, A. K. Fidjeland, G. Ostrovski et al. , “Human-level control through deep reinforcement learning,” nature, vol. 518, no. 7540, pp. 529–533, 2015

  23. [31]

    Reinforcement learning-based model predictive control for discrete-time systems,

    M. Lin, Z. Sun, Y . Xia, and J. Zhang, “Reinforcement learning-based model predictive control for discrete-time systems,”IEEE Transactions on Neural Networks and Learning Systems , vol. 35, no. 3, pp. 3312– 3324, 2024

  24. [32]

    Flatness-based model predictive con- trol for quadrotor trajectory tracking,

    M. Greeff and A. P. Schoellig, “Flatness-based model predictive con- trol for quadrotor trajectory tracking,” in 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), 2018, pp. 6740– 6745

Pith tools

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