{"id":"04f6d84c-cc0c-46e9-a1b7-a04916a931a9","arxiv_id":"2507.12259","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A model-based neural network trained with a cost-based loss predicts co-states and, via a QP, produces constrained control that matches or beats nonlinear MPC on a unicycle at far lower compute.","lead":"A neural network is trained without expert examples to predict the co-state variables of an optimal control problem, and its outputs are converted into control commands by solving a small quadratic program. The authors show this neural co-state regulator runs about 100 times faster than nonlinear MPC on a simulated unicycle and matches or improves convergence in three test cases.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The training co-states are not the constrained-optimal co-states: Section IV.B.2 assumes constraint-independence, which PMP does not provide, and the loss never enforces co-state dynamics or transversality.","rationale":"The reader's weakest-assumption analysis identifies exactly the load-bearing flaw: the network is trained on unconstrained co-states and then used in a constrained QP. This is not a stylistic or consensus disagreement; it is a mismatch between the mathematical object being learned and the object required by PMP when constraints are active. The paper's own abstract claims constrained, real-time optimal control and out-of-training-domain superiority, so the unsupported constraint-independence claim is central. The additional observation that the training loss omits co-state dynamics and the PMP terminal condition reinforces the concern but is not needed to reject the paper as written. The experiments are single-run simulations with no error bars and no comparison of closed-loop cost against a true constrained-optimal solution, so they cannot rescue the central claim. I therefore keep the reader's REJECT verdict unchanged.","tokens_in":10844,"tokens_out":4734,"duration_ms":62416,"concrete_test":"For Case B with z(0) = [-5.24, 4.11, 2.72]^T, solve the constrained OCP (18) by an indirect shooting method that enforces the state/co-state equations (10)-(11), the active-set Hamiltonian minimization (14), and the transversality condition lambda(tf) = 2 S z(tf); record the resulting lambda_c(0). Compare this with the unconstrained co-state lambda_u(0) obtained from Eq. (13) with the same initial state. If |v| or |omega| reaches its bound during the horizon and lambda_c(0) differs from lambda_u(0), the constraint-independence assumption in Section IV.B.2 is invalid for this experiment.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing premise is stated in Section IV.B.2: 'the optimality of the co-state is independent of the input constraints,' so the network is trained with unconstrained controls from Eq. (13) and deployed with the constrained QP in Eq. (17). This is false. In PMP, the co-state dynamics (11) contain the actual optimal control u(t), and when an input bound becomes active, both u(t) and the resulting state trajectory change. Hence the constrained-optimal co-state trajectory generally differs from the unconstrained one. Solving Eq. (17) with the unconstrained co-state therefore minimizes the Hamiltonian with the wrong adjoint variable and does not solve the constrained OCP (8). Independent of the constraint issue, the loss Lstage + Lterminal + beta||Lambda|| does not enforce the co-state dynamics (11) or the transversality condition lambda(tf) = grad_z phi(z(tf)); the L1 regularizer actually drives the terminal co-state toward zero, which is inconsistent with the nonzero quadratic terminal cost used in the experiments. The central claim that the NCR learns PMP-optimal constrained co-states and outperforms MPC therefore rests on an invalid premise.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":11101,"tokens_out":4378,"duration_ms":50262,"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":[{"comment":"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.","section":"§IV.B.2, Eqs. (16)-(17)"},{"comment":"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.","section":"§IV.B.2 and §VI"},{"comment":"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.","section":"§IV.B.2 and Algorithm 1"},{"comment":"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.","section":"§V.A-C and Table I"}],"minor_comments":[{"comment":"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.","section":"§IV.B.1"},{"comment":"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.","section":"§II.B"},{"comment":"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.","section":"Algorithm 1, line 6"},{"comment":"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.","section":"Fig. 6"},{"comment":"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.","section":"Table I"},{"comment":"There are minor typographical issues, including 'pontryagin's' in Section II and 'close-loop' in Section V.A.","section":"Throughout"}],"recommendation":"reject","confidential_remarks":"The manuscript's novelty relative to the authors' prior CoNN work [25] consists mainly of replacing supervised TPBVP labels with an unsupervised loss, but the loss does not actually enforce PMP conditions and the constraint-independence premise is false. The empirical comparison uses a finite-horizon MPC as the 'expert' baseline, which is a weak proxy for optimality; a direct comparison with indirect shooting or a tight-tolerance direct collocation solution would be needed. I would encourage the authors to reframe the method as a direct policy-learning heuristic, remove the PMP-optimality claims, or add the missing co-state dynamics and transversality constraints to the loss."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper has a load-bearing flaw: its training phase ignores input constraints, and the loss never enforces the co-state dynamics or transversality. That means the network is not learning the constrained-optimal co-state trajectory, and the QP at deployment is minimizing a Hamiltonian with the wrong adjoint variable. The central claim of outperforming MPC rests on that invalid premise.\n\nThat said, the paper does something genuinely new: it removes the need for expert co-state trajectories in co-state neural networks by using a rollout cost as the loss. The idea of learning a policy through the lens of PMP, with the co-state as an intermediate variable, is worth exploring. The QP extraction layer is clean, and the unicycle example is appropriately standard.\n\nThe soft spots are serious. Section IV.B.2 states 'the optimality of the co-state is independent of the input constraints,' which is false for any problem where constraints become active. In PMP, the co-state dynamics contain the actual optimal control, so the constrained-optimal co-state differs from the unconstrained one. The loss itself is just the original cost plus an L1 penalty on the co-state; it doesn't include the co-state differential equations or the terminal condition. In fact, the L1 term pushes the terminal co-state toward zero, which contradicts the quadratic terminal cost used in the experiments. So the network is really just a policy network whose hidden output is named 'co-state.'\n\nThe experiments are too thin to support the claims: three single-run simulations, no error bars or repeated trials, and no direct cost comparison. The speed advantage is expected from NN inference versus solving an NLP, and the MPC baseline may not have been tuned to the same stage cost. The claim of 'outperforming MPC' is not convincing. The citation pattern is fine, and the authors do engage with prior learning-based MPC and PMP approximation work.\n\nWho is this paper for? Researchers working on learning-based optimal control and PMP approximations. The idea is a legitimate research direction, but this version is not ready for serious peer review. It needs either a correct treatment of constraints (e.g., train with the constrained QP or include the constraints in the loss) and enforcement of the co-state dynamics and transversality, or a clearly reframed statement that the network is a policy approximator with a regularized cost. If that were done, the paper might be worth revisiting. As is, I would desk reject.\n\nI won't cite it in the near term, but I might bring it to a reading group as a cautionary example of how a plausible-sounding PMP-based learning scheme can go wrong.","headline":"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.","tokens_in":11579,"tokens_out":4814,"would_cite":false,"duration_ms":52343,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"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…","keywords":["neural co-state regulator","Pontryagin's minimum principle","input constraints","quadratic program","model predictive control","unsupervised learning","real-time optimal control","unicycle model"],"falsifier":"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.","tokens_in":1584,"feed_emoji":"🎛️","tokens_out":2452,"duration_ms":101751,"temperature":0.7,"pith_summary":"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.","feed_headline":"Neural co-state network beats nonlinear MPC at 100x speed","feed_subtitle":"PMP-trained network plus a small QP gives fast, smooth constrained control even outside its training domain.","key_machinery":"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.","core_discovery":"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.","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies Pontryagin's Minimum Principle, the Hamiltonian, and the co-state terminal conditions that the loss and the constrained QP are built on.","marker":"[7]"},{"why":"Introduces the co-state neural network this work extends; NCR removes its reliance on expert TPBVP supervision.","marker":"[25]"},{"why":"The temporal-difference learning idea the authors cite as motivation for their unsupervised, PMP-informed loss.","marker":"[30]"},{"why":"Provides the unicycle dynamics and the input bounds ($-1 \\le v \\le 1$, $-4 \\le \\omega \\le 4$) used in the experiments.","marker":"[31]"},{"why":"Defines the nonlinear MPC formulation used as the expert baseline in all comparisons.","marker":"[4]"}],"fun_headline_variants":["Neural co-state network beats MPC at 100x speed","Co-state NN learns constrained control, outperforms MPC","PMP-trained neural net delivers real-time constrained control","Data-driven co-state regulator outruns MPC 100x"],"cache_read_input_tokens":13824,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["Neural co-state network beats MPC at 100x speed","Co-state NN learns constrained control, outperforms MPC","PMP-trained neural net delivers real-time constrained control","Data-driven co-state regulator outruns MPC 100x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000195,"raw_usage":{"total_tokens":1374,"prompt_tokens":980,"completion_tokens":394,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":596,"completion_tokens_details":{"reasoning_tokens":327}},"tokens_in":596,"tokens_out":394,"duration_ms":4788,"temperature":1.0,"reasoning_tokens":327,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:50:10.623237+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies Pontryagin's Minimum Principle, the Hamiltonian, and the co-state terminal conditions that the loss and the constrained QP are built on."},{"cited_title":"Co-state Neural Network for Real-time Nonlinear Optimal Control with Input Constraints","cited_arxiv_id":"2503.00529","evidence_quote":"Introduces the co-state neural network this work extends; NCR removes its reliance on expert TPBVP supervision."},{"cited_title":"Reinforcement learning-based model predictive control for discrete-time systems,","cited_arxiv_id":null,"evidence_quote":"Provides the unicycle dynamics and the input bounds ($-1 \\le v \\le 1$, $-4 \\le \\omega \\le 4$) used in the experiments."},{"cited_title":"Gr ¨une and J","cited_arxiv_id":null,"evidence_quote":"Defines the nonlinear MPC formulation used as the expert baseline in all comparisons."}],"review_version":1}