{"id":"3d6922d3-c0a6-4493-90e0-6794fe4600f3","arxiv_id":"2504.17139","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Opt-ODENet learns safe, stable state-feedback controllers by backpropagating a CLF-based loss through a neural ODE and a differentiable CBF-QP layer, and it learns the CBF class-K parameter during training.","lead":"This paper trains a controller by blending two ideas: a neural ODE models the motion, and a quadratic-programming layer turns safety constraints into hard rules the controller must obey at every step. The method also learns the safety margin automatically, and tests show a simulated unicycle reaching its target without colliding with an obstacle.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (13) omits the state-dependence of the QP layer: ∂τ/∂x ignores ∂τ/∂G·∂G/∂x and ∂τ/∂h·∂h/∂x, so Proposition 4's adjoint gradients are not the true loss gradients.","rationale":"The reader's weakest assumption concerns QP feasibility and KKT invertibility; those are real but partly acknowledged in Appendix 7.2.2 and are configuration-dependent. The chain-rule omission is more load-bearing because it is an actual error in the derivation of the central gradient claim. If Eq. (13) is wrong, Proposition 4 cannot be used to compute the loss gradient, and the paper's stated justification for training via the adjoint method with QP sensitivity fails. The fix is straightforward—add ∂τ/∂G·∂G/∂x + ∂τ/∂h·∂h/∂x, or state that these paths are detached and discuss the resulting bias—so the paper is conditionally acceptable rather than rejectable. The experimental evidence (no collisions with learned κ, faster convergence than fixed baselines) is independent support and could survive the correction. I therefore keep the conditional verdict, with the condition sharpened to include correcting Eq. (13).","tokens_in":16899,"tokens_out":11843,"duration_ms":114446,"concrete_test":"Take a unicycle state x at which the CBF-QP constraint is active (e.g., the QP projection of u_nn lies on the boundary G u = h). Compute ∂τ/∂x by finite differences of the QP solution (9) with respect to x, and compare it with the value given by Eq. (13), ∂τ/∂π · ∂π/∂x, using AD for ∂π/∂x and the KKT sensitivity for ∂τ/∂π. If the two differ beyond numerical tolerance (say 1%), Eq. (13) is incomplete. A stronger end-to-end test is to implement the adjoint update (11) with the current (13) and compare the resulting ∇θ1ℓ and ∇θ2ℓ against gradients from full backpropagation through the Euler-unrolled computation graph on one training batch; any mismatch confirms the missing G(x), h(x,θ2) terms.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Proposition 4 and the adjoint system (11) require ∂F/∂x, expanded in (12) as f_x + g_x τ + g ∂τ/∂x. Equation (13) then states ∂τ/∂x = (∂τ/∂π)(∂π/∂x), i.e. the QP solution τ is differentiated only through the controller network output π(x,θ1). But τ is the minimizer of the CBF-QP layer (9), whose data depend on the state: G = -∂B/∂x g(x) and h = ∂B/∂x f(x) + α(B(x),θ2). The full chain rule for the QP solution therefore includes (∂τ/∂G)(∂G/∂x) + (∂τ/∂h)(∂h/∂x), where ∂τ/∂G and ∂τ/∂h are precisely the KKT sensitivities (20)/(51). These terms are generically nonzero whenever the CBF constraint is active, which is the regime the safety layer is designed for. Omitting them makes (13) incorrect, so the adjoint state p and the gradients μ1(t0), μ2(t0) in (11) do not equal ∇θ1ℓ and ∇θ2ℓ of the loss (8)/(25). The central theoretical claim—that backpropagation through the Neural ODE and QP layer via the adjoint equations computes the training gradient—is therefore unsupported as stated. This is an internal inconsistency in the proof, not a disagreement with prior work; if the implementation uses unrolled automatic differentiation instead, the experiments may still be valid, but then Proposition 4 is not the method being evaluated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Opt-ODENet, a Neural ODE-based control framework in which a Control Barrier Function (CBF) constraint is enforced as a hard requirement through a differentiable Quadratic Programming (QP) layer, while a Control Lyapunov Function (CLF) is used as a training loss to promote convergence to a target. The main theoretical claim is Proposition 4: the loss gradients with respect to the controller network parameters θ1 and the CBF class-K parameters θ2 can be computed by solving an adjoint system whose right-hand side involves, among other terms, the derivative of the QP layer solution τ with respect to its inputs, obtained via KKT sensitivity analysis. The paper presents a unicycle obstacle-avoidance experiment and a multi-car HOCBF experiment, reporting that learning the class-K coefficient κ (13.38 for the unicycle; {18.6, 96.04} for the cars) yields the best safety-performance trade-off among the tested configurations. The long version includes an appendix with additional derivations and experimental details.","tokens_in":17291,"tokens_out":5275,"duration_ms":50395,"significance":"If the central gradient derivation were correct, the paper would contribute a useful design pattern: combining Neural ODEs with differentiable optimization layers to train state-feedback controllers that are safe by construction and whose stability is encouraged by a CLF term. The idea of learning the CBF class-K parameter through gradient-based training rather than fixing it a priori is interesting and is supported by the empirical comparisons in Tables 1. The paper also provides a public code link, which is commendable for reproducibility. However, the load-bearing theoretical claim (Proposition 4) contains a chain-rule gap that undermines the stated adjoint computation, and the hard-safety guarantee is qualified by the documented infeasibility of the QP layer during training. These issues prevent the paper from currently establishing its central claims, although they appear to be locally fixable.","major_comments":[{"comment":"The chain rule for ∂τ/∂x in Eq. (13) is incomplete. The CBF-QP layer (9) has G = -∂B/∂x g(x) and h = ∂B/∂x f(x) + α(B(x), θ2), both of which depend on the state x. The full derivative of the QP solution with respect to x is ∂τ/∂x = (∂τ/∂π)(∂π/∂x) + (∂τ/∂G)(∂G/∂x) + (∂τ/∂h)(∂h/∂x), where ∂τ/∂G and ∂τ/∂h are precisely the KKT sensitivities (20). Eq. (13) omits the last two terms, so the adjoint state p in (11) does not satisfy the correct adjoint equation and μ1(t0), μ2(t0) are not the gradients of the loss ℓ. This is an internal inconsistency in the proof of Proposition 4, not merely a stylistic issue. If the actual implementation uses unrolled automatic differentiation through the ODE solver and the QP layer, then the paper should state that Proposition 4 does not describe the method being evaluated, or should supply the corrected adjoint equations.","section":"Section 4, Eq. (13)"},{"comment":"The paper assumes the CBF-QP layer is feasible and differentiable at every state visited during training, but this is not guaranteed. The appendix (Section 7.2.2) documents that in the car-following HOCBF case with κ={2,1}, 'certain infeasibility happens, which makes the whole training problematic.' Since the forward pass of the QP layer fails when no feasible control exists, the abstract's statement that CBFs 'enforce real-time safety' as hard requirements is not supported by the presented method. The paper should either provide a sufficient condition for feasibility, add a relaxation (such as slack variables) with an explicit safety guarantee, or clearly bound the set of states for which the QP is feasible.","section":"Section 4, Eq. (9); Appendix 7.2.2"},{"comment":"The abstract claims that 'Stability and convergence are ensured through Control Lyapunov Functions (CLFs) in the loss function'. However, Remark 5 states that the convergence expression (26) holds only if there exists a parameter θ* satisfying ℓ(θ*)=0. Gradient-based training provides no guarantee that such a parameter is reached, and the experiments only show empirical convergence losses. The claims in the abstract and Section 6 should be rephrased to reflect the conditional nature of the stability property, e.g., 'encouraged' or 'promoted' rather than 'ensured'.","section":"Abstract; Remark 5"}],"minor_comments":[{"comment":"The caption says 'Distance over time with varying ℓ', but the text refers to varying γ (the CLF decay rate); the caption should be corrected to avoid ambiguity.","section":"Figure 2 caption"},{"comment":"The expression for ∂ℓ/∂G appears to have a typographical inconsistency: the term 'λ∗d⊤u' is ambiguous (the standard OptNet result is λ* du^T plus a term involving D(λ*)dλ z*^T); please clarify the notation.","section":"Appendix 7.1.2, Eq. (51)"},{"comment":"There are many rendering artifacts in the equations (e.g., '⁄𝑥' instead of \\dot{x}, '𝜃' for θ, and fragmented over-dots), which make the paper difficult to read. The authors should ensure the final version is typeset cleanly.","section":"Throughout the manuscript"},{"comment":"The adjoint equation for p in the appendix writes ∂L/∂z in the p dynamics, but the running cost L depends on x, t, and θ; this should be ∂L/∂x for consistency with the main text.","section":"Section 7.1.1, Eq. (32)"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a useful paper with a real gap in the central derivation. The combination—training a Neural ODE controller through a differentiable CBF-QP layer and learning the class-K coefficient end-to-end—is new in the cited literature. The unicycle experiment is credible: learned kappa = 13.38 beats the two fixed values on mean error and avoids collisions, and the authors are appropriately candid in the appendix that infeasibility can occur and make training problematic.\n\nThe soft spot is not experimental; it's Proposition 4. The adjoint equation for p requires ∂F/∂x, which includes g ∂τ/∂x. Equation (13) computes ∂τ/∂x as (∂τ/∂π)(∂π/∂x), but τ is the minimizer of the CBF-QP, whose data G and h depend on x. Unless you include (∂τ/∂G)(∂G/∂x)+(∂τ/∂h)(∂h/∂x), the chain rule is incomplete. Those terms are generically nonzero when the CBF constraint is active, which is exactly the safety regime. So the adjoint state p in (11) is not the true gradient of the loss, and the claim that backprop through the QP layer via the adjoint equations trains the network is unsupported as written. If the implementation actually unrolls the ODE solver and QP layer with automatic differentiation, the experiments may still be sound, but then Proposition 4 is not the method being evaluated. That distinction needs to be stated.\n\nTwo smaller issues. The abstract says stability is 'ensured,' but the exponential convergence claim is conditional on a parameter that drives the CLF loss to zero; no such parameter is shown to exist for these examples. And the empirical section reports one trajectory per setting, no variance across seeds, and no code commit hash, so the numeric claims are hard to check.\n\nAll that said, the paper is a legitimate building block. The HOCBF extension with learned per-constraint kappas is a nice idea, the appendix documents real failure modes rather than hiding them, and the OptNet/adjoint machinery is the right toolkit even if the composition needs correcting. I'd send it to peer review, but the authors need to fix the gradient derivation or re-label it as the exact adjoint only when the QP dependency on x is included, and then add a numerical check of the gradient against a reference.","headline":"A genuinely useful integration of Neural ODEs with differentiable CBF-QP layers, but Proposition 4 omits the QP layer's state dependence, so the main gradient claim needs repair before the paper's guarantees can be taken at face value.","tokens_in":17858,"tokens_out":2479,"would_cite":false,"duration_ms":24668,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A Neural ODE controller can be trained through a differentiable QP layer that projects each command onto a barrier-function safety constraint, with a Lyapunov loss for stability.","keywords":["Neural ODE","differentiable QP layer","control barrier function","control Lyapunov function","safe control","adjoint method","higher-order CBF","optimal control"],"falsifier":"Run the trained controller from an initial state that approaches the boundary of the safe set while the raw network command pushes inward; if at any reached state the constraint set $\\{u : \\dot{B}(x,u)+\\alpha(B(x))\\ge 0\\}$ is empty, the QP has no feasible output and the claimed safety-by-construction guarantee fails at that point.","tokens_in":16675,"feed_emoji":"🛡️","tokens_out":13161,"duration_ms":106489,"temperature":0.7,"pith_summary":"Opt-ODENet claims that a state-feedback controller can be learned end-to-end by passing the network's raw command through a differentiable quadratic-programming (QP) layer that projects the command onto a control barrier function (CBF) constraint — a scalar safety certificate whose positivity defines the set of safe states — so that safety is enforced as a hard requirement. A control Lyapunov function (CLF, a scalar certificate whose decrease implies convergence) is placed in the loss, and the paper proves that gradients flow backward through both the Neural ODE dynamics and the QP solution by combining the adjoint method with differentiation of the QP's KKT system; this makes the CBF's class-$\\mathcal{K}$ coefficient learnable. If correct, the framework designs safe, stabilizing controllers without a nominal controller, reference trajectories, or large datasets. Experiments on a unicycle obstacle-avoidance task and a higher-order-CBF car-following task support the claim.","feed_headline":"Differentiable QP layer keeps Neural ODE controllers collision-free","feed_subtitle":"A Lyapunov loss drives convergence while a barrier-function projection enforces safety, with no demonstrations or nominal controller","key_machinery":"The load-bearing object is the differentiable CBF-QP projection layer: a quadratic program whose objective is $\\tfrac{1}{2}\\|u-u_{nn}\\|^2$ with $u_{nn}=\\pi(x,\\theta_1)$, and whose inequality constraint is the CBF condition $\\dot{B}(x,u)+\\alpha(B(x),\\theta_2)\\ge 0$. The layer is differentiable because its backward pass differentiates the KKT conditions (15)-(20), producing the sensitivities $\\partial u^*/\\partial q$ and $\\partial u^*/\\partial h$ that the adjoint equations (11) need for $\\nabla_{\\theta_1}\\ell$ and $\\nabla_{\\theta_2}\\ell$. The second mechanism is the CLF-based Lyapunov loss $\\ell=\\mathbb{E}\\left[\\int_{t_0}^{t_f} \\max\\{0,\\dot{V}(x,u)+\\gamma V(x)\\}\\,dt\\right]$; when training drives this loss to zero, the closed-loop system inherits exponential convergence of the form $\\Phi(x(t))\\le \\Phi(x(t_0))e^{-\\kappa t}$.","core_discovery":"On the paper's own terms, the central claim is that for a known control-affine system $\\dot{x}=f(x)+g(x)u$ and a known CBF $B$, the policy $u=\\tau(x,\\theta_1,\\theta_2)$ obtained by solving the QP $\\min_u \\tfrac{1}{2}\\|u-\\pi(x,\\theta_1)\\|^2$ subject to $\\frac{\\partial B}{\\partial x}f(x)+\\frac{\\partial B}{\\partial x}g(x)u+\\alpha(B(x),\\theta_2)\\ge 0$ can be trained directly by backpropagating a CLF-based loss through the Neural ODE and through the QP. Proposition 4 establishes the adjoint equations (11) for the parameter gradients, with the QP sensitivity entering through the derivative of the KKT system (equations (15)-(20)). The learned class-$\\mathcal{K}$ coefficient $\\kappa=13.38$ yields collision-free unicycle trajectories with mean error $0.2036$, the best among the tested safe configurations in Table 1, and in the car-following HOCBF setting the learned per-constraint $\\kappa$ gives zero safety violations and better reward than fixed choices.","pith_inferences":["The paper leaves implicit that the same adjoint-plus-KKT derivation would extend to other differentiable optimization layers (for example, a second-order cone or nonlinear projection), broadening the safety mechanism beyond quadratic programs without changing the training scheme.","Because the unicycle case learns only a scalar class-$\\mathcal{K}$ coefficient while the car-following case needs per-constraint coefficients, a natural extension is to make $\\alpha$ state- or constraint-dependent; the multi-constraint result already hints that a single scalar is a bottleneck.","One could test generalization by training in one obstacle layout and deploying in a different one; if the learned $\\kappa$ transfers, it would suggest the QP projection, not memorized geometry, is carrying the safety guarantee."],"forward_implications":["Deployment can use the trained policy as the final controller: every command is the QP projection of the neural output, so no separate runtime safety filter is needed as long as the QP remains feasible.","Because the CBF class-$\\mathcal{K}$ coefficient is learned, the user does not need to hand-tune how strongly safety intervenes; the optimizer finds a value such as $\\kappa = 13.38$ that balances convergence against conservatism.","The approach removes the need for a nominal safe controller, reference trajectories, or large datasets: training solves the optimal control problem directly from sampled initial states for known dynamics.","If the Lyapunov loss reaches zero, the learned closed-loop system converges exponentially at rate $\\kappa$, so the same certificate used in training doubles as a stability guarantee.","Higher-order CBFs extend the framework to constraints with relative degree greater than one, such as car following under acceleration control."],"supporting_citations":[{"why":"Supplies the Neural ODE model of continuous-depth networks on which the controller and adjoint training are built.","marker":"[Chen et al., 2018]"},{"why":"Supplies the differentiable QP layer whose KKT-based backward pass makes end-to-end training through the safety projection possible.","marker":"[Amos and Kolter, 2017]"},{"why":"Supplies the CBF definition and invariance condition that becomes the inequality constraint of the QP layer.","marker":"[Ames et al., 2019]"},{"why":"Supplies the CLF definition whose derivative condition is turned into the Lyapunov loss.","marker":"[Sontag, 1989]"},{"why":"Supplies the implicit function theorem used to differentiate the QP solution through its KKT system.","marker":"[Dontchev and Rockafellar, 2009]"},{"why":"Supplies the Lyapunov-loss construction whose zero-loss point yields the exponential convergence guarantee.","marker":"[Rodriguez et al., 2022]"},{"why":"Supplies higher-order CBFs that let the framework handle safety constraints with relative degree greater than one.","marker":"[Xiao and Belta, 2019]"},{"why":"Serves as the fixed-CBF-parameter baseline that motivates learning the class-K coefficient.","marker":"[Pereira et al., 2021]"},{"why":"Serves as the no-QP baseline whose collisions demonstrate the need for the hard safety layer.","marker":"[Ip et al., 2024]"},{"why":"Serves as the terminal-loss-only baseline that the CLF-based loss is compared against.","marker":"[Böttcher et al., 2022]"}],"fun_headline_variants":["Neural ODE + QP layer: safe control without nominal models","CBF-embedded QP makes Neural ODEs collision-free","Training Neural ODEs with hard safety constraints via QP","Direct optimal control via differentiable QP and Neural ODE","Safe control learning: QP layer enforces barriers in Neural ODE"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the CBF-QP layer is feasible and differentiable at every state visited during training and deployment — the paper's own appendix records a car-following case where infeasibility made training problematic — and, for the exponential-convergence statement, that the Lyapunov loss can actually be driven to zero.","fun_headline_variants_meta":{"raw":{"variants":["Neural ODE + QP layer: safe control without nominal models","CBF-embedded QP makes Neural ODEs collision-free","Training Neural ODEs with hard safety constraints via QP","Direct optimal control via differentiable QP and Neural ODE","Safe control learning: QP layer enforces barriers in Neural ODE"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000707,"raw_usage":{"total_tokens":3190,"prompt_tokens":953,"completion_tokens":2237,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":569,"completion_tokens_details":{"reasoning_tokens":2147}},"tokens_in":569,"tokens_out":2237,"duration_ms":15642,"temperature":1.0,"reasoning_tokens":2147,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:49:06.126603+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the trained controller from an initial state that approaches the boundary of the safe set while the raw network command pushes inward; if at any reached state the constraint set $\\{u : \\dot{B}(x,u)+\\alpha(B(x))\\ge 0\\}$ is empty, the QP has no feasible output and the claimed safety-by-construction guarantee fails at that point.","supporting_citations":[{"cited_title":"A ‘universal’construction of artstein's theorem on nonlinear stabilization","cited_arxiv_id":null,"evidence_quote":"Supplies the CLF definition whose derivative condition is turned into the Lyapunov loss."},{"cited_title":"Lyanet: A lyapunov framework for training neural odes","cited_arxiv_id":null,"evidence_quote":"Supplies the Lyapunov-loss construction whose zero-loss point yields the exponential convergence guarantee."},{"cited_title":"Control barrier functions for systems with high relative degree","cited_arxiv_id":null,"evidence_quote":"Supplies higher-order CBFs that let the framework handle safety constraints with relative degree greater than one."},{"cited_title":"Safe optimal control using stochastic barrier functions and deep forward-backward sdes","cited_arxiv_id":null,"evidence_quote":"Serves as the fixed-CBF-parameter baseline that motivates learning the class-K coefficient."}],"review_version":1}