{"id":"5da6af59-e652-4111-895d-3dfe6e09c869","arxiv_id":"2504.15830","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A finite-horizon optimal control problem yields a control barrier function with explicit class-K gain, larger safe domain, and reuse under time-varying constraints.","lead":"This paper gives a way to build safety functions, called control barrier functions, by solving a short planning problem that looks ahead in time. The resulting safety function works even when obstacles change over time, without recomputing it.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 6 Step 1 fails to turn the finite-horizon bound (17) into a Dini-derivative bound: eq. (18) needs the optimal control to be right-continuous at 0, which is not assumed.","rationale":"The reader's weakest assumption (local Lipschitz continuity of HT) is a real concern and is acknowledged in Remark 3, but my stress-test found a more direct obstruction inside Step 1 of Theorem 6. Inequality (17) is a finite-horizon bound for a fixed ε; converting it into the infinitesimal Dini bound requires (18), which implicitly assumes the optimal control has a right limit at 0 compatible with the chosen value u*(0). Since the paper allows arbitrary bounded measurable controls, this tangency is not guaranteed, and the proof as written is incomplete even when HT is locally Lipschitz. This is load-bearing because Step 1 is the part of the proof that certifies the CBF condition on the exterior of C. The theorem is repairable by adding a right-continuity assumption on the maximizer or by deriving the Dini bound directly through a dynamic programming argument on small times, and the numerical results and available code support the method's practical promise. I therefore keep the reader's conditional verdict rather than moving to rejection; the paper should be accepted only after the Step 1 argument is fixed.","tokens_in":27295,"tokens_out":21157,"duration_ms":200826,"concrete_test":"Independently re-derive the substitution in (18) from the stated assumptions alone. The step replaces φ(σ;x0,u*) by x0+σf(x0,u*(0)); for a measurable control the error is ∫_0^σ[f(φ(s),u*(s))−f(x0,u*(0))]ds, which is generally O(σ), not O(σ²), unless u* is right-continuous at 0. Exhibit a maximizer of (8) for a simple system (e.g., ẋ=u, U={-1,1}, h(x)=x, F={x≥2}, T=2, γ=0.5) that is not right-continuous at 0, and show the two difference quotients in (18) differ by O(1). If no additional argument yields sup_{u∈U} dHT(x0;f(x0,u)) ≥ γ, the Step 1 certificate fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The proof of Theorem 6, Step 1, only establishes the finite-horizon inequality HT(φ(εT;x0,u*)) − HT(x0) ≥ γεT for the fixed ε selected in (11). To pass from this to the Dini-derivative inequality (19d), the proof invokes (18), which asserts that the Dini derivative along the trajectory equals the Dini derivative in the constant direction f(x0,u*(0)). The justification is the Taylor-type expansion φ(σ;x0,u*) = x0 + σf(x0,u*(0)) + O(σ²). Under the stated assumptions, u* is only bounded and measurable, so φ(σ) − x0 − σf(x0,u*(0)) = ∫_0^σ [f(φ(s),u*(s)) − f(x0,u*(0))]ds, and the integrand need not vanish as s→0 unless u* is right-continuous (or has a right Lebesgue point) at 0. The remainder is then O(σ), not O(σ²), so after division by σ it contributes an O(1) term to the liminf, and the equality in (18) fails even when HT is locally Lipschitz. Consequently, the CBF inequality (10) is not proven for the blue region, which is exactly the part of the domain outside C where the certificate is most needed. This gap is independent of the local Lipschitz premise and more basic than the Step 2 issue: equation (18) is the only bridge from the finite-horizon growth bound to the infinitesimal Dini condition.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a method for synthesizing Control Barrier Functions (CBFs) by solving a finite-horizon max-min optimal control problem. The resulting function HT is claimed to be a CBF in the Dini sense on a domain D that may be larger than its zero superlevel set, with an explicit extended class Ke function alpha. The method is designed to handle input constraints, to provide pointwise values of the CBF without computing it globally, and to be robust to time-varying constraints by yielding a shiftable CBF. The paper includes three variants of the construction, implementation remarks with a discrete-time approximation, and simulations for single/double integrators, a kinematic bicycle, and a unicycle.","tokens_in":27617,"tokens_out":8767,"duration_ms":91331,"significance":"If the main theorem is correct, the paper offers a useful addition to the CBF synthesis literature: it avoids a nominal control law, produces an explicit alpha, allows domain D larger than C, and enables time-varying constraint handling through shiftability. The pointwise computability and the accompanying open-source implementation are concrete strengths. However, the central proof of Theorem 6 has a load-bearing gap in the passage from a finite-horizon growth inequality to the infinitesimal Dini-derivative condition, and a second gap in the treatment of the boundedness assumption. These issues affect the formal certificate of safety for the synthesized function as currently written. The overall idea is promising, but the theoretical core needs repair before the claims can be accepted.","major_comments":[{"comment":"The proof does not justify the transition from the finite-horizon inequality (17) to the infinitesimal Dini-derivative inequality (19d). Equation (18) relies on the expansion phi(sigma;x0,u*) = x0 + sigma f(x0,u*(0)) + O(sigma^2), but u* is only bounded and measurable, and no right-continuity (or right Lebesgue point) of u* at 0 is assumed. The remainder integral of f(phi(s),u*(s)) - f(x0,u*(0)) is then generally O(sigma), not O(sigma^2), so equality (18c) fails even when HT is locally Lipschitz. Consequently the CBF inequality (10) is not established on the 'blue region' outside C, which is the critical part of the domain for the certificate. In addition, (17) is stated for the fixed epsilon chosen in (11); taking the limit sigma->0 in (19d) requires the same inequality for all sufficiently small sigma, which the proof does not show.","section":"Section IV.A, Theorem 6, Step 1, Eq. (18)-(19)"},{"comment":"The boundedness assumption on f is not sufficient for the estimate in (22). The assumption 'for all x in C there exists u in U such that ||f(x,u)|| <= M' is existential in x and does not imply that the chosen constant input u at x0 satisfies ||f(phi(s;x0,u),u)|| <= M for s in [0,sigma], especially if the trajectory leaves C. The proof needs f to be bounded on a neighborhood of C (or on C x U with C compact and U compact) and the input chosen so that the bound holds along the trajectory. As written, the lower bound (23) and hence the conclusion (24d) for the 'green region' are not proven.","section":"Section IV.A, Theorem 6, Step 2, Eq. (22)-(23)"},{"comment":"Definition 1 requires the zero superlevel set C = {x | HT(x) >= 0} to be compact, but the paper does not prove compactness of C and does not include it as an explicit assumption. In the numerical examples, H is the complement of a disk and is unbounded, so compactness is not automatic. Without an added compactness or bounded-domain assumption, HT does not formally satisfy Definition 1 as a CBF.","section":"Definition 1 and Theorem 6"}],"minor_comments":[{"comment":"The sentence 'for all t' in [0,T]' after Eq. (17) is confusing because (17) contains no such parameter; it should be clarified that the argument can be repeated for every smaller positive epsilon' if the infinitesimal passage is intended.","section":"Section IV.A, Eq. (17)"},{"comment":"The local Lipschitz continuity of HT is assumed as a premise and is only checked numerically a posteriori; the paper should state more prominently that Theorem 6 is a verification theorem conditional on this regularity, not a construction that guarantees it.","section":"Section IV.A, Remark 3"},{"comment":"Proposition 10 relies on [55, Theorem 4] and assumes differentiability of HT; the paper should note explicitly that the time-varying CBF claim is not self-contained and that differentiability is not verified for the numerical examples.","section":"Section V.B, Proposition 10"},{"comment":"The notation for the p-norm approximation is unclear: 'p \" 0' appears to be a typo, and the text should state that p is chosen sufficiently large (p >> 0). Additionally, the constant h-tilde requires min_{x in D} h(x) to exist and be finite, which needs an explicit boundedness assumption on D and h.","section":"Section VI.B, Eq. (41) and Algorithm 1"}],"recommendation":"major_revision","confidential_remarks":"The central construction is interesting and the numerical part is substantial, but the proof of Theorem 6 needs a genuine repair, not just a wording change. The Step 1 regularity gap is the most serious issue because it breaks the proof of the CBF property on the region outside C. I would suggest the authors either prove the needed Dini-derivative bound under weaker regularity of the optimal input, or add a right-continuity/selection assumption on u* and verify it for the examples."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a solid, useful extension of their earlier finite-horizon predictive CBF, and the time-varying shiftability result is a nice addition. But the proof of Theorem 6 has a hole that is more serious than the reader's note suggests. The stress-test is right. Equation (18) claims the Dini derivative along the trajectory equals the Dini derivative in the constant direction f(x0,u*(0)), via a Taylor expansion φ(σ)=x0+σf(x0,u*(0))+O(σ²). That expansion needs u* to be right-continuous at 0, or at least to have a right Lebesgue point. The assumptions only give bounded measurable input. Without that, the remainder is O(σ), so after dividing by σ you get an O(1) contribution and the liminf equality fails. This isn't cosmetic: Step 1 is exactly the part of the domain outside C where the certificate is most needed. The paper's own Remark 3 acknowledges the Lipschitz issue, which is also unproven, but the Step 1 gap is independent of that.\n\nThe good parts: the gamma-penalty is a simple idea that buys ascent on a larger domain and makes the class-K function explicit; the shiftability application via their own prior work is credible; the simulations cover nonholonomic systems with input constraints and the code is available. Proposition 5 and the C⊂H argument are clean. Step 2's boundedness issue is patchable by choosing the bounded control and using continuity of f. So the method is probably right, but the main theorem as written is not proven.\n\nWho's this for? Anyone working on CBF synthesis, especially with time-varying constraints or nonholonomic systems. It deserves a serious referee; the gap is fixable and the idea is worth publishing. I'd send it to review, but with a clear request to repair the Step 1 argument or state an extra regularity assumption on u* before acceptance.","headline":"A promising predictive CBF synthesis with a real gap in the proof of the main theorem: Step 1 of Theorem 6 needs right-continuity of the optimal control that isn't assumed.","tokens_in":28162,"tokens_out":4188,"would_cite":true,"duration_ms":39266,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A finite-horizon optimal control problem, solved pointwise, produces a control barrier function that remains valid when constraints change over time, without recomputation.","keywords":["control barrier functions","predictive CBF synthesis","time-varying constraints","Dini derivative","Hamilton-Jacobi reachability","finite-horizon optimal control","shiftable control barrier functions","input constraints"],"falsifier":"Compute HT from (8) for a system satisfying Assumptions 1 and 2 in which the finite-horizon value is nonrobust in the sense of the examples cited in [64]; if on a fine grid of D there is any state x where the supremum over u in U of the Dini derivative dHT(x; f(x,u)) is strictly less than -alpha(HT(x)), or where HT is visibly discontinuous, then the claimed certificate in (10) is not delivered by the synthesis.","tokens_in":27069,"feed_emoji":"🛡️","tokens_out":7872,"duration_ms":68925,"temperature":0.7,"pith_summary":"The paper sets out to prove that a control barrier function can be synthesized pointwise as the value of a finite-horizon optimal control problem, and that this value function is itself a CBF in the Dini sense on a domain larger than its safe set. The payoff is reusability: because the CBF is defined through a prediction horizon rather than a fixed constraint shape, adding a time-varying offset within a prescribed bound produces another valid CBF, so moving or expanding constraints do not force a full recomputation. The authors also give an explicit construction of the extended class K function alpha that appears in the CBF inequality, and show that alpha can instead be chosen directly as a design parameter. If the construction holds, safety filters can be built for systems with input constraints and weak controllability, without needing a nominal backup controller.","feed_headline":"Safety barrier adapts to moving constraints without recomputation","feed_subtitle":"A single finite-horizon optimization builds a control barrier function that adapts as constraints change over time.","key_machinery":"The central object is the finite-horizon value function HT defined in (8): HT(x0) = max over input trajectories in U[0,T] of min over t in [0,T] of h(x(t)) - gamma t, with the state trajectory required to hit the known set F at some time within the horizon. The machinery consists of two complementary bounds on its Dini derivative: near the boundary of the safe set the gamma penalty forces an asymmetric ascend of at least gamma along the optimal trajectory, while inside the safe set a Lipschitz bound LM bounds the worst possible descent; these are combined in (20) into an explicit extended class K function alpha so that sup_u dHT(x; f(x,u)) >= -alpha(HT(x)) holds on all of D. The same structure is what makes HT shiftable in the sense of [55], since the ascend margin near the boundary tolerates the addition of a bounded time-varying offset.","core_discovery":"The central claim, formalized as Theorem 6, is that for a Lipschitz constraint function h and a known subset F of an (unknown) forward control invariant set V, the function HT(x0) = max over input trajectories of the minimum over the horizon of h(x(t)) - gamma t, subject to reaching F within time T, is a control barrier function in the Dini sense on the whole domain D. This holds when Assumptions 1 and 2 are satisfied, T >= tau, gamma < delta/T, f is bounded on the safe set, and HT happens to be locally Lipschitz. The negative gamma-t term is what makes the proof work: along the optimal predictive trajectory the value strictly increases at rate at least gamma on a neighborhood of the safe set's boundary, and the boundedness of f plus Lipschitz continuity bounds any possible decrease away from the boundary. Those two bounds are glued into an explicit extended class K function alpha in (20), and a modified synthesis (29) lets alpha itself be chosen by the designer. Since the CBF is defined on D rather than only on its zero super-level set, HT is shiftable: HT + lambda(t) remains a CBF for any time-varying lambda satisfying (36), which is how the method handles time-varying constraints.","pith_inferences":["The pointwise evaluation property suggests a natural testable extension: fit the computed values of HT with a neural network or another regressor and certify the learned function by checking the Dini condition on a grid; the paper only uses linear interpolation in its examples.","The gamma-based ascend margin near the boundary could plausibly confer a form of input robustness: disturbances smaller than the margin would not immediately destroy the CBF inequality, though the paper does not analyze robustness to model error.","The local Lipschitz assumption on HT is the soft spot; a useful extension would be to identify structural conditions on f and U beyond the nonrobust examples cited in [64] under which this regularity is guaranteed a priori rather than verified numerically.","Since HT is computed only on a subset and then interpolated, the actual safety guarantee for interpolated points depends on the interpolation error; establishing a Lipschitz bound on HT would let the interpolation error be bounded and the filter made certifiably safe on the full grid-free domain."],"forward_implications":["If the central theorem is correct, one CBF computation for a static constraint automatically yields valid CBFs for a whole family of time-varying constraints, through HT(x) + lambda(t) with lambda obeying the explicit rate condition (36).","The explicit class K function means the safety margin encoded in the CBF inequality is known, and in the alpha-based version (29) it becomes a direct design parameter rather than an implicit outcome of the computation.","Because the method needs no nominal control law and only a known subset F of an invariant set, it applies to non-locally-controllable and nonholonomic systems such as the bicycle and unicycle examples, and automatically accounts for input constraints.","Longer prediction horizons enlarge the guaranteed safe sets C_lambda,T, so the horizon T can be traded off against the size of the region in which safety is certified.","Pointwise evaluation means the CBF does not have to be gridded over the whole domain; this is what allows the method to be parallelized and to be combined with symmetry reduction in the follow-up work."],"supporting_citations":[{"why":"Supplies the preliminary finite-horizon synthesis and Proposition 4, which justifies the existence of the reaching time tau used in Assumption 2.","marker":"[56]"},{"why":"Provides the definition of shiftable CBFs and the theorem used in Proposition 10 to turn HT into a time-varying CBF.","marker":"[55]"},{"why":"Motivates the Hamilton-Jacobi reachability viewpoint that the value function inherits, and serves as the main alternative that yields time-dependent rather than time-invariant barriers.","marker":"[48]"},{"why":"Grounds Remark 3, where the paper concedes that finite-horizon optimal control value functions can be nonrobust, motivating the local Lipschitz assumption.","marker":"[64]"},{"why":"Sets up the standard differentiable CBF and safety-filter framework that the paper extends to the Dini-sense and predictive setting.","marker":"[2]"},{"why":"Represents the backup-CBF approach that requires a nominal control law, which the present method avoids.","marker":"[46]"}],"fun_headline_variants":["Predictive CBF synthesis adapts to changing constraints without recomputation","One-shot CBF synthesis handles time-varying constraints without recompute","Predictive CBFs: adapt to moving constraints, no recomputation needed","Finite-horizon CBF synthesis adapts to time-varying constraints"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the value function HT is locally Lipschitz continuous; the paper assumes this in Theorem 6 and only verifies it numerically after the fact, noting in Remark 3 that finite-horizon optimal control values can be nonrobust.","fun_headline_variants_meta":{"raw":{"variants":["Predictive CBF synthesis adapts to changing constraints without recomputation","One-shot CBF synthesis handles time-varying constraints without recompute","Predictive CBFs: adapt to moving constraints, no recomputation needed","Finite-horizon CBF synthesis adapts to time-varying constraints"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00098,"raw_usage":{"total_tokens":4212,"prompt_tokens":1050,"completion_tokens":3162,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":666,"completion_tokens_details":{"reasoning_tokens":3085}},"tokens_in":666,"tokens_out":3162,"duration_ms":20392,"temperature":1.0,"reasoning_tokens":3085,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:17:36.900195+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compute HT from (8) for a system satisfying Assumptions 1 and 2 in which the finite-horizon value is nonrobust in the sense of the examples cited in [64]; if on a fine grid of D there is any state x where the supremum over u in U of the Dini derivative dHT(x; f(x,u)) is strictly less than -alpha(HT(x)), or where HT is visibly discontinuous, then the claimed certificate in (10) is not delivered by the synthesis.","supporting_citations":[{"cited_title":"Construction of control barrier functions using predictions with finite horizon,","cited_arxiv_id":null,"evidence_quote":"Supplies the preliminary finite-horizon synthesis and Proposition 4, which justifies the existence of the reaching time tau used in Assumption 2."},{"cited_title":"Examples when nonlinear model predictive control is nonrobust,","cited_arxiv_id":null,"evidence_quote":"Grounds Remark 3, where the paper concedes that finite-horizon optimal control value functions can be nonrobust, motivating the local Lipschitz assumption."},{"cited_title":"Backup control barrier functions: Formulation and comparative study,","cited_arxiv_id":null,"evidence_quote":"Represents the backup-CBF approach that requires a nominal control law, which the present method avoids."}],"review_version":1}