{"id":"5a813b3b-eb29-4b4e-8fa5-4796bfbf8ec4","arxiv_id":"2508.19417","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A rigorous optimal control formulation for a mixed-autonomy platoon with the Bando-FtL model is shown to have a minimizer and is solved by adjoint gradient descent, with simulations showing large reductions in acceleration and fuel use for one to five autonomous vehicles.","lead":"This paper shows how a few autonomous cars can be optimally driven to smooth out stop-and-go traffic waves ahead of a platoon of human-driven cars. The authors prove the optimization problem is solvable and simulate one to five robot cars cutting total acceleration by 70-85% and fuel use by 16-26% on a real highway trajectory.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Numerical effectiveness claim not validated: Bando-FtL constants (vmax, ds, l) are unreported, the human model is uncalibrated, and no benchmark controller is compared; the 70.4%/84.6% reductions in Table 1 may be artifacts of the chosen model.","rationale":"A good-faith reading shows the theoretical core is largely solid: Lemma 4.1 and Theorem 4.3 provide a credible existence proof, and the main mathematical dependencies (Theorem 3.3 from [17]) are cited rather than derived. I found only fixable typos, such as the missing square in the amin formula in eqs. (16)–(17) and a minor index error in the d_min definition in Lemma 4.1. These do not change the verdict. The reader's weakest assumption was the imported Theorem 3.3; while that is a legitimate verification concern, the cited theorem is peer-reviewed and likely correct. The more load-bearing weakness is the numerical validation: the experiments are irreproducible and uncalibrated, and the paper itself admits the lack of a benchmark. Since the abstract and title emphasize applications, this directly threatens the central applied claim. I therefore partially agree with the reader, and I recommend keeping the conditional verdict, with the numerical-validation conditions given priority.","tokens_in":22006,"tokens_out":16906,"duration_ms":187195,"concrete_test":"Rerun Experiment I under two changes: (i) report and fix all model constants (vmax, ds, l) and calibrate α, β, and V to the I-24 car-following data via maximum likelihood on observed gaps and velocities; (ii) compare the optimized controller against an existing baseline (e.g., the PI controller of Stern et al. 2018 or the MPC of Wu & Bayen 2022) on the same leader trajectory. If the L2-acceleration reduction relative to the calibrated all-human baseline drops below roughly 20%, or if a standard controller achieves comparable or better smoothing, then the paper's claim of 'effectiveness' is not established and the numerical section requires major revision.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central applied claim—that optimal control of AVs yields 70.4% and 84.6% reductions in total L2 acceleration (Table 1)—rests on numerical experiments whose human-driving model is neither fully specified nor calibrated. Section 6.2 gives α=0.1 and β=525, but V in eq. (8) depends on unstated vmax and ds, and vehicle length l is also never assigned; initial conditions are likewise omitted. Without these constants, the experiment is not reproducible. More importantly, the Bando-FtL model is not calibrated to the I-24 trajectory; the baseline is the model's own unstable behavior, so the percentage improvement measures smoothing relative to an artificial surrogate, not to real human driving. The conclusion (Section 6.3) explicitly states that benchmarking against existing longitudinal controllers is 'a natural next step,' confirming that no such comparison is made. The abstract also promises experiments on 'distributions of controlled vehicles' that do not appear in the paper. These gaps mean the headline numerical results are currently unsupported as evidence of real-world effectiveness, even though the theoretical existence theorem (Theorem 4.3) is credible and, modulo fixable typos in eqs. (16)–(17), the proof structure is sound.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper studies optimal control of a single-lane mixed-autonomy platoon with a prescribed leader trajectory. Human-driven vehicles follow the Bando-Follow-the-Leader (Bando-FtL) model, while autonomous vehicles have directly controlled acceleration. The authors define an admissible control set with acceleration bounds, nonnegative velocity, and a minimum safety gap for the AVs. Their main theoretical result (Theorem 4.3) asserts existence of a minimizer for a convex running cost and C^1 terminal cost, relying on well-posedness and headway/velocity bounds for Bando-FtL from [17]. They propose a direct adjoint-based gradient descent method with a penalty formulation for state constraints, and report numerical experiments on an I-24 trajectory showing large reductions in total L2 acceleration and fuel consumption as the number of AVs increases. The exposition is generally clear, but several load-bearing proof steps and the numerical validation need correction.","tokens_in":22350,"tokens_out":15480,"duration_ms":165122,"significance":"If completed and corrected, the existence theorem is a meaningful contribution: it transfers the known well-posedness of Bando-FtL into an optimal-control guarantee for a convex-in-control problem under explicit safety constraints. The reliance on [17] is legitimate and not circular, and the one-sided coupling structure is exploited cleanly. The numerical algorithm is standard in spirit, but the adjoint derivation currently has sign errors, and the experiments are under-specified; the applied claims in Table 1 are therefore not yet supported. The framework is potentially useful for benchmarking AV smoothing controllers, provided the numerical section is made reproducible and compared against existing controllers.","major_comments":[{"comment":"The constructed braking control has a dimensional error: amin := -v° / (2d) has units 1/s, not m/s^2. Solving (15) gives amin = -(v°)^2 / (2(x_l(0)-x_1(0)-l-dsafe)) and T* = 2(...)/v°. As written, with v°=10 m/s and d=20 m, the claimed amin=-0.25 m/s^2 stops far too late. The zero-velocity case v°=0 is also not covered. Since U ≠ ∅ is needed for Theorem 4.3, the proof must be corrected; this is fixable but not merely cosmetic.","section":"§3, Corollary 3.4, Eq. (16)-(17)"},{"comment":"The final chain J(u*) = lim_n J(y_n,u*) ≤ lim_n J(y_n,u_n) is not justified as stated. The pointwise inequality J(y_n,u*) ≤ J(y_n,u_n) need not hold for a weak-star limit. The correct argument is to use convexity and the established uniform convergence of states to show J(u*) ≤ liminf_n J(u_n). Please rewrite the limiting step explicitly. This is load-bearing because the existence conclusion depends on it.","section":"§4.1, Theorem 4.3 proof"},{"comment":"The variational derivative of the constraint term in H should be ζ^T(δ\\dot y - f_y δy - f_u δu), not ζ^T(δ\\dot y + f_y δy + f_u δu). With the printed signs, the adjoint equation and the gradient G_i = L_{u_i} - ζ_{i+|I|} are incorrect: for L(u)=u^2, S=0, one obtains G=2u, which is the direction of steepest ascent, not descent. This sign error affects the core of Algorithm 1 and the numerical results. Please verify and correct the adjoint derivation.","section":"§5.1, Eq. (30)-(32), (40)"},{"comment":"The numerical validation is not reproducible and does not support the headline reductions. The optimal velocity function (8) depends on vmax and ds, and the dynamics depend on l, but none are reported; initial conditions are also omitted. The Bando-FtL parameters α=0.1, β=525 are not calibrated to the I-24 data, and no baseline controller is compared (the conclusion acknowledges this). Since the objective being optimized is the same L2-acceleration metric reported in Table 1, the acceleration reductions are not independent evidence; the fuel metric is independent but inherits the uncalibrated dynamics. The abstract also promises experiments on 'distributions of controlled vehicles', but only one fixed distribution is tested.","section":"§6.2, Table 1"}],"minor_comments":[{"comment":"Typo: 'saftey' should be 'safety'.","section":"Definition 2.7"},{"comment":"The grid indexing is inconsistent: p is called a positive real in (47), but it is a count; in (53), t_p=T should presumably be t_r, and the quantifier uses j while the set index is i/r.","section":"§5.3, Eq. (47)-(53)"},{"comment":"The first-variation formula should include \\tilde u(t) in the L_u term and evaluate the terminal term at \\tilde u(T); as written the expression is dimensionally inconsistent.","section":"§5.1, Eq. (28)"},{"comment":"The Bando-FtL model is attributed to [12], but [12] is the Delle Monache et al. reference; the original Bando model is [4] and the Bando-FtL well-posedness is [17]. Please check the citation.","section":"Definition 3.1 and references"},{"comment":"The text says box constraints on the control are not imposed in the experiments. This means Theorem 4.3, which assumes controls in U with finite amin, amax, does not directly cover the numerical problem; a remark should clarify the theoretical status of the unconstrained instance.","section":"§6.1"}],"recommendation":"major_revision","confidential_remarks":"The core theoretical framework is sound and the paper is potentially publishable after the corrections above. The sign error in §5.1 appears typographical but must be fixed before the numerical claims can be trusted. I do not see a circularity problem with using Theorem 3.3 from [17]; the dependence should be stated more explicitly. The numerical section needs full parameter reporting and at least one benchmark comparison to justify the reported percentage reductions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The existence theorem is the real contribution. Theorem 4.3 gives a minimizer for the Bando-FtL optimal control problem via weak-star compactness plus a continuity lemma that uses the uniform headway bounds from the model's well-posedness. That fixes a genuine gap in [52], whose compactness claim is wrong for box-constrained controls in L^p. The fix is clean: keep weak-star convergence and add convexity in the control. This part is solid and publishable.\n\nThe numerical section is not yet support for the Table 1 claims (70-85% acceleration, 16-26% fuel). The experiment is under-specified: vmax, ds, and l are never stated, initial conditions are omitted, and the Bando-FtL parameters are chosen, not calibrated, to the I-24 trajectory. The baseline is the model's own unstable behavior, so the percentages are relative to an artificial human surrogate, not to real driving. No benchmark against existing controllers is given; the conclusion says that's future work. The abstract promises distribution experiments that are not in the paper. This is a real weakness in the applied framing, not a nitpick. The fuel metric is independent of the optimized objective, which earns some credit, but it inherits the same unvalidated model.\n\nThe small stuff: eqs. (16)-(17) are missing squares on v0; likely a typo, but dimensionally off as printed. The reliance on [17] is fine, since it's a published, parameter-free well-posedness result; not circular. The whole theory sits on the assumption that the headway bounds hold for any admissible leader; that's explicit, and if it fails the proof collapses. Acceptable as an assumption, but worth stating as a limitation in the final version.\n\nWho this is for: specialists in ODE optimal control for car-following, and mixed-autonomy researchers who want a theoretical baseline. The theory deserves a serious referee; the numerics need a revision pass: report all constants, benchmark against a simple controller, calibrate or justify parameters, ship code and data, and either add the promised experiments or cut them from the abstract.\n\nMy recommendation: send it to review. Conditional acceptance is the right call, with the numerical revisions as conditions. The theoretical core holds up.","headline":"Solid existence theory fixes a gap in prior work; the numerical claims are not yet supported by the experiments as reported.","tokens_in":22899,"tokens_out":3844,"would_cite":true,"duration_ms":41040,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":null,"created_at":"2026-08-05T15:48:11.643367+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":null,"supporting_citations":[],"review_version":1}