{"id":"5fcc46f3-4db7-4c85-84d0-b0848cfe616b","arxiv_id":"2502.03695","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"CiMPCC adds a curvature-derived reference velocity term to MPCC and reports 11.4-12.5% faster lap times on a 1:10 scale autonomous race car.","lead":"A racing-car controller tweak adds the track's curvature to the MPCC optimizer's velocity target, slowing the car in sharp turns and speeding it on straights. In 17 laps on a self-built 1:10 scale test car, lap time fell 11.4-12.5% versus two baseline planners.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Lap-time gain may reflect asymmetric reference-velocity tuning rather than the curvature-integration mechanism, since MPCC's reference is 3 m/s while CiMPCC's upper bound is 3.8 m/s from an expert lap.","rationale":"The paper has real strengths: physical-vehicle experiments, 17 laps, code release, and a clear mechanism. However, the experimental comparison is not controlled with respect to the reference-velocity settings: the baseline MPCC is given a constant 3 m/s reference while CiMPCC is given an expert-derived 3.8 m/s upper bound. This asymmetry is a confound. The curvature-aware cost term may still help, but the paper's reported 11.8% mean lap-time improvement cannot be cleanly attributed to it from the presented data. The reader's conditional verdict is appropriate; I would not reject because the method is plausible and the code is available for further testing, but I would demand the ablation/fair-baseline experiment before accepting a general claim.","tokens_in":10541,"tokens_out":5785,"duration_ms":53696,"concrete_test":"Run an ablation and a fair baseline: (1) Set MPCC's reference velocity to uref_vp = 3.8 m/s (and uref_vl = 4.18, with R2 re-tuned to match CiMPCC's straight-line behavior) and repeat the 17-lap test on the same track; if the gap drops below ~2%, the curvature mapping is not the main source of improvement. (2) Disable the curvature dependence in CiMPCC by setting g(K_n) constant (e.g., beta = 0.5) while keeping the same vbar/v; compare lap times to isolate the effect of the curvature profile.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (abstract; Table I) is that the proposed curvature-integrated MPCC reduces lap time by 11.4-12.5% vs. MPCC and RDM+OTG. The most load-bearing assumption is that this improvement is caused by the curvature-to-velocity mapping in Eq. (9)-(10), not by the reference-velocity tuning used in the comparison. In Section V, the MPCC baseline is run with uref = [3.3, 0, 3] m/s (Table II), while CiMPCC is given vbar_vp = 3.8 m/s, 'obtained as the fastest velocity at which the expert drives the vehicle to complete a single lap,' and vbar_vl = 4.18 m/s. The paper does not test MPCC with the same aggressive straightaway reference, nor does it report the mapping parameters alpha and w or any uncertainty/error bars. If MPCC had been given vbar_vp = 3.8 m/s, it might achieve comparable straight-line speed and a much smaller lap-time gap; then the claimed benefit would be attributable to a better hand-tuned reference velocity, not to the curvature-integration mechanism. This also raises transferability: the method's performance depends on track-specific expert-derived bounds, which may not generalize.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes Curvature-Integrated MPCC (CiMPCC), a local trajectory planning method for autonomous racing that augments the MPCC cost function with a reference velocity profile obtained by mapping a smoothed, normalized curvature of the racetrack centerline through a Gaussian function. The method is validated on a 1:10 scale F1TENTH vehicle on a custom track over 17 laps. The authors report a 11.4%–12.5% reduction in lap time and a 14.9%–17.3% increase in mean velocity compared with a traditional MPCC baseline and an RDM+OTG global planning baseline. Code is made available.","tokens_in":10846,"tokens_out":2660,"duration_ms":25021,"significance":"If the claimed improvements are attributable to the proposed mechanism, the paper offers a simple, practical modification to MPCC that could benefit autonomous racing and related local planning problems. The use of a physical vehicle, the release of code, and the real-time computation analysis are concrete strengths. However, the central claim depends on a fair experimental comparison and on the reproducibility of the curvature-to-velocity mapping, both of which are currently not fully established.","major_comments":[{"comment":"The comparison between MPCC and CiMPCC is confounded by asymmetric reference-velocity tuning. MPCC uses uref = [3.3, 0, 3] m/s, while CiMPCC is given vbar = [4.18, 3.8] m/s, where 3.8 m/s is 'the fastest velocity at which the expert drives the vehicle.' Since the new term in Eq. (10) explicitly penalizes deviations from vbar, CiMPCC is effectively handed a more aggressive speed reference than MPCC. The paper does not report an experiment in which MPCC is given the same vbar (or any reference above 3.3 m/s). Without such a control, the lap-time gain may reflect the expert-derived reference speed rather than the curvature-integration mechanism itself.","section":"Section V, Table II and Eq. (10)"},{"comment":"The two key tuning parameters of the proposed mapping are not reported: the MAF window width w in Eq. (7) and the exponent coefficient alpha in Eq. (9). Both directly determine the shape of the reference velocity profile and thus the planned velocity. Without these values, the method is not reproducible and it is impossible to assess whether the chosen mapping is a reasonable or an unusually well-tuned instance.","section":"Section IV-B and IV-C"},{"comment":"The experimental results are presented as max/min/mean over 17 laps with no standard deviations, confidence intervals, or statistical tests. The claimed 11.4%–12.5% lap-time improvement therefore lacks an assessment of run-to-run variability. Given that 17 laps is a modest sample, error bars or a paired significance test are needed to support the central quantitative claim.","section":"Section V, Table I"},{"comment":"The paper does not compare CiMPCC against the closest prior work, the curvature-aware MPCC of Lyons and Ferranti (reference [9]), which is discussed in the introduction and related work. Since that method also incorporates curvature into MPCC, a direct comparison is necessary to support the novelty and practical advantage of CiMPCC. Without it, the contribution over the existing curvature-aware MPCC is not demonstrated.","section":"Section II and Section V"},{"comment":"The reference velocity bounds vbar and v are derived from a single expert lap on the same test track and then used as fixed parameters in the cost function. This raises a transferability concern: the method's performance may depend on having an expert demonstration on every new track, and the paper provides no analysis of sensitivity to these bounds or to the choice of alpha and w. A small sensitivity study (e.g., varying alpha and vbar) would materially strengthen the claim that the method generalizes beyond the single reported track.","section":"Section V, Eq. (10) and Table II"}],"minor_comments":[{"comment":"The discrete curvature formula in Eq. (5) is not the standard finite-difference curvature of a plane curve; the authors should clarify the approximation used and whether the denominator uses the chord length or something else.","section":"Section III, Eq. (5)"},{"comment":"The baseline method is inconsistently named: 'RDM+OTG' in Table I and Fig. 7(c) appears as 'RMD+OTG' in Fig. 7(c) caption and in the text near Fig. 7 ('RMD+OTG'). Please unify the notation.","section":"Throughout"},{"comment":"The sentence 'The trajectory planned by RDM+OTG is tracked by the proposed controller in [28]' is ambiguous; it should read 'the controller proposed in [28]'.","section":"Section V"},{"comment":"Fig. 6(a) is referenced as showing 'control inputs of MPCC and CiMPCC in a complete lap,' but the y-axis label is 'Velocity [m/s]' and the legend indicates vl. Consider clarifying that this is the planned longitudinal velocity, not the steering input.","section":"Fig. 6"},{"comment":"The notation v with underbar and overbar is overloaded: v is used both as a general velocity vector and as the lower bound, and the boldface is not used consistently. Please introduce distinct symbols (e.g., v_low and v_high) to avoid confusion.","section":"Section IV-C, Eq. (10)"}],"recommendation":"major_revision","confidential_remarks":"The core idea is reasonable and the physical experiments are a positive feature, but the central quantitative claim is currently not convincingly attributed to the curvature-integration mechanism because of the asymmetric reference-velocity tuning and missing parameters. The paper would be significantly improved by adding a same-reference MPCC baseline, reporting alpha and w, including uncertainty statistics, and comparing against [9]. These are within the scope of a revision; hence major_revision rather than reject."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the paper ships real evidence: 17 laps on a 1:10 F1TENTH-style car, code on GitHub, and a clear description of a simple MPCC modification. Second, the headline lap-time gain is probably inflated by an asymmetric comparison: CiMPCC is given a reference velocity of 3.8 m/s obtained from an expert lap, while the MPCC baseline is tuned to a 3.3 m/s reference. That alone could explain a large chunk of the 11–12% gap.\n\nWhat is actually new: the specific mechanism — normalize smoothed centerline curvature, map it through a Gaussian, and add a velocity-tracking term to the MPCC cost — is simple and easy to reproduce. The paper describes the implementation well and reports computation-time statistics that show real-time feasibility. That is more than many racing papers do. The closest prior work, curvature-aware MPCC [9], is cited but not used as a baseline, which is a missed opportunity.\n\nThe soft spots are real but not fatal. First, the reference-velocity asymmetry: Table II shows MPCC with uref = [3.3, 0, 3] while CiMPCC gets vbar = [4.18, 3.8]. The authors say both were finetuned, but they never test MPCC with the same aggressive reference, so the claimed improvement is not cleanly attributable to the curvature mechanism. Second, the mapping parameters alpha and the MAF window w are not reported; those directly shape the velocity profile. Third, the velocity bounds vbar and v are fitted to the expert lap on this single track, which makes transferability unclear. Fourth, no error bars or significance tests, though the 17-lap spread is tight enough to suggest the effect is consistent.\n\nWho this is for: people working on local trajectory planning for autonomous racing, especially on small-scale platforms. They will get a clear, reproducible idea and a reasonable starting point for further experiments. The paper is not a major theoretical advance, but it is honest engineering with real hardware results.\n\nRecommendation: send it to peer review. A serious referee should ask for the missing ablations — MPCC with a higher reference, alpha and w values, and a comparison against [9] — before the quantitative claims are accepted. As written, the central mechanism is plausible but not fully isolated from the tuning asymmetry.","headline":"Solid incremental MPCC extension with real vehicle experiments, but the headline lap-time gain is likely inflated by an asymmetric reference-velocity comparison.","tokens_in":11359,"tokens_out":2144,"would_cite":false,"duration_ms":21197,"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":"This paper claims that adding the racetrack centerline's smoothed curvature to the MPCC objective as a Gaussian reference-velocity profile reduces autonomous racing lap time by 11.4%–12.5%.","keywords":["autonomous racing","model predictive contouring control","trajectory planning","curvature-aware planning","velocity profile optimization","local planning","1:10 scale vehicle","Gaussian curvature mapping"],"falsifier":"Run an ablation on the same physical track and vehicle with the curvature mapping disabled (set $\\beta$ constant) while keeping the expert-derived velocity bounds in the cost; if lap time remains near the CiMPCC value, the curvature integration is not the cause of the reported gain. A second check is to repeat the comparison on a second track with different curvature statistics and see whether the 11–12% improvement persists.","tokens_in":10324,"feed_emoji":"🏎️","tokens_out":7006,"duration_ms":57302,"temperature":0.7,"pith_summary":"Traditional model predictive contouring control (MPCC) plans a racing trajectory online by maximizing progress along the racetrack centerline, but it ignores how sharply the centerline bends. This paper claims that feeding the centerline's smoothed, normalized curvature into the MPCC cost function—through a Gaussian mapping to a reference velocity profile—lets a 1:10 scale autonomous vehicle brake earlier into sharp corners and carry more speed out of them. On a physical track with four large curvature turns, the resulting CiMPCC method reports a mean lap time 11.8% shorter than the baseline MPCC and 11.7% shorter than a global path-velocity-decomposition planner, with mean velocity higher by 15.2% and 17.2%. If the claim transfers, curvature-aware online planning is a cheap, parameter-light way to close part of the gap between local and global racing trajectory methods.","feed_headline":"Curvature-aware MPCC cuts lap time 11 to 12 percent","feed_subtitle":"The curvature-aware cost makes the racer brake earlier and carry more speed out of each corner.","key_machinery":"The central object is the normalized smooth curvature (NSC) of the centerline, $K_i^n=(K_i-K_{\\min})/(K_{\\max}-K_{\\min})$, produced by moving-average smoothing of the discrete curvature and min-max normalization. A Gaussian map $g(K_i^n)=e^{-\\alpha (K_i^n)^2}$ converts NSC into a reference velocity weight; the current value $\\beta=g(K_{\\text{cur}}^n)$ blends the cost between an aggressive velocity bound $\\bar{v}$ and a safe bound $\\underline{v}$. This continuous, bounded mapping is what lets the optimizer slow down in sharp bends without stalling and accelerate on straights, and it is the component the paper adds to the standard MPCC objective.","core_discovery":"The paper's central claim is that integrating the racetrack centerline's normalized smooth curvature (NSC) into the MPCC objective improves racing performance. The curvature is smoothed with a moving average filter, normalized to $[0,1]$, and mapped through $g(K_i^n)=e^{-\\alpha (K_i^n)^2}$ to a reference overall velocity. That reference enters the cost as a blending term: the planned velocity is pulled toward the aggressive upper bound when curvature is low and toward a safe lower bound when curvature is high, with a lower truncation coefficient that prevents over-braking. The authors assert that this mechanism, validated in seventeen laps on a physical 1:10 scale vehicle, reduces mean lap time by 11.8% versus MPCC and 11.7% versus RDM+OTG while raising mean velocity by 15.2% and 17.2%.","pith_inferences":["Editorial inference: the reported 11–12% gain likely depends on how much headroom the expert-driven aggressive velocity bound leaves; on a track where the expert baseline is already near the vehicle's limit, the curvature mapping would have less to add.","A testable extension would replace the fixed expert-derived bounds with an online-adapted bound, such as one updated from the achieved velocity of the previous lap, which would separate the mapping's contribution from the bound-tuning contribution.","The same normalized-curvature idea could transfer to overtaking by computing curvature of a reference race line rather than the centerline, which the paper lists as future work.","On full-scale vehicles, the kinematic model used here would need a dynamic model; the curvature-velocity mapping should still work, but the safe and aggressive bounds would need to encode tire friction limits."],"forward_implications":["On tracks with sharp curvature, CiMPCC should produce lap-time gains over plain MPCC whenever the expert-derived velocity bounds reflect the track's speed limits.","The curvature-to-reference-velocity term modifies only the objective, so it can be added to other model predictive contouring formulations without changing their constraints.","Because over 95% of CiMPCC solve times are under 0.0206 seconds, the method is compatible with real-time control on small-scale vehicles.","The planned velocity profile stays continuous, so it can be passed to lower-level controllers without a separate smoothing step.","CiMPCC's advantage over global planners is largest at corner entry and exit, where local curvature information lets it respond to the current vehicle state."],"supporting_citations":[{"why":"Supplies the base MPCC formulation—contouring error, lag error, projected velocity maximization, and control smoothness—that CiMPCC extends.","marker":"[8]"},{"why":"Identifies the curvature-ignoring failure mode of standard MPCC and proposes a curvature-aware variant that the paper positions as untested at racing speeds.","marker":"[9]"},{"why":"Provides the moving average filter used to smooth the raw centerline curvature before normalization.","marker":"[21]"},{"why":"Supplies the Gaussian mapping form adapted here to create the curvature-to-reference-velocity profile.","marker":"[22]"},{"why":"Provides the 1:10 scale open-source vehicle and software stack used for the physical experiments.","marker":"[23]"},{"why":"Supplies the nonlinear programming solver that runs the CiMPCC optimization in real time on the vehicle's computer.","marker":"[24]"},{"why":"Supplies the minimum-curvature path planner used in the RDM+OTG global-planning baseline.","marker":"[10]"},{"why":"Supplies the Frenet-frame optimal trajectory generator used as the velocity-planning half (OTG) of the comparison baseline.","marker":"[27]"},{"why":"Supplies the pursuit controller used to track the RDM+OTG trajectory in the comparison.","marker":"[28]"}],"fun_headline_variants":["Curvature-aware MPCC speeds up laps by 11–12%","CiMPCC: Curvature planning trims lap times by 12%","Racing AI uses track curvature to cut lap time 12%","New MPCC variant slashes lap time 11-12% on F1TENTH","Curvature-integrated MPCC wins 12% faster lap times"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper's lap-time gain rests on the assumption that the aggressive and safe velocity bounds, taken from the fastest expert-driven lap on this one track, are a fair and transferable benchmark rather than the main source of the improvement.","fun_headline_variants_meta":{"raw":{"variants":["Curvature-aware MPCC speeds up laps by 11–12%","CiMPCC: Curvature planning trims lap times by 12%","Racing AI uses track curvature to cut lap time 12%","New MPCC variant slashes lap time 11-12% on F1TENTH","Curvature-integrated MPCC wins 12% faster lap times"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00072,"raw_usage":{"total_tokens":3253,"prompt_tokens":984,"completion_tokens":2269,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":600,"completion_tokens_details":{"reasoning_tokens":2166}},"tokens_in":600,"tokens_out":2269,"duration_ms":15230,"temperature":1.0,"reasoning_tokens":2166,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T04:04:34.338781+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run an ablation on the same physical track and vehicle with the curvature mapping disabled (set $\\beta$ constant) while keeping the expert-derived velocity bounds in the cost; if lap time remains near the CiMPCC value, the curvature integration is not the cause of the reported gain. A second check is to repeat the comparison on a second track with different curvature statistics and see whether the 11–12% improvement persists.","supporting_citations":[{"cited_title":"Optimization-based au- tonomous racing of 1: 43 scale rc cars,","cited_arxiv_id":null,"evidence_quote":"Supplies the base MPCC formulation—contouring error, lag error, projected velocity maximization, and control smoothness—that CiMPCC extends."},{"cited_title":"Curvature-aware model predictive contour- ing control,","cited_arxiv_id":null,"evidence_quote":"Identifies the curvature-ignoring failure mode of standard MPCC and proposes a curvature-aware variant that the paper positions as untested at racing speeds."},{"cited_title":"The scientist and engineer’s guide to digital signal processing,","cited_arxiv_id":null,"evidence_quote":"Provides the moving average filter used to smooth the raw centerline curvature before normalization."},{"cited_title":"Deep drifting: Autonomous drifting of arbitrary trajectories using deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the Gaussian mapping form adapted here to create the curvature-to-reference-velocity profile."},{"cited_title":"Race driver model,","cited_arxiv_id":null,"evidence_quote":"Supplies the minimum-curvature path planner used in the RDM+OTG global-planning baseline."},{"cited_title":"Optimal trajectory generation for dynamic street scenarios in a frenet frame,","cited_arxiv_id":null,"evidence_quote":"Supplies the Frenet-frame optimal trajectory generator used as the velocity-planning half (OTG) of the comparison baseline."},{"cited_title":"Model-and acceleration-based pursuit controller for high-performance autonomous racing,","cited_arxiv_id":null,"evidence_quote":"Supplies the pursuit controller used to track the RDM+OTG trajectory in the comparison."}],"review_version":1}