{"id":"1b901d1c-c344-4fb5-a6f3-3eb041f779b1","arxiv_id":"2505.05157","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An online forward-backward velocity planner with spatial-domain sampling lets an autonomous race car adapt to reduced grip in simulation, improving section times by up to 1.42 s.","lead":"This paper presents a method for computing a race car's speed profile online, reacting to grip changes such as tire temperature or rubber buildup, and feeds it to a local trajectory planner that samples in space rather than time. The approach is tested in simulation on the Yas Marina circuit and shows second-level time gains over a fixed offline speed profile.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The 1.42 s gain is guaranteed only within the point-mass gg-diagram model; transient effects, acknowledged in §IV-C, could invalidate the feasibility claim.","rationale":"The reader's weakest-assumption analysis and my own converge on the same point: the point-mass quasi-steady-state model, with feasibility defined by gg-diagrams, omits transient vehicle dynamics. The paper itself flags this in Section IV-C, so the concern is acknowledged rather than hidden. The central claim of a 1.42 s improvement while 'remaining feasible' is meaningful only if that simplified model captures the vehicle's true capability limits. Because the simulation imposes perfect tracking of trajectories that already satisfy the planner's gg-diagram check, the result is self-consistent within the model but says nothing about whether the same trajectories would be feasible or stable for the real vehicle. This is a load-bearing concern for the paper's real-world applicability, though not for its internal mathematical correctness. Other issues, such as missing code/data, undisclosed cost weights, and a single deterministic run, are secondary and would mainly affect reproducibility. The apparent conflict between the abstract's multi-vehicle robustness claim and the Section IV-B limitation that feasibility is only guaranteed on the race line is a scope issue that does not directly undermine the central 1.42 s result, which is a single-vehicle, race-line-following scenario. A high-fidelity dynamic simulation test would settle the transient-dynamics concern: if the online velocity profile still remains feasible and the time gain persists, the claim is robust; if not, the gain is a point-mass artifact. Given that the authors explicitly acknowledge the limitation and position the work as a foundation, conditional acceptance is the appropriate verdict, and my assessment does not change the reader's conclusion.","tokens_in":10503,"tokens_out":4725,"duration_ms":49103,"concrete_test":"Replace the perfect-tracking, point-mass-consistent simulation in Section III-B with a dynamic single-track (bicycle) model that includes yaw inertia and tire relaxation, using the same online velocity profile and spatial sampling planner. Run the 2000–2600 m sector with α = 0.7 and check whether the executed trajectory violates the gg-diagram limits (e.g., the combined acceleration leaves the diamond) or exhibits unstable states (sideslip growth) at the abrupt acceleration transitions near s = 2520 m. If violations appear or the sector time degrades materially, the 'remains feasible' claim is point-mass-only and the headline improvement is not physically grounded.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central quantitative claim (Section III-B, Table II) is that the online velocity profile yields a 1.42 s faster sector because the sampling planner 'remains feasible' while tracking it. Feasibility is defined exclusively by the point-mass quasi-steady-state gg-diagram (Section II-B, Eq. 4). The simulation imposes perfect trajectory tracking (Section III), so any trajectory passing the planner's gg-diagram check is guaranteed to appear feasible in simulation: the result is internally consistent by construction. The paper's own Discussion IV-C concedes that transient effects are not integrated and 'might lead to unstable driving states while remaining feasible in the point-mass formulation.' The load-bearing assumption is therefore that the point-mass gg-diagram sufficiently represents the vehicle's actual acceleration limits at racing speed. If transients (tire force buildup, load transfer, actuator lag) cause the true combined acceleration to exceed the gg-diagram even when the point-mass profile lies inside it, the 1.42 s improvement is an artifact of the simplified model, and the 'feasibility' claim does not transfer to the physical racing environment the paper targets. This is not an internal inconsistency, but it is the weakest link in the central claim's real-world validity.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an online velocity profile generation method for autonomous racing on three-dimensional tracks, combining a forward-backward (FW-BW) solver over a fixed race line with an apex-detection procedure, and integrates the resulting velocity profile into a sampling-based local planner through a spatial-domain trajectory sampling strategy. The central claims are that the online profile adapts to changing grip conditions (modeled by a scaling factor alpha) and that, when used by the sampling planner, it enables feasible tracking and improves sector time by 1.42 s compared with an offline profile (Section III-B, Table II). Additional experiments address lateral deviation from the race line and runtime behavior.","tokens_in":10780,"tokens_out":4466,"duration_ms":50592,"significance":"If the claims hold, the contribution is practically relevant: real-time adaptation of reference velocity profiles to grip changes is a recognized gap in autonomous racing, and the spatial sampling extension is a reasonable way to align local trajectories with brake points and apexes. The FW-BW formulation in Section II-D is standard and correctly applied, the spatial transformation in Eq. (11) is mathematically sound, and the paper is unusually candid about its limitations, especially in Section IV-C. The deterministic simulation setup and clear parameter tables support reproducibility of the reported scenarios. However, the central quantitative claim rests on a single simulation run within a point-mass quasi-steady-state model, and the real-time claim is weakened by the reported runtime exceeding the planning interval.","major_comments":[{"comment":"The average runtime of the entire planning step is reported as 114 ms, while Table I fixes the simulation time per planning step at 100 ms. This means the average computation time exceeds the planning period, so the statement that the approach demonstrates 'feasibility for online application' and 'real-time capable calculation' is not directly supported. Please report the runtime distribution (maximum, percentile), clarify how a 114 ms average is consistent with a 100 ms planning step (e.g., asynchronous execution, lookahead, or budget-free simulation), or revise the real-time claim accordingly.","section":"Section III-D and Table I"},{"comment":"Feasibility throughout the paper is defined by the point-mass quasi-steady-state gg-diagram in Eq. (4), and the central comparison in Section III-B reports that the sampling planner with the online profile 'remain[s] feasible.' The paper's own Section IV-C acknowledges that transient effects 'are not integrated into the online velocity profile, which might lead to unstable driving states while remaining feasible in the point-mass formulation.' Consequently, the 1.42 s improvement and the feasibility statement are not evidence that the profile is safe or feasible for the physical vehicle; they are only evidence of feasibility within the simplified model. The authors should either validate the claimed improvement with a higher-fidelity vehicle dynamics model (or a transient-feasibility check) or explicitly and consistently scope the central claim to the point-mass model.","section":"Section II-B, Eq. (4) and Section IV-C"},{"comment":"The main quantitative result, the 1.42 s sector-time improvement, is based on a single deterministic simulation scenario with one grip-scaling value (alpha = 0.7) and one initial condition. No sensitivity analysis is provided with respect to alpha, velocity optimization horizon, cost weights, start state, or track section. Without such variation, the claim of 'robust performance' and the generality of the 1.42 s improvement are not established. Please add parameter sweeps or at least a few additional scenarios (e.g., different alpha values, different initial velocity offsets, and a different track section) to support the claimed robustness.","section":"Section III-B, Table II and Section III-C"}],"minor_comments":[{"comment":"The notation in Eq. (11) is confusing because 's' denotes both the arc-length coordinate in the spatial domain and the progress coordinate of the reference line; for instance, the left-hand side 'ddot s(s)' mixes the two roles. A clearer notation, such as using sigma or q for the spatial coordinate, would improve readability.","section":"Eq. (11)"},{"comment":"In Eq. (6), the symbol hat a_y is used for the admissible lateral acceleration, but in the body text and Eq. (13) it is also used for the actual lateral acceleration. This conflation makes the fixed-point iteration harder to follow. Please define hat a_y explicitly as the lateral acceleration limit implied by the gg-diagram for the current velocity.","section":"Eq. (6) and Appendix"},{"comment":"The abstract and contributions list describe the approach as suitable for 'competitive single- and multi-vehicle scenarios,' but Section IV-B states that the velocity profile is guaranteed feasible only on the race line and that the approach is 'only partially suitable for multi-vehicle scenarios.' Please align the abstract and contribution statements with this limitation.","section":"Abstract and Section I-C"},{"comment":"Minor editorial issues: 'Subotits et al.' appears twice in Section I-B and should be 'Subosits et al.,' and the reference to Lipp and Boyd is missing venue and year information. These should be corrected.","section":"References and typos"}],"recommendation":"major_revision","confidential_remarks":"The derivation and implementation appear sound, and the limitations section is unusually honest. The main issue for the journal is that the central empirical claim is supported by a single deterministic simulation in a simplified model, and the runtime data contradict the real-time claim unless clarified. With additional sensitivity experiments and a corrected runtime assessment, the paper could become acceptable. I would not recommend rejection because the core methodology is coherent and the limitations are explicitly acknowledged by the authors."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a solid engineering paper, not a breakthrough. The new bit is combining an online FW-BW velocity profile (scaled by a grip factor α) with a spatial-domain sampling strategy so the local planner brakes and accelerates relative to apex locations rather than a fixed time horizon. The math in Section II-D is standard and correctly applied; the spatial transformation in (11) is right; the apex detection via fixed-point iteration on curvature is reasonable. Runtime of 43 ms for the profile generation is believable for online use. The citation pattern looks fine — the authors lean on their own prior work where it is directly the baseline ([6], [24]), and that is legitimate.\n\nThe paper is honest: Section IV-C admits transient effects are not integrated and could lead to unstable states, and Section IV-B concedes the feasibility guarantee holds only on the race line, which is a big caveat for the multi-vehicle claims in the abstract.\n\nThe main soft spot is the quantitative evidence. The 1.42 s and 1.18 s improvements come from single deterministic simulation runs with perfect tracking and exactly known opponent positions, so there are no error bars or sensitivity analysis. The cost weights in (12) are described as 'tuned' but not listed, which makes the comparison hard to reproduce. No code or data are released. These are fixable but they matter if anyone wants to rely on the numbers.\n\nThe stress-test note about transients is on point but doesn't sink the paper. The feasibility check is defined by the point-mass gg-diagram, and since the simulation imposes perfect tracking, the experiment is internally consistent by construction. The paper does not claim transient feasibility — it explicitly flags it as future work. So the central claim holds within the stated model. The concern is about external validity, not internal validity.\n\nI'd send this to review. It's a worthy conference/workshop paper, the limitations are acknowledged, and the spatial sampling idea is worth seeing in print even if the evaluation is a bit thin. The authors should be pushed to release code/data and run multiple seeds or perturb α.","headline":"A credible integration of online grip-adaptive velocity profiles with spatial sampling for local racing planners; the 1.42 s gain is real within the point-mass model, but the paper's own limitations section narrows its scope.","tokens_in":11279,"tokens_out":3007,"would_cite":false,"duration_ms":25546,"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":"An online velocity profile keeps a sampling race planner feasible when grip drops, saving 1.42 s per section.","keywords":["autonomous racing","velocity profile generation","sampling-based motion planning","spatial trajectory sampling","gg-diagram","grip adaptation","forward-backward solver","apex detection"],"falsifier":"Run the same planner with the online profile on a track with measured grip reduced to α = 0.7 in a high-fidelity vehicle simulation or on a real car; if the executed trajectory exceeds the vehicle's actual tire limits, loses grip, or goes unstable in a section where the gg-diagram check passed, the point-mass feasibility claim is refuted.","tokens_in":10312,"feed_emoji":"🏎️","tokens_out":6037,"duration_ms":58708,"temperature":0.7,"pith_summary":"The authors aim to establish that a local sampling-based planning algorithm for autonomous race cars should not rely on a once-computed offline speed profile, because track grip changes. They propose recomputing the speed profile online along the same offline race line, scaling the vehicle's acceleration limits by a grip factor, and passing that updated profile to the local planner. The key comparison shows that when grip is reduced, the planner using the online profile brakes later and accelerates earlier through a chicane, remaining dynamically feasible and completing the section 1.42 s faster than the same planner using the offline, grip-blind profile. The paper also claims that sampling trajectories in the spatial domain, anchored to apexes, tracks the race-line profile better than temporal sampling when the vehicle starts off the profile.","feed_headline":"Grip-aware online velocity profile wins a section by 1.42 s","feed_subtitle":"A sampling planner that recomputes its speed profile stays feasible when grip drops; spatial sampling fixes braking points.","key_machinery":"The load-bearing machinery is the pair of the forward-backward solver and the gg-diagram feasibility model, together with a spatial sampling overlay. A gg-diagram is a speed-dependent diamond of allowed longitudinal and lateral accelerations, scaled by a grip factor α; the solver integrates the fastest acceleration forward from the start and the maximum deceleration backward from each detected apex, and the feasible profile is the pointwise minimum. Apexes are located not simply by maximum curvature but by fixed-point iteration on admissible velocity along search intervals. The local planner then generates candidate trajectories over a fixed arc length S instead of a fixed time T, so brake points and apexes line up with the track geometry instead of drifting with time.","core_discovery":"At the paper's core is the claim that a velocity profile fixed at planning time is the wrong reference for a sampling-based race planner under changing track conditions. The authors recompute the feasible speed profile along the offline race line in real time from a three-dimensional track model, using a forward-backward solver constrained by velocity-dependent gg-diagrams that are shrunk by a grip scaling factor α. When the grip is reduced to α = 0.7 on a test section, the planner using this online profile follows the race-line velocity and path while staying feasible, whereas the planner using the offline profile brakes early, deviates from the path, and loses 1.42 s over the section. In addition, when the car starts below the reference speed or off the race line, trajectories generated in the spatial domain place braking and apex points correctly, while temporal trajectories decelerate too early and accelerate before the apex.","pith_inferences":["This implies a testable full-lap extension: feed a grip map rather than two constant α values into the same profile generator, and the method should show smaller but still positive lap-time gains anywhere the offline profile is out of date.","Because the profile is only reoptimized on the fixed race line, the method's own logic predicts its advantage mostly in single-vehicle or follow-the-line situations; in close multi-vehicle racing where the path shifts laterally, the speed profile would need to be updated on the actual path to keep the guarantee.","The observed abrupt acceleration switches point to a concrete add-on: a jerk penalty or low-pass smoothing on the forward-backward output would trade some lap time for actuator feasibility, a trade the current experiments do not yet measure.","The spatial sampling idea, being a parameterization change, could be applied to any Frenet-frame sampling planner, not just the racing one used here."],"forward_implications":["When grip changes during a lap, a planning stack that recomputes the speed profile on the race line can keep the local planner inside its acceleration limits without switching to a more conservative global plan.","Spatial-domain trajectory sampling is better suited than temporal sampling for tracking a speed profile from a starting state that is off the profile, because it aligns braking and apex points with track geometry.","The speed profile generation is cheap enough for real-time use, with an average runtime of about 43 ms, and its cost grows linearly with the number of apexes in the horizon.","The method is expected to transfer to other tracks and vehicles as long as a fixed race line and a valid gg-diagram model are available."],"supporting_citations":[{"why":"Supplies the offline three-dimensional race line and the road-frame angular velocity relation that the online profile is computed along.","marker":"[6]"},{"why":"Provides the 3D track representation, including the parametrized spine and road frame, used to describe the drivable surface.","marker":"[7]"},{"why":"Supplies the velocity-dependent gg-diagram extended by apparent gravity that defines the acceleration feasibility set.","marker":"[5]"},{"why":"The underlying sampling-based local planner whose temporal trajectory generation is extended to the spatial domain.","marker":"[24]"},{"why":"Basis for the relative jerk-optimal trajectory generation in a Frenet frame that the sampling step adapts.","marker":"[21]"},{"why":"An offline forward-backward boundary-speed method with critical-point detection that motivates the solver structure and its extension to transient constraints.","marker":"[19]"},{"why":"Establishes the forward-backward solver technique for minimum-time velocity profiles on a fixed path.","marker":"[9]"}],"fun_headline_variants":["Online speed profile beats offline by 1.42 s in grip drop","Grip-aware online velocity wins 1.42 s over offline","Online velocity planning beats offline by 1.42 s under grip loss","Adaptive speed profile gains 1.42 s vs offline on grip drop"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole feasibility argument rests on a point-mass quasi-steady-state model that ignores transient load transfer and actuator dynamics, so a profile that passes the gg-diagram check can still be unsafe in the real car; the paper says so itself.","fun_headline_variants_meta":{"raw":{"variants":["Online speed profile beats offline by 1.42 s in grip drop","Grip-aware online velocity wins 1.42 s over offline","Online velocity planning beats offline by 1.42 s under grip loss","Adaptive speed profile gains 1.42 s vs offline on grip drop"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001736,"raw_usage":{"total_tokens":6798,"prompt_tokens":817,"completion_tokens":5981,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":433,"completion_tokens_details":{"reasoning_tokens":5902}},"tokens_in":433,"tokens_out":5981,"duration_ms":39387,"temperature":1.0,"reasoning_tokens":5902,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T23:10:50.355642+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same planner with the online profile on a track with measured grip reduced to α = 0.7 in a high-fidelity vehicle simulation or on a real car; if the executed trajectory exceeds the vehicle's actual tire limits, loses grip, or goes unstable in a section where the gg-diagram check passed, the point-mass feasibility claim is refuted.","supporting_citations":[{"cited_title":"On- line Time-Optimal Trajectory Planning on Three-Dimensional Race Tracks,","cited_arxiv_id":null,"evidence_quote":"Supplies the offline three-dimensional race line and the road-frame angular velocity relation that the online profile is computed along."},{"cited_title":"Optimal Control of a Formula One Car on a Three-Dimensional Track—Part 1: Track Modeling and Identification,","cited_arxiv_id":null,"evidence_quote":"Provides the 3D track representation, including the parametrized spine and road frame, used to describe the drivable surface."},{"cited_title":"A three-dimensional free-trajectory quasi- steady-state optimal-control method for minimum-lap-time of race vehicles,","cited_arxiv_id":null,"evidence_quote":"Supplies the velocity-dependent gg-diagram extended by apparent gravity that defines the acceleration feasibility set."},{"cited_title":"Sampling- Based Motion Planning with Online Racing Line Generation for Autonomous Driving on Three-Dimensional Race Tracks,","cited_arxiv_id":null,"evidence_quote":"The underlying sampling-based local planner whose temporal trajectory generation is extended to the spatial domain."},{"cited_title":"Optimal trajectory generation for dynamic street scenarios in a Frenet Frame,","cited_arxiv_id":null,"evidence_quote":"Basis for the relative jerk-optimal trajectory generation in a Frenet frame that the sampling step adapts."},{"cited_title":"Development of vehicle dynamics tools for motorsports,","cited_arxiv_id":null,"evidence_quote":"An offline forward-backward boundary-speed method with critical-point detection that motivates the solver structure and its extension to transient constraints."},{"cited_title":"Minimum curvature trajectory planning and control for an autonomous race car,","cited_arxiv_id":null,"evidence_quote":"Establishes the forward-backward solver technique for minimum-time velocity profiles on a fixed path."}],"review_version":1}