{"id":"b42bed62-b6ac-4e41-a850-191c4ae77e72","arxiv_id":"2504.18439","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"The BME Formula Racing Team details the perception, SLAM, planning, and control modules of their Formula Student Driverless car, FRED-003C, which placed third in the 2023 Formula Student Germany electric-driverless category.","lead":"This paper describes the software stack used by a university team to drive a small autonomous race car around a cone-bounded track, and it reports the competition results the stack achieved. A generalist might read it to see how a student team assembles perception, mapping, planning, and control modules into a working self-driving vehicle.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (4) track-width reward is discontinuous and negative above 5 m as printed, so the described planner cannot be the one that lapped unless the equation is corrected.","rationale":"The reader's weakest assumption identifies Eq. (4) as the critical technical inconsistency, and my independent reading agrees. This is the most load-bearing concern because it sits inside the planning module that selects the actual racing line. The paper is an experience report whose central evidence is the third-place FSG finish and 41.02 s fastest lap; that external outcome is credible and should not be dismissed. However, the claim that the described stack is the stack that achieved that result requires the printed equations to be accurate. Eq. (4) is not accurate as printed: the right-side Gaussian is centered at the wrong point, creating a discontinuity and uniformly negative scores above 5 m. Without code or exact parameters, the reader cannot tell whether the equation is a copy-paste error or the actual reward function. The other equations I inspected, such as the pose motion residual in Eq. (2), can be read as a standard odometry residual if v denotes distance traveled and omega denotes half the heading increment, so they are less clearly wrong. The proper response is to keep the CONDITIONAL verdict: the external result supports the system's existence, but the paper's reproducibility and internal consistency need correction before it can serve as the intended starting point for other student teams.","tokens_in":9502,"tokens_out":8449,"duration_ms":92407,"concrete_test":"Re-evaluate Eq. (4) at x=5 and x=6 to confirm the discontinuity and the negative right-tail values. Then compare the printed equation against the actual planner source or a logged reward-table dump from the FSG 2023 runs: if the right branch is centered at x=5 in the source, the equation is a typo and the correctness concern reduces to missing reproducibility data; if the source matches Eq. (4), run the centerline planner on a recorded cone map with a track segment wider than 5 m and verify that it still yields a feasible, competitive centerline consistent with the reported 41.02 s lap.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central planning equation, Eq. (4) in Section III-C, is internally inconsistent with its stated purpose. The text says the reward promotes the optimal track width range of 3 to 5 m and uses a Gaussian as a smoothing term, but both the x<3 and x>5 branches are centered at x=3. Evaluating the right branch at x=5 gives 1.3*exp(-((5-3)/0.4)^2/2) - 0.3, which is approximately -0.3, while the middle branch gives exactly 1.0. The reward therefore has a hard discontinuity at the upper boundary of the intended optimal range. For every width above 5 m the right branch approaches -0.3, so wider track segments are never promoted. If the actual planner uses Eq. (4) as printed, the claimed smoothing behavior does not hold and the planner would aggressively penalize any track wider than 5 m. If it is a typo, the true center of the right tail is not stated, and because no code, parameter values, or reward-table logs are released, the reader cannot determine which version ran in the 41.02 s lap. Since this reward function is the core of the centerline search, the paper's description of the planning module is not reproducible as written. The external competition result supports that some version of the system worked, but it does not confirm that the printed equations describe the working system.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper describes the autonomous software stack of the BME Formula Racing Team's FRED-003C vehicle used in the Formula Student Driverless competition. The stack is organized into state estimation (a FastSLAM/Graph-SLAM hybrid and a LiDAR-inertial SLAM), perception (YOLOv5 camera detection, LiDAR clustering with intensity-based colour classification, and LiDAR-camera fusion), planning (Delaunay-triangulation-based centreline search with a reward function, smoothing, spline fitting, and velocity-profile optimization), and control (a blended Stanley/Pure Pursuit lateral controller with yaw damping and a P-controller for longitudinal actuation). The paper reports a third-place overall finish in the electric-driverless category at Formula Student Germany 2023, with an average trackdrive lap time of 41.30 s and a fastest lap of 41.02 s, and it presents simulation comparisons for the smoothing and control methods.","tokens_in":9844,"tokens_out":6659,"duration_ms":65085,"significance":"If the system description is accurate, the paper is a useful engineering reference for student autonomous-racing teams: it shows a complete, competition-validated pipeline, discusses practical tuning considerations, and compares design alternatives such as smoothing methods and lateral controllers. The paper's main strengths are its direct grounding in competition results and its honest reporting of limitations, such as the steady-state error of the longitudinal P-controller. However, several equations that are central to the described modules contain inconsistencies, notably Eq. (4), Eq. (2), and the FastSLAM resampling statement. As printed, the manuscript does not allow a reader to reproduce or verify the algorithms, even though the competition result provides credible evidence that a version of the system worked. These issues are correctable in revision.","major_comments":[{"comment":"The reward function as printed cannot implement the behaviour described in the text. Both the x<3 and x>5 branches are the same Gaussian centred at x=3, so at the upper boundary of the intended optimal range the score jumps from exactly 1 at x=5 to 1.3*exp(-((5-3)/0.4)^2/2) - 0.3 ≈ -0.3 just above 5 m, and for all larger widths the score approaches -0.3. This contradicts the statement that the optimal track-width range is 3 to 5 m and the claim that the Gaussian acts as a smoothing term. Please correct the right-hand branch (presumably to centre it at x=5) or provide the actual reward parameters; without this, the printed planning algorithm is not the one that produced the reported lap times.","section":"III-C, Eq. (4)"},{"comment":"In the pose-motion residual, the translational terms use v cos(pθ,prev + ω) and v sin(pθ,prev + ω), which imply that ω is the total heading change over the interval, but the third component subtracts 2ω from the orientation difference. No time step appears, and the factor of 2 is not explained. Please state whether ω is an angular velocity or a heading increment and define the residual consistently; as written, the graph-optimization objective is not mathematically well-defined.","section":"III-A2, Eq. (2)"},{"comment":"Table II reports that the combined controller has ITAE = 143.0 and RMS = 0.201 m, both worse than the Stanley controller with ITAE = 50.0 and RMS = 0.075 m, while the text claims that the combined controller 'leverages the strengths of both approaches' and achieves 'improved recovery characteristics.' If ITAE and RMS are standard error metrics for which lower is better, the table does not support the qualitative claim. Please define ITAE, explain why the combined controller is preferred despite its higher tracking error, or revise the claim so that the text and table are consistent.","section":"III-D1, Table II and surrounding text"},{"comment":"The paper states that 'resampling selects the highest weighing particle to output the corresponding map and pose.' Standard particle-filter resampling draws particles randomly with replacement according to their normalized weights; deterministically selecting the highest-weight particle would collapse the particle distribution and is not the FastSLAM algorithm as commonly understood. Please clarify whether the highest-weight particle is used only for output while resampling is performed in the standard way, or describe the actual resampling procedure used in the implementation.","section":"III-A1, FastSLAM description"}],"minor_comments":[{"comment":"The sentence 'For the first time in the team's history, Formula Student East and the overall third place...' is incomplete and is immediately followed by a rephrased duplicate ('In 2023 at FSG, the team also got Formula Student East and the overall third place...'). Remove the duplicate and fix the sentence.","section":"IV, Results"},{"comment":"Please define all symbols and check units in Eq. (7); as written, θ_ss = m r_traj v / C_y (1 + l_F/l_R) does not obviously produce a dimensionless angle, and C_y is described as tire stiffness on the front axle without specifying units or whether rear-axle stiffness is needed.","section":"III-D1, Eq. (7)"},{"comment":"The axes of Figure 7 are unlabeled. Add axis labels and numeric ticks consistent with Eq. (4) so that the reader can verify the plotted reward function.","section":"III-C, Figure 7"},{"comment":"It is unclear how the yaw-damping term kd,yaw(rmeas - rtraj) enters the steering law; please specify whether it is added to Eq. (10) and how rmeas is obtained from the vehicle's sensors or estimator.","section":"III-D1, Eq. (13)"},{"comment":"The text says that the combined controller is stable because the Stanley and Pure Pursuit controllers are each stable, but stability of the individual controllers does not automatically imply stability of a velocity- and curvature-dependent convex combination; please either provide a supporting argument or soften the stability claim.","section":"III-D1, stability discussion"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is a system description rather than a methodological advance; whether this fits the journal is an editorial decision. The main technical issues in Eqs. (2) and (4), the FastSLAM resampling statement, and the Table II contradiction are fixable in revision, and the competition results provide credible evidence that a version of the system worked. I would not reject on the basis of the identified issues if the authors correct them and provide the missing definitions. I also note that the paper does not release code or data, which is acceptable for this type of paper but makes the equation-level errors more consequential."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is exactly what it claims to be: a detailed write-up of a student-built autonomous racing stack that actually performed in competition. The third-place finish at FSG 2023 and the 41.02 s fastest lap are real external evidence that the system worked. The value is primarily for other student teams and for documenting an engineering path, not for advancing the science of autonomous racing. Most modules are established methods: FastSLAM, FAST LIO2, YOLOv5, Delaunay-based centerline planning, Stanley and Pure Pursuit. The reward-based centerline search and the curvature-dependent controller blending are variations of known ideas, not first-principles derivations. That is fine for a systems paper, but it should be framed that way.\n\nWhat the paper does well: the architecture is clearly explained, the sensor fusion and color-classification details are concrete, and the authors are honest about limitations (e.g., the P controller's steady-state error). The comparison of smoothing methods and the controller study are useful, even if the numbers are simulation-only. The paper is also honest about what it is: an experience report with competition results as validation.\n\nThe soft spots are real but not fatal. The most serious is Eq. (4), the track-width reward. As printed, the right tail is centered at x = 3, not x = 5. That makes the reward discontinuous at the upper boundary of the intended 3–5 m optimal range and negative for every width above 5 m. If that is the actual function, the planner would penalize wide track segments, contradicting the stated goal. If it is a typo, the true center is unknown because no code or parameters are released. Either way, the planning module is not reproducible as written. This is the core of the centerline search, so it has to be fixed. Eq. (2) also has an unexplained factor of 2 on the angular velocity; that is minor and likely a typo. Table II is odd: the combined controller has higher ITAE and max lateral error than Stanley alone yet a faster lap time. That is plausible—lower error does not always mean faster—but the authors should explain it. Missing code, parameters, and error bars on simulation statistics are typical for this kind of paper but still limit reproducibility.\n\nThe citation pattern looks fine. The paper cites the relevant prior work (Kabzan, Alvarez, FastSLAM, FAST LIO2, etc.) and does not overclaim novelty. The self-references are to the team's own BSc thesis, which is appropriate in context.\n\nWho is this for? Student teams entering FSD, and researchers who want a concrete example of a working system with modest sensors. It does not deserve a desk reject, but it does need a serious referee who can check the equations and push for more transparency. I would send it to review with a clear request to fix Eq. (4), comment on Eq. (2), and release at least the reward parameters or a simple reference implementation.\n\nSerious thinker: yes. The engineering is coherent and the competition result speaks for itself.","headline":"A credible student-engineering paper whose central planning equation appears to be misprinted; worth reviewing after corrections.","tokens_in":10425,"tokens_out":1376,"would_cite":false,"duration_ms":14541,"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":"The paper claims that a modular autonomous racing stack—particle-filter mapping with graph optimisation, LiDAR-camera cone fusion, reward-based centerline planning, and blended kinematic steering—is enough to race, as shown by a…","keywords":["autonomous racing","Formula Student Driverless","simultaneous localisation and mapping","LiDAR-camera fusion","cone detection","centerline planning","velocity profile planning","lateral control"],"falsifier":"Evaluate Equation (4) at $x=5$ from both sides: the middle branch gives $f(5)=1$, while the right branch gives $1.3\\exp(-(2/0.4)^2/2)-0.3\\approx 0.30$, so the printed reward jumps at the boundary it is supposed to smooth, and at $x=5.5$ it is about $-0.29$ even though the text describes 3–5 m as optimal. Plotting or instrumenting the planner's actual reward output on track segments wider than 5 m would settle whether the formula is a typo and whether the smoothing property really holds.","tokens_in":9333,"feed_emoji":"🏁","tokens_out":13312,"duration_ms":115397,"temperature":0.7,"pith_summary":"The paper presents the autonomous software stack of a Formula Student race car and argues that the stack is enough to race competitively at the 2023 Formula Student Germany electric-driverless event, where it placed third with a fastest trackdrive lap of 41.02 s. The authors' case is that a modular pipeline—particle-filter mapping with graph optimisation, LiDAR-camera cone detection and colour classification, reward-based centerline planning, and a lightweight blended lateral controller—can deliver this performance without dynamic model-based control or heavy optimisation solvers. The paper backs the claim with lap-time and path-tracking data, a comparison of trajectory smoothing methods, and simulated controller comparisons. If the central claim is right, this is a practical template for student teams and a plausible starting point for full-scale autonomous racing development.","feed_headline":"41.02-second lap earns third place at Formula Student","feed_subtitle":"Modular SLAM, cone fusion, reward planning, and blended steering were enough to race on an unseen track.","key_machinery":"The central mechanism is the centreline search over midpoints of a Delaunay triangulation of the detected cones. At each step the planner scores candidate midpoints with a weighted sum of reward factors—track width, cone colour, angle change, spacing, and a predictive-model term—and greedily selects the highest-scoring midpoint; the track-width reward is designed to keep the trajectory inside the 3–5 m band with a Gaussian transition instead of a hard rule. A secondary mechanism is the curvature-dependent steering blend: as path curvature grows, more weight shifts toward the geometric look-ahead law, and as the corner exits, weight shifts back to the kinematic cross-track-error law, which the paper reports reduces overshoot in tight corners. The trajectory smoothing (moving average plus simplification) and $C^2$ cubic-spline fitting are what turn the discrete midpoint chain into a reference with low curvature variation.","core_discovery":"On the paper's own terms, the central discovery is that a deliberately modular, low-complexity stack can reach race performance. The state estimator combines FastSLAM-style particle filtering with pose-graph optimisation over landmark observation and pose motion residuals; the perception system detects track-bounding cones from LiDAR clusters and uses camera bounding boxes for colour classification and false-detection rejection; the planner searches the midpoints of a Delaunay triangulation of the cone map with a weighted reward function and then smooths, spline-fits, and velocity-profiles the resulting path; and the controller blends two kinematic steering laws with a curvature-dependent weight plus a yaw-damping term. The paper reports that this system gave the team its first podium at Formula Student Germany, with an average trackdrive lap of 41.30 s and a fastest lap of 41.02 s.","pith_inferences":["An editorial reading of the smoothing table is that the combined smoother's lap-time gain comes almost entirely from suppressing curvature variation (CVR drops from order $10^{-1}$ to $10^{-3}$), which suggests the effective objective the planner is optimising may be curvature smoothness rather than raw path shortness.","A reader wanting to reuse the stack should treat the printed Equation (4) as provisional: if the right-tail center is a typo for $x=5$, the reward is symmetric and the described behavior follows; if not, the planner needs the 5 m boundary case audited before relying on wide-track sections.","The paper's lateral-controller comparison is reported only in simulation, so the on-track 41.02 s lap cannot by itself isolate the contribution of the combined controller; an on-vehicle ablation of the three steering laws on the same track would separate those effects."],"forward_implications":["A complete, competitive driverless stack can be assembled from accessible methods—particle-filter SLAM, geometric controllers, and rule-based reward functions—without model predictive control or commercial optimisation solvers.","LiDAR-only perception with camera-supplemented colour classification is robust enough for competition, because the fusion step falls back to LiDAR whenever a projected cone lies outside the camera's field of view.","The reward-based centerline planner can build a drivable trajectory online from an unseen cone layout, and the smoothing-and-spline step can keep curvature variation low enough for high-speed lap times.","The velocity profile generated under dynamic constraints can be tracked with a simple longitudinal P controller and a blended steering law, consistent with the reported 41.02 s best lap.","The same architecture can serve as a springboard to full-scale autonomous racing, since the development path described in the paper continued beyond the Formula Student car."],"supporting_citations":[{"why":"Supplies the complete baseline driverless racing stack that defines the architecture this paper presents.","marker":"[2]"},{"why":"Supplies the particle-filter/EKF landmark-based SLAM method used as the state-estimation front-end.","marker":"[7]"},{"why":"Supplies the improved proposal-distribution variant of that SLAM method, adopted in the implementation.","marker":"[8]"},{"why":"Supplies the LiDAR-inertial odometry framework the LiDAR SLAM is adapted from.","marker":"[10]"},{"why":"Supplies the real-time object detection network used for camera-based cone detection.","marker":"[12]"},{"why":"Supplies the cone-image dataset used to train the camera detector.","marker":"[13]"},{"why":"Supplies the slope-robust ground segmentation method used in LiDAR preprocessing.","marker":"[15]"},{"why":"Supplies the optimal-complexity velocity planning algorithm underlying the forward-backward solver.","marker":"[17]"},{"why":"Supplies the kinematic steering law, steady-state yaw compensation, and stability guarantee used by the lateral controller.","marker":"[18]"},{"why":"Supplies the idea of blending two path-tracking steering laws, which the combined controller implements.","marker":"[19]"}],"fun_headline_variants":["Modular stack wins Formula Student podium","41.02s lap from a student-built autonomous stack","Low-complexity autonomy beats unseen tracks","FastSLAM and reward planning race to third place"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the printed track-width reward function in Equation (4) matches the planner's real implementation, because as printed its right-hand Gaussian is centered at $x=3$ rather than $x=5$, making the reward discontinuous at 5 m and negative above about 5.9 m—which would contradict the paper's description of smooth promotion and penalisation around the 3–5 m optimal range.","fun_headline_variants_meta":{"raw":{"variants":["Modular stack wins Formula Student podium","41.02s lap from a student-built autonomous stack","Low-complexity autonomy beats unseen tracks","FastSLAM and reward planning race to third place"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0002,"raw_usage":{"total_tokens":1337,"prompt_tokens":868,"completion_tokens":469,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":484,"completion_tokens_details":{"reasoning_tokens":410}},"tokens_in":484,"tokens_out":469,"duration_ms":4611,"temperature":1.0,"reasoning_tokens":410,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:15:51.124210+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Evaluate Equation (4) at $x=5$ from both sides: the middle branch gives $f(5)=1$, while the right branch gives $1.3\\exp(-(2/0.4)^2/2)-0.3\\approx 0.30$, so the printed reward jumps at the boundary it is supposed to smooth, and at $x=5.5$ it is about $-0.29$ even though the text describes 3–5 m as optimal. Plotting or instrumenting the planner's actual reward output on track segments wider than 5 m would settle whether the formula is a typo and whether the smoothing property really holds.","supporting_citations":[{"cited_title":"AMZ Driverless: The Full Autonomous Racing System","cited_arxiv_id":"1905.05150","evidence_quote":"Supplies the complete baseline driverless racing stack that defines the architecture this paper presents."},{"cited_title":"Fastslam: a factored solution to the simultaneous localization and mapping problem,","cited_arxiv_id":null,"evidence_quote":"Supplies the particle-filter/EKF landmark-based SLAM method used as the state-estimation front-end."},{"cited_title":"Fastslam 2.0: An improved particle filtering algorithm for simultaneous localization and mapping that provably converges,","cited_arxiv_id":null,"evidence_quote":"Supplies the improved proposal-distribution variant of that SLAM method, adopted in the implementation."},{"cited_title":"Fast-lio2: Fast direct lidar-inertial odometry,","cited_arxiv_id":null,"evidence_quote":"Supplies the LiDAR-inertial odometry framework the LiDAR SLAM is adapted from."},{"cited_title":"ultralytics/yolov5: v7.0 - yolov5 sota realtime instance segmentation,","cited_arxiv_id":null,"evidence_quote":"Supplies the real-time object detection network used for camera-based cone detection."},{"cited_title":"Fsoco: The formula student objects in context dataset,","cited_arxiv_id":null,"evidence_quote":"Supplies the cone-image dataset used to train the camera detector."},{"cited_title":"A slope-robust cascaded ground segmentation in 3d point cloud for autonomous vehicles,","cited_arxiv_id":null,"evidence_quote":"Supplies the slope-robust ground segmentation method used in LiDAR preprocessing."},{"cited_title":"An optimal complexity algorithm for minimum-time velocity planning,","cited_arxiv_id":null,"evidence_quote":"Supplies the optimal-complexity velocity planning algorithm underlying the forward-backward solver."},{"cited_title":"Path following controller for au- tonomous vehicles,","cited_arxiv_id":null,"evidence_quote":"Supplies the idea of blending two path-tracking steering laws, which the combined controller implements."}],"review_version":1}