{"id":"d3aa7da5-1af9-49e7-b953-5241b58b742f","arxiv_id":"2504.14747","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":8,"one_line_summary":"An autonomous driving planner that blends a dynamic risk field, Frenet-frame quintic trajectories, and adaptive particle swarm optimization is demonstrated only in two hand-picked simulations.","lead":"This paper combines artificial potential fields, Frenet coordinates, and an improved particle swarm optimizer to plan lane changes for an autonomous car on curved roads. The authors report two simulation cases but no real-world experiments, despite claiming them.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Frenet dynamic transformation Eq. (6) is not the standard Frenet-Serret expression; a constant-speed circular path yields a spurious tangential acceleration, so trajectories using this transformation are not dynamically valid.","rationale":"The reader's weakest assumption is on target: the Frenet transformation is the load-bearing technical link, and a mathematical error there invalidates all downstream trajectory claims. I independently re-derived the standard Frenet-Serret equations and found concrete discrepancies in Eq. (5) and Eq. (6). The validation gaps the reader notes (no real-world experiment, only two hand-picked simulations, no error bars in Fig. 6, undeclared PSO decision variables) are real and independently support REJECT, but they are primarily evidentiary or reproducibility problems. The Frenet transformation is a correctness problem: if the implementation follows the printed equations, the resulting trajectories are not dynamically feasible even before considering the risk field or optimizer. Therefore the reader's REJECT verdict stands unchanged.","tokens_in":8363,"tokens_out":8843,"duration_ms":81529,"concrete_test":"Use the paper's 64-m-radius circular reference path. Construct a ground-truth trajectory with constant speed v=10 m/s and constant lateral offset d=1 m, so the correct Frenet states are sdot = v/(1-kappa*d), ddot = 0, sddot = 0, ddot_ddot = 0. Compute global positions, velocities, and accelerations from standard circular kinematics, then feed them into Eq. (6). If the returned sddot or ddot_ddot is nonzero (the printed formulas give sddot = kappa*sdot^2/(1-kappa*d) > 0), the transformation is invalid. As a second check, compute d from Eq. (5) for the same position under forward and reversed velocity; if the sign changes, the lateral offset is not geometrically defined.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section II-D's Frenet transformation is mathematically unsound. Eq. (5) defines d = sign(xdot*yddot - xddot*ydot) * distance-to-nearest-point. That sign is a kinematic proxy (related to angular acceleration), not the geometric side of the reference path; a stationary vehicle would get d=0 for any off-path position, and reversing velocity at the same position flips the lane side. More importantly, Eq. (6) is not the standard Frenet-Serret dynamic transformation. Starting from r = p(s) + d n(s) with e_t=(cos psi, sin psi) and n=(-sin psi, cos psi), the standard derivation gives sddot = [a_t + kappa_s sdot^2 d + 2*kappa*sdot*ddot]/(1-kappa*d) and ddot_ddot = a_n - kappa*sdot^2*(1-kappa*d), where a_t = xddot cos psi + yddot sin psi and a_n = -xddot sin psi + yddot cos psi. The printed Eq. (6) instead adds kappa*sdot^2 to the sddot numerator and subtracts only kappa*sdot^2 from the ddot_ddot term, omitting the 1-kappa*d factor and all kappa_s and 2*kappa*sdot*ddot coupling terms. Consequently, on the paper's own 64-m-radius road, a vehicle traveling at constant speed with d=0 would be reported as having nonzero longitudinal and lateral accelerations. Since the quintic trajectory generator and IPSO cost function are never shown to operate independently of this transformation, the dynamic feasibility claim in Eq. (19) is not supported.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes an integrated trajectory planning framework for autonomous lane changes on curved two-lane roads, combining an adaptive risk field based on artificial potential fields, a Frenet-frame quintic polynomial trajectory generator, and an improved particle swarm optimizer (IPSO). The paper claims that the framework maintains safety margins and comfort while converging faster than PSO, GA, and Active Set methods, and states in the abstract that it is validated through extensive simulations and real-world scenarios. Section V, however, reports only two hand-picked Matlab test cases plus one qualitative wider-road case and a computation-time bar chart, with no real-world experiments.","tokens_in":8840,"tokens_out":7543,"duration_ms":63665,"significance":"If the framework worked as described, it would be a useful engineering contribution to interactive lane-change planning on curved roads, particularly the idea of a dynamic risk field that adjusts lane-change intention based on surrounding vehicle behavior. The paper's clear decomposition of the pipeline into risk field, Frenet formulation, quintic generation, and IPSO is a strength. However, the reported results are not reproducible without numerous omitted parameter values and undefined terms, and the central dynamic transformation appears to be mathematically incorrect, so the claimed validation is not currently supported. The paper provides no code or machine-checkable proofs to offset these gaps.","major_comments":[{"comment":"The dynamic Frenet coordinate transformation in Eq. (6) is mathematically incorrect. The standard Frenet kinematics with r = p(s) + d n(s) gives sddot = (a_t + 2 κ sdot ddot)/(1 - κ d) and dddot = a_n - κ sdot^2 (1 - κ d), where a_t = xddot cosψ + yddot sinψ and a_n = yddot cosψ - xddot sinψ. Equation (6) instead adds κ sdot^2 to the sddot numerator and subtracts only κ sdot^2 from the dddot term, omitting the 2 κ sdot ddot coupling and the (1 - κ d) factor. For uniform circular motion at constant speed with d = 0, a_t = 0 but the printed formulas yield nonzero sddot and dddot. Since the feasibility constraints in Eq. (19) and the cost function in Eq. (24) are evaluated in this frame, the claim of dynamically feasible trajectories is not supported. Additionally, the lateral-offset sign in Eq. (5) uses sign(xdot yddot - xddot ydot), which is not the geometric side of the reference path; it vanishes for a stationary vehicle and flips when the velocity direction reverses at the same position.","section":"Section II-D, Eq. (6)"},{"comment":"The abstract states that the framework is 'validated through extensive simulations and real-world scenarios', but Section V contains only two hand-picked Matlab test cases, one qualitative wider-road example with no quantitative measures, and a single computation-time bar chart. No real-world experiments, sensor data, statistical analysis, or comparison of trajectory safety and comfort against baseline planners are reported. The load-bearing validation claim in the abstract is therefore unsupported by the evidence in the manuscript.","section":"Section V and Abstract"},{"comment":"The model depends on many parameters that are never specified: α, β, γ0, σ0, λ, ξ, vsafe, u_b,threshold, u_c,threshold, the cost weights w1..w5, and the IPSO bounds and learning-factor endpoints. Moreover, the lane-change feasibility function Φ(Ξ) introduced in Eq. (14) is never defined. Because the lane-change trigger conditions in Eq. (27) rely on thresholds that are not given, the reported lane-change times and safety margins are consequences of undisclosed tunings rather than independent predictions, making the validation circular and the paper unreproducible.","section":"Sections III-IV, Eq. (27) and Eq. (14)"},{"comment":"The claim that IPSO converges faster than PSO, GA, and Active Set is not substantiated. Fig. 6 reports a single bar per algorithm with no variance or confidence intervals, the label 'Case 3' does not correspond to either of the two test cases listed in the text, and no convergence criterion or tolerance is stated. The reported difference between IPSO (0.004 s) and PSO (0.005 s) is likely within run-to-run noise, and no statistical test is provided. There is also no end-to-end comparison of the planner's safety or comfort performance against baseline planners, so the central planning performance claim is not evaluated.","section":"Section V, Fig. 6"},{"comment":"The dynamic feasibility constraints in Eq. (19) include limits on lateral acceleration, yaw rate, side-slip angle, and steering angle, but the manuscript provides no vehicle model or mapping between the quintic polynomial trajectory and the quantities δ, β, and ψdot. Without such a mapping, the penalty-based constraint handling in Section IV-D cannot demonstrate that the generated trajectories satisfy Eq. (19), and the 'dynamically feasible' characterization in the abstract is not justified.","section":"Section II-E, Eq. (19)"}],"minor_comments":[{"comment":"There is a typo 'iniital' for 'initial', and 'radium' should be 'radius'; the symbol 'S2' used in the analysis of Case 1 is never defined.","section":"Section V, first paragraph"},{"comment":"The text says 'Figs. 1(a) through (d) show the smooth curves of velocity, acceleration, steering angle, and yaw rate' and 'Fig. 2(a) illustrates the trajectories', but the figure captions show the opposite assignment: Fig. 1 contains trajectories and field values, while Fig. 2 contains the smooth curves. The figure references appear to be swapped.","section":"Section V, text around Figs. 1-2"},{"comment":"The weighting matrix W in the norm term ∥Xego - Xref∥^2_W is not defined, and the symbols α and β are reused with different meanings in Eq. (12), which is confusing.","section":"Section III-A, Eq. (9)"},{"comment":"The term x_te,max in J1 is never defined, and the objective of maximizing terminal position for a lane-change maneuver is ambiguous; it should be stated whether this encodes progress, target-lane completion, or something else.","section":"Section IV-C, Eq. (24)"},{"comment":"The steering-angle limit δmax = 2° is extremely restrictive for a lane change on a road with a 64-m inner radius, and no source or justification is provided for this constraint.","section":"Section II-E, Eq. (19)"},{"comment":"The time points t=10,20,30,40,50 in Fig. 5 appear to be simulation steps, but the text reports 'lane-changing time is 18' without stating whether this is seconds or time steps; the simulation timestep is never defined.","section":"Section V, Fig. 5 and text"},{"comment":"The caption refers to 'Case 3', which is not among the two test cases defined earlier in the section, and the comparison lacks a dedicated GA citation (Ref. [23] is a trajectory-planning paper, not a genetic-algorithm reference).","section":"Section V, Fig. 6"},{"comment":"Lower-case u_a, u_b, and u_c are used in the trigger conditions without formal definition; the text should state explicitly that these denote the field values U_A, U_B, and U_C evaluated at the ego state.","section":"Section IV-E, Eq. (27)"}],"recommendation":"reject","confidential_remarks":"The abstract's claim of real-world validation is not supported by the manuscript's content, which is a serious discrepancy. More importantly, the incorrect Frenet transformation in Eq. (6) invalidates the reported trajectory results; any future submission would need to re-derive the transformation, re-run all experiments, disclose all parameters, and provide the missing definitions, including Φ(Ξ). Given the scope of these corrections, I recommend rejection rather than major revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"First thing to know: the stress-test note is correct. Eq. (6) is not the standard Frenet dynamic transformation, and it is load-bearing. Second thing: the abstract promises “extensive simulations and real-world scenarios,” but the full text contains only two hand-picked Matlab cases. That gap alone would force a major revision.\n\nOn the positive side, the paper is clearly organized and the architecture is a sensible recombination: risk field for lane-change intent, Frenet coordinates for road alignment, quintic polynomials for smoothing, and IPSO for cost optimization. The two simulation cases illustrate the intended behavior, and the paper is readable. The PSO/GA/Active Set comparison is thin but at least reports absolute compute times, and the adaptive inertia weight is a standard but legitimate improvement.\n\nThe load-bearing problem is Section II-D. Starting from r = p(s) + d n(s), the correct expressions are sddot = (a_t + 2κ sdot ddot)/(1 − κ d) and dddot = a_n − κ(1 − κ d) sdot^2. The paper’s Eq. (6) omits the 2κ sdot ddot coupling term and adds κ sdot^2 in the sddot numerator, so a vehicle traveling at constant speed on a circular path would be reported as having nonzero longitudinal and lateral accelerations. On the paper’s own 64-m-radius road, that error is not negligible, and it invalidates the dynamic feasibility claims in Eq. (19).\n\nThe validation is much thinner than claimed. Only two hand-picked cases are shown; there is no real-world experiment, no randomized benchmark, no sensitivity study, and no error bars on the convergence comparison. The field coefficients and thresholds in Eqs. (9)–(14), the lane-change thresholds in Eq. (27), and the cost weights in Eq. (23) are all undisclosed, so the reported lane-change times and safety margins are outputs of the model rather than independent predictions. The lane-change feasibility function Φ(Ξ) in Eq. (14) is introduced but never defined.\n\nWho is this for? Someone assembling a baseline planner might skim the integration idea, but the paper is not reliable as a reference. The core transformation error and the missing validation mean the results cannot be trusted. A revised version with corrected kinematics, full parameter tables, and real or at least randomized simulation data would be worth another look.\n\nRecommendation: desk reject. It should not consume referee time in its current form.","headline":"A readable integration of known planner components, but the Frenet dynamic transformation is wrong and the promised real-world validation is absent.","tokens_in":9354,"tokens_out":4598,"would_cite":false,"duration_ms":41496,"reading_group":"no","serious_thinker":"no","would_accept_peer_review":false},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"This paper claims that an adaptive risk field, Frenet coordinates, and an improved particle swarm optimizer can together produce safe, comfortable lane changes on curved roads, and supports the claim with simulations of a two-lane…","keywords":["autonomous driving","interactive driving","curvy road","risk field","quintic polynomial curve","particle swarm optimization","Frenet coordinates","trajectory planning"],"falsifier":"Run a round-trip test on the published transformation: take a known circular-arc trajectory of radius 64 m with a known lateral offset, convert the global state to Frenet coordinates with Eq. (6), convert back with Eqs. (7)-(8), and compare with the original global state. If the round-trip error grows with curvature, or if the sign of the lateral offset does not match a point whose left or right side is known, the transformation is not correct and the reported trajectories are not truly road-aligned.","tokens_in":8156,"feed_emoji":"🚗","tokens_out":7355,"duration_ms":60231,"temperature":0.7,"pith_summary":"This paper tries to establish that a single planner can handle interactive lane changes on curved roads by combining three components usually used separately: an artificial-potential-field risk field for safety, Frenet coordinates for road-aligned motion, and an improved particle swarm optimizer (IPSO) for trajectory selection. The proposed planner decides when to change lanes by comparing repulsive, attractive, and lane-change risk fields, generates smooth candidate paths with quintic polynomials in the Frenet frame, and picks the best path under comfort and vehicle-dynamics constraints. Simulations on a two-lane road with inner radius 64 m and outer radius 70 m show the ego vehicle completing lane changes while maintaining safety margins and keeping lateral acceleration below 0.4g. The authors also report that IPSO converges in about 0.004 seconds, faster than standard PSO, a genetic algorithm, and an active-set method. If the framework works as described, it offers a parameter-based alternative to learning-based planners for curved-road interactive driving.","feed_headline":"Risk-field planner safely changes lanes on curved roads","feed_subtitle":"It combines Frenet coordinates and an improved swarm optimizer so lane changes stay smooth and safe.","key_machinery":"The central object is the combined risk-field-Frenet-IPSO planner. The risk field is a sum of three potential functions: lane-keeping attraction Ua, front-vehicle repulsion Ub, and lane-change field Uc with a feasibility function Φ(Ξ). The Frenet frame, using arc length s and lateral offset d, converts curved-road planning into a road-aligned problem, and the quintic polynomial y(t) = a0 + a1 t + a2 $t^{2}$ + a3 $t^{3}$ + a4 $t^{4}$ + a5 $t^{5}$ supplies smooth candidate lateral trajectories. IPSO then searches the polynomial parameter space using an adaptive inertia weight w(t) and time-varying cognitive and social coefficients c1(t) and c2(t), with a penalty function handling constraints. The mechanism that carries the argument is the coupling: the risk fields decide whether and when to change lanes, and the Frenet-based quintic generator plus IPSO decides how, balancing safety and comfort within dynamic limits.","core_discovery":"The central claim is that interactive lane changing on curved roads can be solved by an integrated framework in which a dynamic risk field, adapted from artificial potential fields, quantifies the threat from front, rear, and adjacent vehicles and triggers lane changes when three conditions hold: the front repulsive field exceeds its threshold, the lane-change risk field is below its threshold, and the attraction field favors the adjacent lane. In the Frenet frame of the reference path, a quintic polynomial generates smooth lateral-position profiles from boundary conditions, and IPSO with adaptive inertia and learning factors minimizes a weighted cost of terminal position, lateral acceleration, yaw rate, side-slip, and steering angle subject to dynamic constraints. According to the paper, this yields safe, comfortable, dynamically feasible lane changes on a 64-70 m radius curved road, with faster convergence than PSO, GA, and Active Set methods.","pith_inferences":["Implicit in the paper: because the risk-field layer is agnostic to the road's curvature except through the Frenet frame, the same trigger logic should transfer to non-circular roads, such as highway curves or mountain roads, once a reference path is given; the paper only demonstrates circular geometry.","A testable extension: replace the fixed threshold values in the lane-change triggering condition with adaptive thresholds learned from traffic density or time-to-collision, and measure whether earlier or later triggers reduce the cost function; the paper does not specify how the thresholds were set.","The convergence comparison is measured on a single scenario with one small difference between IPSO and PSO (0.004 vs 0.005 s); an independent check on multiple scenarios with random initializations would show whether the adaptive parameters, rather than chance, produce the speed-up."],"forward_implications":["The planner can complete interactive lane changes on a curved two-lane road with inner radius 64 m and outer radius 70 m while keeping lateral acceleration within 0.4g and yaw rate within friction limits.","Lane-change timing adapts to surrounding traffic: when the rear vehicle is faster than the adjacent vehicle, the repulsive field from the closing gap grows and triggers an earlier lane change.","The IPSO optimizer with adaptive inertia and time-varying learning factors converges to a feasible trajectory in about 0.004 seconds, faster than standard PSO (0.005 s), GA (0.038 s), and Active Set (0.052 s) on the tested case.","Because the risk fields are recomputed from the current vehicle states, the framework can adjust lane-change intention online as the front, rear, and adjacent vehicles change speed or position."],"supporting_citations":[{"why":"Provides the artificial-potential-field method that the proposed risk field adapts for vehicle interactions.","marker":"[19]"},{"why":"Supplies the integrated decision-making-and-motion-planning idea the framework builds on.","marker":"[20]"},{"why":"Supplies the Frenet-coordinate path-planning approach used for curved roads.","marker":"[21]"},{"why":"Supports using Frenet coordinates in urban autonomous driving.","marker":"[22]"},{"why":"Provides the genetic-algorithm baseline in the convergence comparison.","marker":"[23]"},{"why":"Provides the active-set baseline in the convergence comparison.","marker":"[24]"}],"fun_headline_variants":["Adaptive field planner handles curved-road lane changes safely","Swarm-optimized risk field makes curved-road lane changes safe","Curved-road lane changes made safe with adaptive field planner","Integrated APF-Frenet-IPSO planner ensures safe lane changes on curves","Safe curvilinear lane changes via dynamic risk field and IPSO"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Frenet coordinate transformation in Section II-D, specifically Eq. (6), correctly maps global positions, velocities, and accelerations to road-aligned coordinates on curved reference paths; if that transformation is wrong or numerically unstable, every trajectory produced by the downstream optimizer is invalid.","fun_headline_variants_meta":{"raw":{"variants":["Adaptive field planner handles curved-road lane changes safely","Swarm-optimized risk field makes curved-road lane changes safe","Curved-road lane changes made safe with adaptive field planner","Integrated APF-Frenet-IPSO planner ensures safe lane changes on curves","Safe curvilinear lane changes via dynamic risk field and IPSO"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000555,"raw_usage":{"total_tokens":2631,"prompt_tokens":923,"completion_tokens":1708,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":539,"completion_tokens_details":{"reasoning_tokens":1620}},"tokens_in":539,"tokens_out":1708,"duration_ms":10950,"temperature":1.0,"reasoning_tokens":1620,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T11:41:18.873162+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a round-trip test on the published transformation: take a known circular-arc trajectory of radius 64 m with a known lateral offset, convert the global state to Frenet coordinates with Eq. (6), convert back with Eqs. (7)-(8), and compare with the original global state. If the round-trip error grows with curvature, or if the sign of the lateral offset does not match a point whose left or right side is known, the transformation is not correct and the reported trajectories are not truly road-aligned.","supporting_citations":[{"cited_title":"An integrated framework of decision making and motion planning for autonomous vehicles considering social behaviors,","cited_arxiv_id":null,"evidence_quote":"Supplies the integrated decision-making-and-motion-planning idea the framework builds on."},{"cited_title":"Research on path planning for autonomous vehicle based on frenet system,","cited_arxiv_id":null,"evidence_quote":"Supplies the Frenet-coordinate path-planning approach used for curved roads."},{"cited_title":"Lf-net: A learning-based frenet planning approach for urban autonomous driving,","cited_arxiv_id":null,"evidence_quote":"Supports using Frenet coordinates in urban autonomous driving."},{"cited_title":"Optimal trajectory planning and robust tracking using vehicle model inversion,","cited_arxiv_id":null,"evidence_quote":"Provides the genetic-algorithm baseline in the convergence comparison."},{"cited_title":"An active-set algorithm for norm constrained quadratic problems,","cited_arxiv_id":null,"evidence_quote":"Provides the active-set baseline in the convergence comparison."}],"review_version":1}