{"id":"55ffcfec-f81e-47cd-9f57-21d6b6f7b991","arxiv_id":"2506.04684","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"An MPC framework with curvature-based weight tuning tracks reference trajectories on a real drive-by-wire vehicle with reported cross-track errors below about 3 meters, but key model details and timing data are missing.","lead":"This paper describes a model predictive controller that tracks planned paths for an autonomous vehicle, using a linear-parameter-varying vehicle model and weight tuning based on path curvature. The authors test the controller on six trajectories in simulation and on a real drive-by-wire vehicle at a university test track, reporting small path errors.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Eq. (20) propagates A and B as constants over the whole horizon, but Eqs. (12)-(13) make them functions of the state and of the control input delta; without either a time-varying prediction or an explicit nonlinear f(x,u) in CasADi, the LPV derivation is internally inconsistent and does not…","rationale":"The reader correctly identifies the weakest loaded assumption: the LPV prediction derivation requires A_d and B_d to be constant over the horizon, but the paper's own equations make them depend on the state and on the control input delta. This is a genuine internal inconsistency, not a matter of consensus: Eqs. (12)-(13) contain state and input terms, Eq. (20) treats A and B as fixed matrices, and Section IV-E explicitly switches between f(x_k, u_k) and A_d x_k + B_d u_k. If the implementation uses f(x_k, u_k), then the reported experiments can still be evidence for a working nonlinear MPC, but the paper's stated LPV-based contribution and the derivation in Eqs. (17)-(20) are unsupported. If the implementation uses frozen A_d and B_d, the prediction model is mathematically wrong for aggressive trajectories where x_dot, y_dot, psi, and delta vary substantially over the horizon. Either way, the central technical claim needs clarification or revision. The real-vehicle experiments, the ROS-based modular architecture, and the Table III metrics are meaningful engineering evidence, but they do not resolve this model inconsistency because the experiments cannot show which prediction model was actually used. I agree with the reader's CONDITIONAL verdict: the paper is not beyond salvage, since a clear statement of the actual prediction model, released code, or a corrected quasi-LPV formulation would address the issue, but as written the contradiction is load-bearing. No adjustment to the reader's verdict is needed.","tokens_in":15408,"tokens_out":5062,"duration_ms":63060,"concrete_test":"Ask the authors for the CasADi formulation, or reimplement it from Section IV-E, and print the symbolic expression used as the dynamics constraint. If it is f(x_k, u_k) from Eqs. (1)-(8), rerun trajectory T1 with the constraint replaced by the affine form x_{k+1} = A_d x_k + B_d u_k, with A_d and B_d evaluated once per sample from Eqs. (12)-(13) and held constant over the horizon, and compare the max CTE to Table III. If max CTE changes by more than 0.3 m or the solver becomes infeasible, the reported accuracy depends on the nonlinear propagation and the LPV derivation is not load-bearing. If the code already uses the affine form, recompute the same T1 experiment while re-evaluating A_d and B_d at every predicted state and input along the horizon; a large mismatch between predicted and actual states confirms that the frozen-matrix expansion in Eq. (20) is invalid.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim, real-time LPV-based NMPC with high tracking accuracy, rests on the discrete-time affine prediction in Eq. (17) and the closed-form expansion in Eq. (20). That expansion is valid only if A_d and B_d are fixed over the prediction horizon. They are not: Eq. (12) contains 1/x_dot, y_dot, and sin(delta), and Eq. (13) contains sin(delta) and cos(delta), with delta itself one of the optimization variables u_k. The paper even states in Section III-A that 'A is a function of X and B is the function of u.' Hence A^k x0 and A^(k-1-i) B u_i in Eq. (19) are not well-defined matrix powers once the state and input vary along the horizon. The CasADi description in Section IV-E is contradictory: it says the prediction model is x_{k+1} = f(x_k, u_k), but the formal optimization problem lists the dynamics as x_{k+1} = A_d x_k + B_d u_k. If the actual code uses f(x_k, u_k), the experiments demonstrate nonlinear MPC, not the LPV formulation that the title and contributions claim; if it uses frozen A_d and B_d, the predictor ignores known nonlinear couplings, and the reported 'minimal errors' under aggressive maneuvers are not explained by the presented model. This is an internal inconsistency, not merely a disagreement with a particular control paradigm.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Linear Parameter Varying (LPV) based Nonlinear Model Predictive Control (NMPC) framework for trajectory tracking in autonomous vehicles. The authors formulate a bicycle model with lateral tire forces in an LPV state-space form, introduce a curvature-based method for tuning the MPC weight matrices, and implement the controller in a ROS-based architecture with parallel state estimation and optimization. The system is evaluated in simulation and in real-time experiments on a drive-by-wire vehicle at the TiHAN testbed, using six trajectories of varying curvature and speed. Performance is reported in terms of cross-track error, lateral error, and orientation error for both simulation and real-time runs. The central claims are that the LPV formulation enables real-time computation and that the curvature-based tuning generalizes across trajectories.","tokens_in":15808,"tokens_out":7991,"duration_ms":90972,"significance":"The experimental effort is a clear strength: the authors demonstrate a complete MPC stack on a real vehicle, integrating GPS/IMU state estimation, CasADi optimization, and ROS communication, and they report detailed per-trajectory error metrics. The idea of scheduling MPC weights on path curvature is practically appealing and, if properly validated, could be a useful engineering contribution. However, the LPV derivation is internally inconsistent, the discrete-time prediction model is not specified with adequate rigor, and the tuning evaluation is in-sample. As presented, the paper does not support the stated LPV-based contribution or the generalization claim about tuning. The paper would be suitable for a major revision if the authors can resolve these issues and add appropriate validation.","major_comments":[{"comment":"The prediction formula x_k = A^k x0 + sum_{i=0}^{k-1} A^{k-1-i} B u_i is valid only if A_d and B_d are constant over the prediction horizon. However, the continuous-time matrices in Eqs. (12)-(13) are explicit functions of the state (1/x_dot, y_dot, psi) and of the control input delta. Since delta is one of the optimization variables, the closed-form matrix powers in Eq. (20) are not well-defined. The manuscript itself states in Sec. III-A that 'A is a function of X and B is the function of u.' This is an internal inconsistency that must be resolved: either define a genuinely parameter-varying model with a scheduling variable independent of u, or use the nonlinear propagation f(x_k, u_k) in CasADi and remove the LPV claim from the contribution.","section":"Sec. IV-B, Eqs. (19)-(20)"},{"comment":"The state vector is defined inconsistently. The text in Sec. III-A lists x = [x, y, psi, v]^T, while Eq. (11) defines x = [x_dot, y_dot, psi, psi_dot, X, Y]^T. The cost function in Eq. (22) uses four error states (x_dot, psi, X, Y), but the system matrices in Eqs. (12)-(13) are 6x6 and 6x2. The dimensions of A_d, B_d, C, Q, S, and the error vector are therefore ambiguous. Specify the exact state dimension and ensure that all matrices and cost-function terms conform to that dimension.","section":"Sec. III-A and Eq. (11)"},{"comment":"The discretization procedure is not described. The sentence 'The discrete system matrices A_d and B_d are derived directly from the continuous-time counterparts' leaves open whether forward Euler, zero-order hold, or a matrix exponential was used. This matters because the continuous matrices are state- and input-dependent, and a standard LTI discretization does not apply. The sampling time T_s is also never given. Without this information, the prediction model cannot be reproduced, which directly affects the validity of the reported simulation and real-time results.","section":"Sec. IV-A"},{"comment":"The weight matrices Q and S are tuned using the curvature of the same six trajectories on which the controller is then evaluated. This makes the error metrics in Table III in-sample results and does not provide evidence that the curvature-based tuning generalizes to unseen paths. The actual weight values and the curvature thresholds of the 'generalized framework' are not reported. To substantiate the generalization claim, the authors should either evaluate on held-out trajectories or provide a cross-validation study, and they should disclose the tuned weights and thresholds.","section":"Sec. V-C and Table III"},{"comment":"The real-time claim is not quantified. No computation time per MPC iteration is reported, even though real-time performance is a stated contribution. Moreover, no baseline controller (e.g., PID, linear MPC, or a standard NMPC) is compared, so the 'robustness' and 'high accuracy' claims are relative only to the reference trajectory and not to a competing method. Please add timing data and at least one baseline comparison.","section":"Sec. VI"}],"minor_comments":[{"comment":"The text after Eq. (18) says 'the discrete-time state-space equations are expressed in Eq 18', but the state equation is Eq. (17). Please correct the cross-reference.","section":"Sec. IV-B"},{"comment":"Several references are duplicated: [8] and [14] are the same paper, [10] and [23] are the same paper, and [13] and [25] are the same paper. Please deduplicate the bibliography.","section":"References"},{"comment":"The (2,2) entry of matrix A appears to have a typo: 'C_alpha_f + C_alpha_f cos(delta)' is likely meant to be 'C_alpha_f + C_alpha_r cos(delta)'. Please check the notation carefully.","section":"Eq. (12)"},{"comment":"The sentence 'A is a function of X and B is the function of u' is unclear and grammatically awkward; moreover, the scheduling variable v in A(v) and B(v) is never defined. Please clarify the scheduling parameter and its role in the LPV model.","section":"Sec. III-A"},{"comment":"The columns 'Max. long. Acceleration' and 'Max. lat. Acceleration' are given without units. Please add units (m/s^2) for completeness.","section":"Table II"},{"comment":"In the definition of s_i, the word 'parameters' should be singular: 'the parameter s represents the cumulative arc length.'","section":"Sec. V-C"},{"comment":"The y-axis labels 'lateral error' and 'orientation error' are undefined; please specify whether these are signed errors, absolute errors, or normalized values, and give units.","section":"Figs. 7-8"}],"recommendation":"major_revision","confidential_remarks":"The experimental platform and the reported tracking results, taken at face value, indicate a functioning MPC implementation on a real vehicle. The main concerns are the internally inconsistent LPV derivation, the unspecified discretization, and the in-sample tuning evaluation. In my judgment, a major revision could remedy these issues if the authors re-frame the contribution as a nonlinear MPC engineering study with a clearly specified prediction model and add held-out validation. The duplicate references and the multiple citations to the authors' own group should also be cleaned up during revision."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—the short version: this is an honest engineering paper with real experiments, but the LPV formulation in the math does not match what the experiments likely did, and the tuning is in-sample. I'd send it out rather than desk reject, but the authors have work to do.\n\nWhat's new: they build a ROS-based MPC pipeline for a drive-by-wire vehicle, run six trajectories in simulation and real-time, report CTE/orientation errors, and propose a curvature-based heuristic for choosing Q and S weights. That curvature-tuning rule is a small but real practical contribution, and the experimental setup is credible: IIT Hyderabad test track, Jetson AGX Orin, GPS/IMU localization, parallel ROS nodes for state estimation and optimization. Good that they separate simulation vs real-time numbers.\n\nThe big problem is the model inconsistency. The paper says A(v) and B(v) are functions of X and u, and indeed Eqs. (12)-(13) contain sin(delta), cos(delta), and 1/x_dot. But the prediction expansion in Eq. (19)-(20) treats A and B as constant over the horizon. You can't raise a state-dependent matrix to powers and factor it out of a sum when the state and input are changing. The authors even write x_{k+1}=f(x_k,u_k) in the CasADi section, and then list the formal dynamics as A_d x_k + B_d u_k. Those two descriptions are contradictory. If the implementation uses f(x,u), then what they actually ran is nonlinear MPC, not LPV-MPC, and the title's LPV contribution is a presentation layer. If they froze A_d and B_d, then the predictor ignores the couplings the model itself says matter. Either way, the derivation doesn't support the claim.\n\nSoft spot two is evaluation. The Q and S weights are tuned using curvature of the same six trajectories on which performance is reported (Section V-C, Table III). That's in-sample fitting, not a test of the tuning rule. There's no held-out trajectory, no comparison to fixed weights, no sensitivity analysis. Mean CTEs of 0.16-0.26 m on a real vehicle are fine, but we don't know whether the curvature rule adds anything over a single well-tuned weight set.\n\nAlso minor: no discretization scheme or horizon length is specified, and the real-time latency claim is never quantified.\n\nOn the citation pattern: it's broad, relevant, and doesn't look like self-citation padding. The related work is a bit list-y but fair.\n\nWho this is for: practitioners doing vehicle MPC deployment who want a template for ROS implementation and a curvature-weighting idea. The control theorist will be annoyed by the math gap. It deserves a serious referee because the experimental base is real and the issues are fixable in revision: rewrite the model section honestly (say it uses nonlinear propagation), add held-out validation and fixed-weight baseline, disclose the weights and discretization. I'd recommend conditional acceptance with major revision.","headline":"A real vehicle experiment that deserves a referee, but the LPV derivation has a load-bearing gap and the tuning is evaluated in-sample.","tokens_in":16252,"tokens_out":2165,"would_cite":false,"duration_ms":24801,"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":"A curvature-tuned, LPV-based model predictive controller tracks a real autonomous vehicle with mean cross-track error below 0.3 m on six test trajectories.","keywords":["model predictive control","linear parameter varying","trajectory tracking","autonomous vehicles","curvature-based tuning","real-time control","vehicle dynamics"],"falsifier":"Inspect the optimizer's internal dynamics: if the implemented predictor is the nonlinear function x_{k+1}=f(x_k,u_k) rather than the constant-matrix A_d x_k + B_d u_k of Eq. (20), then the LPV machinery is not what achieves the reported accuracy. A direct test is to run the controller on a sharp turn (e.g., trajectory T1) with the LPV prediction model forced to use the frozen matrices, and compare the predicted next state against the actual vehicle state; a divergence larger than the reported cross-track error would falsify the claim that the LPV representation alone supports the tracking performance.","tokens_in":15210,"feed_emoji":"🚗","tokens_out":10741,"duration_ms":101718,"temperature":0.7,"pith_summary":"The paper sets out to show that a model predictive controller, built on a linear parameter varying (LPV) representation of a vehicle's nonlinear dynamics and tuned by path curvature, can track reference trajectories on a real autonomous vehicle in real time. The proposed framework runs state estimation and control optimization in parallel, and it is tested on six trajectories spanning quiet roads to aggressive, high-speed maneuvers. The reported results claim close agreement between simulation and real-time behavior, with mean cross-track errors below 0.3 meters in all six real-world tests and maximum cross-track errors below 2.9 meters even on the sharpest course. If these claims hold, the work matters because it offers a practical recipe for transferring MPC-based trajectory tracking from simulation to a physical vehicle with modest computational hardware.","feed_headline":"Curvature-tuned MPC keeps a real vehicle's mean path error under 0.3 m","feed_subtitle":"Six test runs, from quiet to aggressive, show simulation and real-time tracking line up.","key_machinery":"The load-bearing mechanism is the LPV state-space model of the vehicle, in which the nonlinear kinematic-bicycle equations are packed into velocity-dependent matrices A(v) and B(v), so the optimizer sees a 'linear-like' prediction model at each step. A second named mechanism is the curvature-based tuning rule: the path's sharpest curvature selects the diagonal entries of the state-weight matrix Q and the terminal-weight matrix S, so sharp turns are penalized more heavily. The resulting receding-horizon optimization is solved symbolically at each sample, and the whole pipeline is split into concurrent processes so that state estimation and control computation run in parallel and only the first control input is applied to the vehicle.","core_discovery":"On the paper's own terms, the central discovery is that an NMPC controller whose internal model is the LPV form of the kinematic bicycle dynamics—velocity-dependent system matrices A(v) and B(v)—can be made to track six distinct reference paths on a drive-by-wire vehicle with mean cross-track error at or under about 0.26 m and mean orientation error at or under 0.13 rad, while keeping the maximum cross-track error below about 2.9 m. The authors attribute this performance to a curvature-based weight-tuning scheme: the sharpest curvature of each path selects the state-weight (Q) and terminal-weight (S) matrices, so sharper turns are penalized more heavily in the cost function. The claimed outcome is that simulation and real-time results align closely, meaning the controller transfers from a virtual model to the actual vehicle with only modest degradation from sensor noise and actuation delay.","pith_inferences":["Editorial inference: the constant-matrix prediction formula in Eq. (20) is not consistent with the state- and input-dependent matrices A and B in Eqs. (12)–(13); if the deployed solver actually propagates states with the nonlinear function f(x_k,u_k), then the LPV derivation is not the operative predictor and the reported performance is produced by the nonlinear model instead.","Editorial inference: a direct testable extension, which the paper lists as future work, is to make the curvature-based tuning online and receding-horizon, so Q and S adapt to local curvature rather than to the path's single sharpest point.","Editorial inference: the reported real-time maximum cross-track errors imply that on roads narrower than about 3 m the controller would need either stronger position weights or tighter bounds; a narrow-track test would directly probe how much of the claimed tracking performance survives a harder geometry."],"forward_implications":["Both fixed-speed and variable-speed operation are handled by the same controller, so a single MPC can slow the vehicle into sharp turns and accelerate on straights without reconfiguration.","Curvature-based tuning is claimed to generalize to new trajectories: compute the path's sharpest curvature, choose the matching weight matrices, and run the controller without manual re-tuning.","The reported closeness between simulation and real-time results means the development loop can be largely closed on a virtual vehicle, with real-vehicle runs used only to fine-tune for actuation delay.","On the most aggressive trajectory (T1), the real-time maximum cross-track error is 2.86 m, so the worst-case excursions are substantially larger than the mean errors across all six paths."],"supporting_citations":[{"why":"Supplies the LPV modeling approach used to write the nonlinear vehicle equations in velocity-parameterized matrix form.","marker":"[19]"},{"why":"Shows a real-time LPV-based MPC implementation, supporting the claim that this model class is tractable at control rates.","marker":"[20]"},{"why":"Provides the lateral tire force expressions (cornering stiffness, slip-like terms) used in the vehicle dynamics equations.","marker":"[34]"},{"why":"Establishes the kinematic bicycle model that the state-space equations are derived from.","marker":"[35]"},{"why":"Defines the cross-track, lateral-error, and orientation-error metrics used to report tracking accuracy.","marker":"[36]"},{"why":"Provides the CasADi symbolic optimization library that formulates and solves the MPC cost function in real time.","marker":"[38]"}],"fun_headline_variants":["Real-time LPV-MPC keeps mean track error under 0.3 m on real car","Curvature-tuned MPC: mean path error 0.26 m in live tests","LPV-MPC matches simulation to reality with 0.26 m mean error","Weight tuning by curvature drives MPC tracking error below 0.3 m","MPC with curvature-based weights: real car tracks within 0.3 m"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the nonlinear vehicle dynamics can be represented, over each prediction horizon, by a linear model with fixed matrices A_d and B_d; if the real motion during a sharp turn departs from this frozen model, the controller's predicted states no longer match the vehicle.","fun_headline_variants_meta":{"raw":{"variants":["Real-time LPV-MPC keeps mean track error under 0.3 m on real car","Curvature-tuned MPC: mean path error 0.26 m in live tests","LPV-MPC matches simulation to reality with 0.26 m mean error","Weight tuning by curvature drives MPC tracking error below 0.3 m","MPC with curvature-based weights: real car tracks within 0.3 m"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000735,"raw_usage":{"total_tokens":3265,"prompt_tokens":907,"completion_tokens":2358,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":523,"completion_tokens_details":{"reasoning_tokens":2249}},"tokens_in":523,"tokens_out":2358,"duration_ms":18033,"temperature":1.0,"reasoning_tokens":2249,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T10:35:08.880480+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Inspect the optimizer's internal dynamics: if the implemented predictor is the nonlinear function x_{k+1}=f(x_k,u_k) rather than the constant-matrix A_d x_k + B_d u_k of Eq. (20), then the LPV machinery is not what achieves the reported accuracy. A direct test is to run the controller on a sharp turn (e.g., trajectory T1) with the LPV prediction model forced to use the frozen matrices, and compare the predicted next state against the actual vehicle state; a divergence larger than the reported cross-track error would falsify the claim that the LPV representation alone supports the tracking performance.","supporting_citations":[{"cited_title":"In: IEEE International Conference on Mechatronics and Automation, Niagara Falls (2005)","cited_arxiv_id":null,"evidence_quote":"Provides the lateral tire force expressions (cornering stiffness, slip-like terms) used in the vehicle dynamics equations."},{"cited_title":"Katriniok, et al., ”Optimal vehicle dynamics control for combined longitudinal and lateral autonomous vehicle guidance,” in Proc","cited_arxiv_id":null,"evidence_quote":"Establishes the kinematic bicycle model that the state-space equations are derived from."},{"cited_title":"Rokonuzzaman, N","cited_arxiv_id":null,"evidence_quote":"Defines the cross-track, lateral-error, and orientation-error metrics used to report tracking accuracy."},{"cited_title":"Andersson, J","cited_arxiv_id":null,"evidence_quote":"Provides the CasADi symbolic optimization library that formulates and solves the MPC cost function in real time."}],"review_version":1}