{"id":"8a6175b4-3701-45ca-a1fc-0f746f16a3f8","arxiv_id":"2411.13885","paper_version":1,"verdict":"REJECT","confidence":"LOW","novelty_score":3.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"The paper claims that adding Frenet coordinates to DDPG reduces lateral tracking error in Gazebo simulations, but the evidence is qualitative, underspecified, and not reproducible.","lead":"These authors propose putting a road-relative coordinate system, the Frenet frame, inside a standard reinforcement-learning controller for self-driving cars, and they report that it tracks reference paths more accurately than two baseline controllers. The experiments are Gazebo simulations with no code, no error bars, and no reward function, so the result is not yet verifiable.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The training objective for the claimed DDPG controller is never defined: Eq. (6) is presented as a 'loss' but is never connected to the Bellman/Q objective that DDPG actually optimizes, so the reported comparison cannot support the central claim.","rationale":"The paper's contribution is an empirical comparison, and its central claim is a performance assertion. Performance assertions are only meaningful given a fully specified training objective and evaluation protocol. The missing reward function is the single most load-bearing gap because it is a precondition for the method to be DDPG at all. The reader's weakest assumption, concerning the curvature terms in Eq. (1), is secondary: even if the transformation is perfect, an undefined objective invalidates the result. My concern is therefore a correctness/evidentiary risk, not a disagreement with consensus. The paper contains no machine-checked proofs, no code, no hyperparameters, and no statistical analysis; those are real gaps, and the paper itself does not claim otherwise. If the missing details were supplied and the comparison were reproduced, the claim could be salvaged, but as written the evidence is insufficient. I therefore keep the reader's REJECT verdict; no new information changes it, though I arrive at it via a different load-bearing concern than the one named as the reader's weakest assumption.","tokens_in":6438,"tokens_out":5000,"duration_ms":51142,"concrete_test":"Obtain, or require the authors to specify, the exact reward function r_t(s_t, a_t) used in Eq. (2) and the precise place Eq. (6) enters training. The decisive check is to re-run one scenario with a documented reward (e.g., r_t = -L_t with L_t from Eq. (6)) and explicit episode/discount settings, then report mean and standard deviation of lateral error over at least 10 random seeds for Frenet-DDPG, Cartesian DDPG, and PPO. If the reported advantage disappears, or if no reward function can be provided because Eq. (6) was used directly as a supervised loss, the central claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that a DDPG policy operating on Frenet states achieves lower lateral error than Cartesian DDPG and PPO. For that claim to be true, a DDPG agent must be trained, which requires a reward signal and a policy-gradient/critic update. The paper defines neither. Section II-B gives the standard DDPG update equations (Eqs. (2)-(5)), but Section II-C then introduces Eq. (6), L = l + l_dot + l_ddot, as 'the loss function,' saying the goal is to minimize lateral error, velocity, and acceleration. This is never connected to the DDPG objective: the actor in Eq. (4) maximizes the critic's Q-value, while the critic in Eq. (3) minimizes a Bellman residual. Unless Eq. (6) is the reward, or is otherwise used inside the Q-target, it cannot be the thing DDPG minimizes. The experimental section records only 'average losses' and one qualitative trajectory (Fig. 4), with no lateral-error numbers, no seeds, and no reward specification. Even if the Frenet transformation in Eq. (1) is implemented exactly, the absence of a defined training signal means the reported 'DDPG+Frenet' method could be a direct supervised error minimizer rather than a DDPG policy, making the central comparison uninterpretable and the claim unsupported.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes combining a Serret-Frenet coordinate transformation with the deep deterministic policy gradient (DDPG) algorithm for vehicle trajectory tracking. It claims that a DDPG controller operating on Frenet-state observations tracks predefined paths with high accuracy and stability, significantly reducing lateral error relative to Cartesian-state DDPG and PPO in Gazebo simulations. The paper presents the Frenet transformation in Eq. (1), standard DDPG update equations in Eqs. (2)-(5), and a 'loss function' in Eq. (6), followed by a qualitative experimental section with training-loss and tracking-error figures.","tokens_in":6753,"tokens_out":5768,"duration_ms":59364,"significance":"If the central claim were fully supported, a Frenet-state DDPG controller that reliably reduces lateral error compared with Cartesian-state DDPG and PPO in a realistic simulator would be a useful empirical contribution to autonomous driving research. The choice of baselines and the use of a local frame for lateral control are well motivated. However, as submitted, the paper does not define the learning objective actually optimized by DDPG, provides no quantitative performance values, and gives no experimental protocol that would allow reproduction. The contribution is therefore prospective rather than established; no code, data, or machine-checked results are supplied that would mitigate the missing details.","major_comments":[{"comment":"The central learning objective is undefined. Equation (6), L = l + l_dot + l_ddot, is introduced as 'the loss function,' but no connection is made to the DDPG objective in Eqs. (2)-(5): the critic in Eq. (3) minimizes a Bellman residual, and the actor in Eq. (4) maximizes the critic's Q-value. No reward function r_i is defined anywhere in the paper. If Eq. (6) is intended to be the reward, that must be stated and substituted into Eq. (2), but the paper never says this. As written, the reported 'DDPG+Frenet' procedure could be a supervised minimizer of lateral-error terms rather than a policy-gradient method, which makes the comparison and the central claim uninterpretable.","section":"II-C, Eq. (6)"},{"comment":"The central quantitative claim is not supported by any reported numbers. Figure 4(b) is described qualitatively as showing the smallest and smoothest lateral error, but the manuscript reports no lateral-error values, no standard deviations, no number of trials or seeds, and no statistical tests. Figure 3 plots 'average losses' that are never defined in relation to the DDPG updates of Eqs. (2)-(5). Without such quantitative results, the claimed 'significantly reducing lateral errors' cannot be assessed.","section":"III-B, Fig. 4"},{"comment":"The paper's motivation and implementation appear contradictory. The text says the Frenet coordinate system avoids considering road curvature, yet Eq. (1) explicitly includes the reference-path curvature k_r and its derivative k_r'. The authors must state whether curvature terms enter the implemented state; if they do not, the transformation is not a proper Frenet frame, and if they do, the accompanying motivating sentence is incorrect.","section":"II-A, Eq. (1)"},{"comment":"The evaluation metric overlaps with the training objective in a way that is not acknowledged. Since Eq. (6) minimizes lateral position, velocity, and acceleration, and Fig. 4(b) evaluates lateral error, part of the reported success is by construction. The comparison against Cartesian-state DDPG and PPO provides some independent grounding, but only if the same reward or loss specification is given for all baselines; no such specification appears.","section":"II-C and III"}],"minor_comments":[{"comment":"The experimental protocol is incomplete: road shape, vehicle model, sampling frequency, action bounds, neural-network architectures, learning rates, discount factor, soft-update coefficient, replay buffer size, exploration-noise schedule, and number of training episodes are all missing. The statement that the 'SGD optimizer' was used is insufficient for reproduction of DDPG.","section":"III"},{"comment":"Equation (1) is typeset in garbled form with unpaired parentheses and split expressions, and the notation is inconsistent (for example, dot derivatives versus derivatives with respect to s). A clean, self-contained derivation of the Frenet transformation is needed.","section":"II-A, Eq. (1)"},{"comment":"The figures are not self-explanatory: Fig. 3 does not define what 'loss value' means, and Fig. 4(b) has no axis labels or units on the lateral-error plot.","section":"III, Figs. 3 and 4"},{"comment":"Several references, such as [19], [20], [25], [29], and [31], are unrelated to trajectory tracking or Frenet coordinates and are never discussed in the body; the bibliography should be aligned with the content of the paper.","section":"References"}],"recommendation":"reject","confidential_remarks":"The reference list appears padded with unrelated papers; this is not the basis of my recommendation but may merit editorial attention. The submission is closer to an extended abstract than a full journal article, and the missing reward specification and quantitative evaluation are fundamental rather than cosmetic issues."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read it so you don't have to. The Frenet-coordinate state representation for DDPG is a reasonable idea, and the comparison against Cartesian-state DDPG and PPO is the right experiment to run. That is the extent of what is solid here.\n\nThe central claim is unsupported because the training objective is never defined. Section II-C introduces Eq. (6), L = l + l_dot + l_ddot, as \"the loss function,\" but never connects it to the DDPG critic Bellman update or the actor's Q-maximization in Eqs. (2)-(5). The paper reports \"average losses\" and one qualitative trajectory, but no reward function, no hyperparameters, no seeds, no quantitative lateral-error values, and no error bars. That is not a minor omission: if Eq. (6) is the reward or part of a supervised objective, the algorithm isn't DDPG as described; if it is not used anywhere, the reported training curves don't correspond to any well-defined optimization. Either way, the comparison cannot be interpreted.\n\nThere is also an internal contradiction in the coordinate transform. The text says Frenet coordinates avoid considering road curvature, yet Eq. (1) explicitly includes k_r and k_r'. The implementation must resolve this or the state representation isn't a proper Frenet frame. The reference list is another red flag: many citations (snore detection, pill identification, stock prediction, football formation) are unrelated to trajectory tracking or RL, which suggests padding rather than scholarship.\n\nTo their credit, the authors did run four method variants in Gazebo, and the qualitative figure is at least plausible. But the absence of any training-signal definition and any quantitative metrics means the paper cannot be evaluated as a technical contribution. It is not ready for peer review in this form; a serious referee would need a complete method specification and actual error statistics.\n\nI wouldn't cite this. If you want a cautionary example of how not to report RL experiments, it's usable, but not for the results.","headline":"The DDPG training objective is never defined, so the central trajectory-tracking claim is unsupported — a desk reject despite a sensible starting idea.","tokens_in":7291,"tokens_out":2234,"would_cite":false,"duration_ms":22285,"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":"A DDPG controller that expresses its state in Frenet coordinates tracks a predefined path with lower lateral error than Cartesian-state DDPG and PPO.","keywords":["DDPG","Frenet coordinates","trajectory tracking","path tracking","autonomous driving","reinforcement learning","lateral error","robot navigation"],"falsifier":"Re-run the comparison on a curved reference path using an independently verified Serret-Frenet transformation that includes the curvature terms of Eq. (1), and report the full distribution of lateral error over many episodes and seeds. If Frenet-state DDPG does not show a clear reduction in lateral error relative to Cartesian-state DDPG, or if dropping the curvature terms changes the result, the central claim is falsified.","tokens_in":6220,"feed_emoji":"🚗","tokens_out":9370,"duration_ms":80930,"temperature":0.7,"pith_summary":"This paper proposes a trajectory-tracking controller for autonomous driving in which the reinforcement-learning state is expressed in Frenet coordinates: distance traveled along the reference path, lateral deviation from it, and their velocities and accelerations. The claim is that this path-relative representation lets a DDPG agent separate along-road from cross-road motion, so the learned policy follows a predefined path with smaller and smoother lateral error. The authors support the claim with simulation training curves and tracking comparisons in which Frenet-based DDPG keeps the lowest loss and the smallest lateral deviation against Cartesian-state DDPG and PPO. If correct, the result is an algorithm-level improvement: no new network architecture or reward design is needed, only a change of coordinates for the state the policy sees.","feed_headline":"Frenet coordinates cut DDPG's path-tracking lateral error","feed_subtitle":"A simulation study reports lower and smoother lateral deviation than Cartesian-state DDPG and PPO on the same reference path.","key_machinery":"The load-bearing object is the Serret-Frenet coordinate map in Eq. (1), which rewrites the vehicle's Cartesian state relative to the reference path in terms of longitudinal coordinate $s$, lateral offset $l$, and their derivatives, including the reference-path curvature $k_r$ and its derivative $k_r'$. The DDPG agent receives this re-expressed state, chooses continuous steering and acceleration actions, and is trained against the loss $L = l + \\dot{l} + \\ddot{l}$, which concentrates the gradient on lateral behavior. The paper's argument is that this path-relative frame makes the policy's gradient point more directly at reducing cross-track error, so the same Actor-Critic machinery with experience replay and target networks converges to tighter and smoother tracking than it does with Cartesian states.","core_discovery":"The central discovery claimed is that replacing the Cartesian state representation of a DDPG trajectory-tracking agent with a Serret-Frenet representation materially improves tracking performance. Concretely, the paper converts the vehicle's position, heading, velocity, and acceleration into the Frenet coordinates $(s, \\dot{s}, \\ddot{s}, l, \\dot{l}, \\ddot{l}, l', l'')$, where $s$ is arc length along the reference path and $l$ is the lateral offset, and uses this state as the input to the DDPG actor and critic. The training objective is the loss $L = l + \\dot{l} + \\ddot{l}$, which penalizes lateral deviation, lateral velocity, and lateral acceleration. In the reported simulation experiments, the Frenet-state DDPG agent tracks the reference path more closely, keeps the lowest loss during training, and produces a smoother and smaller lateral-error curve than Cartesian-state DDPG and PPO.","pith_inferences":["Editorial inference: the performance gain most likely comes from making the state invariant to the path's absolute position and heading, so a policy trained in Frenet coordinates should generalize to unseen roads without retraining; the paper does not test this.","Editorial inference: the same coordinate change should also help other continuous-action reinforcement-learning algorithms such as PPO or soft actor-critic; the paper only tests PPO in Cartesian form, not PPO with Frenet states.","Editorial inference: the loss $L = l + \\dot{l} + \\ddot{l}$ adds quantities with different units and no weighting, so part of the reported gain may be implicit reward shaping; testing squared or weighted variants would show whether the benefit is robust.","Editorial inference: a direct extension is to benchmark Frenet-state DDPG against a model-predictive controller on the same paths, which would reveal how close the learned policy comes to a controller-based floor on lateral error."],"forward_implications":["Frenet-coordinate state encoding can be dropped into existing DDPG path-tracking pipelines without changing the network architecture or reward design.","The reported smoother lateral-error curve implies the learned policy is stable enough to follow a reference path without oscillation, a prerequisite for passenger comfort and safety.","Because the controller uses the vehicle's speed and position in the path-relative frame, steering can adapt to the current velocity rather than treating the path as a fixed geometric curve.","The method extends DDPG trajectory tracking to what the paper calls complex environments, supporting path planning and control for autonomous vehicles and mobile robots."],"supporting_citations":[{"why":"Supplies the DDPG algorithm, specifically its continuous-action Actor-Critic formulation, on which the proposed controller is built.","marker":"[14-18]"},{"why":"Provides the Frenet-coordinate frame in which the vehicle's Cartesian state is re-expressed relative to the reference path.","marker":"[23]"},{"why":"Provides the local-reference-frame trajectory-tracking background that motivates using Frenet coordinates.","marker":"[24]"},{"why":"Establishes the advantage of separating lateral and longitudinal control, the design rationale for the Frenet state representation.","marker":"[25]"},{"why":"Applies the lateral/longitudinal separation idea to path tracking, supporting the paper's choice of a path-relative state.","marker":"[26]"},{"why":"Supports converting position and velocity information into a Frenet framework for trajectory tracking.","marker":"[27-29]"}],"fun_headline_variants":["Frenet state slashes DDPG lateral tracking error","Frenet coordinates improve DDPG path-tracking precision","DDPG with Frenet state tracks paths more precisely","Frenet coords reduce DDPG's lateral deviation","Frenet beats Cartesian for DDPG tracking"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparison depends on Eq. (1) being a correctly implemented Frenet transformation, including the reference-path curvature $k_r$ and its derivative $k_r'$; the text at one point says road curvature need not be considered even though Eq. (1) explicitly contains curvature terms, so a sloppy implementation would not actually be using a true Frenet frame.","fun_headline_variants_meta":{"raw":{"variants":["Frenet state slashes DDPG lateral tracking error","Frenet coordinates improve DDPG path-tracking precision","DDPG with Frenet state tracks paths more precisely","Frenet coords reduce DDPG's lateral deviation","Frenet beats Cartesian for DDPG tracking"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001336,"raw_usage":{"total_tokens":5405,"prompt_tokens":890,"completion_tokens":4515,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":506,"completion_tokens_details":{"reasoning_tokens":4432}},"tokens_in":506,"tokens_out":4515,"duration_ms":29486,"temperature":1.0,"reasoning_tokens":4432,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:46:36.269192+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the comparison on a curved reference path using an independently verified Serret-Frenet transformation that includes the curvature terms of Eq. (1), and report the full distribution of lateral error over many episodes and seeds. If Frenet-state DDPG does not show a clear reduction in lateral error relative to Cartesian-state DDPG, or if dropping the curvature terms changes the result, the central claim is falsified.","supporting_citations":[{"cited_title":"Enhanced Credit Score Prediction Using Ensemble Deep Learning Model","cited_arxiv_id":"2410.00256","evidence_quote":"Provides the Frenet-coordinate frame in which the vehicle's Cartesian state is re-expressed relative to the reference path."},{"cited_title":"Long -term visual simultaneous localization and mapping: Using a bayesian persistence filter -based global map prediction[J]","cited_arxiv_id":null,"evidence_quote":"Provides the local-reference-frame trajectory-tracking background that motivates using Frenet coordinates."}],"review_version":1}