{"id":"33f233d5-0da8-423a-8434-b27450775bd3","arxiv_id":"2507.09714","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":12,"one_line_summary":"A unified iLQR-based racing controller that blends historical lap data with soft obstacle-avoidance penalties overtakes more simulated opponents than LMPC baselines at lower compute.","lead":"This paper proposes IteraOptiRacing, a single planning-and-control algorithm for autonomous race cars that reuses data from previous laps to overtake multiple moving opponents. It combines iterative LQR optimization with soft obstacle-avoidance penalties and reports faster overtaking and lower computation than three LMPC baselines in simulation.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Overtaking and collision-free claims are validated only against pre-generated, non-interactive opponent trajectories; Eq. (10) and the selection logic in §III-D assume those predictions are correct, so the racing claim does not yet cover reactive competitors.","rationale":"Reader's weakest assumption correctly identifies the non-interactive, known opponent trajectories as the key gap between the demonstrated system and the advertised 'racing' capability. My pass did not find an internal inconsistency in the optimization or a fatal flaw in the implementation; for the stated simulation setup, the reported numbers are a reasonable first validation and the comparison with three LMPC variants is informative. The concern is about external validity: the central claim's scope ('competing with other racing cars') exceeds the evidence (scripted moving obstacles). The paper itself flags the limitation, which is why this is a conditional acceptance issue rather than a rejection. I would not change the reader's CONDITIONAL verdict; I would ask the authors to either soften the headline claims or add a reactive-opponent experiment. The concrete test above is the minimal check that would settle whether the concern lands.","tokens_in":19415,"tokens_out":6059,"duration_ms":74493,"concrete_test":"Re-run the 100 randomized M-shape scenarios (same speeds, same initial positions) with surrounding vehicles switched from pre-generated open-loop trajectories to closed-loop PID agents whose target lateral deviation depends on the ego's position, e.g., d_target(t) = d_random(t) + c·(sc - sc_p) when |sc - sc_p| < 5 m, while (15) still randomizes velocity. Record overtaking success and collisions for IteraOptiRacing and Baselines 1–3; if the success-rate gap or the 'no failures' result changes materially, the conclusion depends on the non-interaction assumption laid out in §IV-A.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV-A and Remark 5 make the load-bearing assumption explicit: all surrounding vehicles follow pre-generated deterministic trajectories, known to the ego and unaffected by it. Every safety and selection mechanism—the obstacle penalty (8)–(9), the collision check (10), and the best-trajectory selection in §III-D—uses the predicted opponent states x_{p,t+k|t}. This is fine for obstacle avoidance against scripted moving obstacles, but the abstract and contributions claim 'competing with other racing cars' and 'collision-free' overtaking of multiple moving vehicles. Real racing opponents observe the ego and react (blocking, defending, changing lanes), so the predicted x_{p,t+k|t} used in (10) will be wrong at execution time and the selected 'collision-free' trajectory can become unsafe. The randomized PID opponents in §IV-A do not close this loop: their target velocities and lateral deviations are independent of the ego's state. The paper acknowledges the limitation in §IV-A and §V, but the headline claims are stated without this caveat. Therefore the central empirical claim—that IteraOptiRacing overtakes multiple competing vehicles more quickly and collision-free—is not yet supported for interactive opponents, and the 'time-optimal' claim is likewise only relative to these scripted scenarios.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes IteraOptiRacing, a unified planning and control framework for autonomous racing built on i2LQR. At each time step, K nearest historical states are used as candidate terminal targets; for each candidate, an iLQR problem with affine time-varying dynamics minimizes a quadratic terminal tracking cost plus control-effort and control-rate penalties, with obstacle avoidance encoded as an exponential soft constraint and weights adaptively adjusted until a discrete collision-check condition is met. A best trajectory is then selected by collision, reachability, convergence, and cost-to-go criteria. The method is evaluated in a high-fidelity simulator against three LMPC-based baselines in randomized scenarios with multiple pre-generated PID-driven opponents, reporting higher overtaking success and lower computation time.","tokens_in":19811,"tokens_out":6319,"duration_ms":72494,"significance":"If the central claims held, this would be a useful real-time, parallelizable planner-controller that avoids the nonsmooth planner-controller switching of prior work and handles multiple moving obstacles while improving lap time over iterations. The empirical comparison is fairly extensive: 300 randomized tests across speed ranges plus three track shapes, with consistent improvements in overtaking counts and roughly an order-of-magnitude lower mean solve time compared with the three LMPC variants. The authors should be credited for evaluating against three distinct baselines, reporting hyperparameters in Table II, and making the non-interactive assumption explicit in Section IV-A. However, the headline claims of 'collision-free and time-optimal generated trajectories' and of overtaking 'other racing cars' are not established: the cost function does not minimize time, no safety certificate is provided, and all experiments use deterministic, non-reactive opponent trajectories.","major_comments":[{"comment":"The claim that IteraOptiRacing produces time-optimal trajectories is not supported by the problem formulation. The stage cost in Eq. (5) penalizes control effort and control rate and the terminal cost in Eq. (6) penalizes quadratic deviation from a historical target zg; neither term is a time cost or a cost-to-go. The selection rule in Section III-A2 chooses the reachable candidate with the smallest cost-to-go h(zg), but because condition (12) only requires the terminal state to be within epsilon of zg, the stored cost-to-go of zg is not a valid bound on the remaining time from the actually reached terminal state. The LMPC monotonic-improvement argument in Appendix A relies on a convex-hull terminal constraint and a convex cost-to-go, neither of which is present in Eqs. (4)-(6). Remark 1 defers the obstacle-free equivalence to prior work [51], but that result is not reproduced here and cannot substitute for a derivation.","section":"III-A2, III-C1, Remark 1"},{"comment":"The empirical support for the collision-free overtaking claim is limited to non-interactive, pre-generated opponent trajectories. Section IV-A states that all surrounding-vehicle trajectories are pre-generated and stored, that the vehicles are controlled by PID with target velocities and lateral deviations independent of the ego state, and that 'this work does not consider any interaction between the ego vehicle and other surrounding vehicles, such as Stackelberg games'; Remark 5 reaffirms that the safe-boundary checks in Eq. (10) rely on these deterministic predictions. Real racing opponents react to the ego car, so the predicted states x_{p,t+k|t} used in Eqs. (8)-(11) will generally be wrong, and a trajectory selected as collision-free at planning time can become unsafe at execution time. The abstract and contributions nevertheless claim collision-free overtaking of 'other racing cars' without this caveat; this overstatement needs to be removed or supported by experiments with reactive opponents or by a robustness guarantee against prediction error.","section":"IV-A, Remark 5"},{"comment":"Even with perfect predictions, the 'collision-free' property is not formally guaranteed. Obstacle avoidance is encoded as a soft exponential barrier in Eqs. (7)-(9), and Remark 2 states that weights are adjusted until condition (10) is satisfied 'or the maximum iteration number is reached'; the manuscript does not specify what happens if the maximum is reached and (10) remains violated, nor does it prove that the iterative weight adjustment converges to a trajectory satisfying (10). In addition, Eq. (10) is checked at discrete prediction nodes, so absence of collisions between nodes is not certified. Section III-D selects only among trajectories that pass the check but does not define a fallback if no candidate qualifies. A rigorous claim of 'collision-free generated trajectories' requires either a hard constraint, a certificate, or an explicit failure-handling policy.","section":"Remark 2, Eq. (10), Section III-D"}],"minor_comments":[{"comment":"The layout of the four success-rate categories in Table III is confusing: the superscripts are only placed on the first number of each block, making it difficult to map percentages to the (a)-(d) categories; please format each category as a separate labeled cell or row.","section":"Table III"},{"comment":"Equation (11) uses both vx,i and vx,p for the same quantity; unify the notation for the p-th surrounding vehicle's speed.","section":"Eq. (11)"},{"comment":"Equation (12) compares a squared norm to epsilon, while the text describes epsilon as a tracking ratio; the inequality should be written consistently, e.g., as a norm bound on the error rather than a squared norm compared to a non-squared threshold.","section":"Eq. (12)"},{"comment":"The values of epsilon and gamma in Table II are given without units or a sensitivity analysis, although the overtaking-range criterion (11) depends directly on them; a brief discussion of their influence on the results would improve reproducibility.","section":"Table II and Eq. (11)"},{"comment":"There is an internal inconsistency in the target-selection ordering: Section III-A2 says candidates are selected 'with cost-to-go values prioritized in descending order,' while Section III-B states that points with smaller cost-to-go are used first; the intended ordering should be stated consistently.","section":"III-A2 and III-B"},{"comment":"References [4] and [42] are the same paper ('A benchmark comparison of imitation learning-based control policies for autonomous racing') and should be merged into a single citation.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"The paper is a reasonable engineering contribution if reframed as deterministic multi-obstacle avoidance in a known simulation environment. The main risk is that the abstract and contributions promise more than the assumptions deliver, particularly regarding time optimality and collision-free behavior against interactive opponents. I would ask the authors to either add reactive-opponent experiments or explicitly rescope all such claims, and to replace the appeal to [51] with a direct derivation or reproduction of the obstacle-free equivalence."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Plain English: this is a real empirical contribution with a narrow scope, and the paper's own disclaimers in the body are more honest than the abstract. The new algorithmic part is modest—an adaptive weight scheme and a trajectory selection rule on top of the authors' i2LQR—but the evaluation is the main event: 300 randomized scenarios across three track shapes, three speed ranges, nine obstacle vehicles, against three LMPC-based baselines. The results are consistent: IteraOptiRacing overtakes more scripted opponents and does it at roughly 28 Hz on a desktop CPU while the baselines run at 2-3 Hz. That is useful, concrete progress.\n\nWhere it gets soft: the words 'collision-free' and 'time-optimal' in the abstract are not backed by any derivation. The cost function in (4a) is a quadratic tracking cost plus control effort; time optimality is supposed to emerge from selecting target states with low cost-to-go, but there is no proof that this works in obstacle environments, only the empirical comparison. Collision-free is conditional on the predicted opponent trajectories being correct; the paper explicitly says those are pre-generated and deterministic (Remark 5, Section IV-A). The baselines share this assumption, so the comparison is fair, but the contribution claim 'competing with other racing cars' overstates what is tested—these are scripted moving obstacles, not reactive competitors. To the paper's credit, the limitation is acknowledged in the body and conclusion, just not in the headline claims.\n\nThe bigger technical gap is that the adaptive weighting and selection scheme is never compared against the plain i2LQR from the authors' CDC paper. Remark 1 cites that paper for equivalence to LMPC in obstacle-free conditions, but nothing here reproduces it. Without that comparison, I can't tell how much of the gain comes from the new pieces versus i2LQR itself. Also, the hyperparameters are numerous; they're fixed across all tests, which is good, but there's no sensitivity analysis, so I'd worry about overfitting to the specific randomized scenario distribution.\n\nBottom line: the empirical claim as scoped—overtaking pre-programmed vehicles faster than LMPC variants—is supported. The broader claims are not. This deserves a serious referee; it's a competent systems paper, and the flaws are fixable with a revised abstract, an i2LQR baseline, and ideally one reactive-opponent case. I would not cite it in my own work in the next year, but I would read a revision.","headline":"Solid but narrow empirical study; the abstract overpromises what the evidence actually supports.","tokens_in":20338,"tokens_out":3225,"would_cite":false,"duration_ms":36363,"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 single iterative controller, IteraOptiRacing, claims to plan overtakes and optimize lap time simultaneously in real time.","keywords":["autonomous racing","iterative linear quadratic regulator","i2LQR","model predictive control","overtaking","obstacle avoidance","lap time optimization","real-time planning and control"],"falsifier":"Run the same randomized overtaking tests on the L-, M-, and elliptic tracks with opposing vehicles whose velocity and lateral deviation are updated based on the ego car's current position (for example, actively moving to block), and compare one-lap overtake success rates; if IteraOptiRacing no longer beats the three LMPC baselines, the paper's claim fails in reactive traffic.","tokens_in":19232,"feed_emoji":"🏎️","tokens_out":5644,"duration_ms":60111,"temperature":0.7,"pith_summary":"IteraOptiRacing is an attempt to collapse the two jobs of autonomous racing—following a fast line and overtaking rivals—into one optimization loop that runs in real time. The paper argues that by iterating on the ego car's own historical lap data, a single i2LQR controller can both improve lap time and avoid multiple moving cars, eliminating the abrupt switch between a time-optimal planner and an overtaking replanner seen in earlier methods. In high-fidelity simulations with nine randomized opponents on three tracks, it reports overtaking more cars, never failing where a baseline succeeds, and computing overtaking solutions in about 0.036 seconds on average, against 0.309–0.485 seconds for the three LMPC baselines. The value of the claim, if true, is that competitive racing does not need separate planning and control layers or expensive game-theoretic reasoning to handle multi-car traffic.","feed_headline":"One controller overtakes rivals and cuts lap time in real time","feed_subtitle":"Iterative optimal control with historical lap data unifies planning and control, beating three LMPC baselines in simulations.","key_machinery":"The load-bearing object is the target terminal set: at each time step, K-nearest past states drawn from the ego car's stored historical laps become candidate terminal targets for K separate iLQR problems. Obstacle avoidance is inserted into the stage cost as an exponential soft constraint whose weight is increased, and whose terminal-tracking weights are divided down, whenever the open-loop trajectory violates the safety boundary (10). An affine time-varying dynamics model linearized around the previous open-loop prediction makes each candidate problem cheap, and independence of the K problems makes them parallelizable. This single optimization loop is what replaces the planner/controller switch.","core_discovery":"The paper's central claim is that IteraOptiRacing, built on the Iterative Linear Quadratic Regulator for Iterative Tasks (i2LQR), unifies the planner and the controller so that one local optimization produces trajectories that are simultaneously time-optimal on the basis of historical laps and collision-free against multiple moving vehicles. At each time step the algorithm builds a target terminal set from the K-nearest historical states, solves one independent iLQR problem per candidate target, encodes obstacle avoidance as a soft exponential barrier inside the cost, and dynamically relaxes terminal-tracking weights whenever the produced trajectory violates the safety boundary. The best trajectory is then selected by collision and reachability checks using the minimal cost-to-go. The paper validates this design in randomized simulations and reports that it overtakes more of the nine surrounding vehicles than LMPC with local replanning, LMPC with slacked target state, and LMPC with slack on the convex hull, at a consistently lower computational cost.","pith_inferences":["Editorial inference: because the K candidate iLQR problems are independent, the same history-driven terminal-set mechanism could be applied to reactive opponents by substituting an online predictor for the pre-generated trajectories, a test the paper does not run.","Editorial inference: the adaptive weight schedule exposes a tunable trade-off between safety margin and lap-time optimality that could be mapped explicitly as a Pareto frontier rather than fixed hyperparameters.","Editorial inference: the framework needs only a stored history set and a differentiable dynamics model, so it could transfer to other iterative multi-agent tasks such as drone racing or warehouse vehicle coordination, though the paper only demonstrates car racing."],"forward_implications":["No separate overtaking layer is needed: the same optimization that improves lap time also generates collision-free overtaking trajectories around multiple moving cars.","The K nearest-state candidate problems are solved independently, so parallel computing directly reduces wall-clock time, supporting the reported mean solve time of 0.036 seconds during overtaking.","Because obstacles are handled as soft constraints that relax terminal tracking only when safety is at risk, overtaking maneuvers can be generated without deadlock or constraint infeasibility.","In all reported randomized tests, whenever any baseline overtook all nine cars, IteraOptiRacing also did; and there were cases where only IteraOptiRacing succeeded, so the paper claims superiority over the baselines in these settings."],"supporting_citations":[{"why":"Supplies the i2LQR algorithm that IteraOptiRacing extends from iterative tasks to racing with dynamic obstacles.","marker":"[51]"},{"why":"Provides the learning-based MPC formulation, the affine time-varying dynamics model, and the historical cost-to-go idea both the algorithm and baselines build on.","marker":"[19]"},{"why":"Defines the LMPC with local replanning baseline and the nonsmooth planner-controller switch the paper claims to avoid.","marker":"[28]"},{"why":"Supplies the two relaxed-LMPC baselines (slacked target state and slack on convex hull) used for comparison.","marker":"[54]"},{"why":"Provides the constrained iLQR exponential barrier method used to turn state and obstacle constraints into cost terms.","marker":"[49]"},{"why":"Establishes the original LMPC framework and its iterative lap-time improvement guarantee that the paper adopts for cost-to-go.","marker":"[52]"},{"why":"Supplies the dynamic bicycle model and Pacejka tire model used in the simulator and algorithm.","marker":"[53]"}],"fun_headline_variants":["One iteration-based controller plans, avoids, and races faster","Historical laps guide a real-time unified racing controller","i2LQR unifies planning and control to beat LMPC baselines","Real-time racing controller uses past laps for optimal overtakes","Unified iLQR framework outperforms LMPC in racing sims"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the surrounding vehicles' future trajectories are pre-generated and known in advance, and that they do not react to the ego car's maneuvers.","fun_headline_variants_meta":{"raw":{"variants":["One iteration-based controller plans, avoids, and races faster","Historical laps guide a real-time unified racing controller","i2LQR unifies planning and control to beat LMPC baselines","Real-time racing controller uses past laps for optimal overtakes","Unified iLQR framework outperforms LMPC in racing sims"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000174,"raw_usage":{"total_tokens":1260,"prompt_tokens":900,"completion_tokens":360,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":516,"completion_tokens_details":{"reasoning_tokens":272}},"tokens_in":516,"tokens_out":360,"duration_ms":4059,"temperature":1.0,"reasoning_tokens":272,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T17:50:08.951814+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same randomized overtaking tests on the L-, M-, and elliptic tracks with opposing vehicles whose velocity and lateral deviation are updated based on the ego car's current position (for example, actively moving to block), and compare one-lap overtake success rates; if IteraOptiRacing no longer beats the three LMPC baselines, the paper's claim fails in reactive traffic.","supporting_citations":[{"cited_title":"i2lqr: Iterative lqr for iterative tasks in dynamic environments,","cited_arxiv_id":null,"evidence_quote":"Supplies the i2LQR algorithm that IteraOptiRacing extends from iterative tasks to racing with dynamic obstacles."},{"cited_title":"Learning how to autonomously race a car: a predictive control approach,","cited_arxiv_id":null,"evidence_quote":"Provides the learning-based MPC formulation, the affine time-varying dynamics model, and the historical cost-to-go idea both the algorithm and baselines build on."},{"cited_title":"Autonomous racing with multiple vehicles using a parallelized optimization with safety guarantee using control barrier functions,","cited_arxiv_id":null,"evidence_quote":"Defines the LMPC with local replanning baseline and the nonsmooth planner-controller switch the paper claims to avoid."},{"cited_title":"Minimum time learning model predictive control,","cited_arxiv_id":null,"evidence_quote":"Supplies the two relaxed-LMPC baselines (slacked target state and slack on convex hull) used for comparison."},{"cited_title":"Constrained iterative lqr for on-road autonomous driving motion planning,","cited_arxiv_id":null,"evidence_quote":"Provides the constrained iLQR exponential barrier method used to turn state and obstacle constraints into cost terms."},{"cited_title":"Learning model predictive control for iterative tasks. a data-driven control framework,","cited_arxiv_id":null,"evidence_quote":"Establishes the original LMPC framework and its iterative lap-time improvement guarantee that the paper adopts for cost-to-go."},{"cited_title":"Rajamani, V ehicle dynamics and control","cited_arxiv_id":null,"evidence_quote":"Supplies the dynamic bicycle model and Pacejka tire model used in the simulator and algorithm."}],"review_version":1}