{"id":"11059008-767e-487f-a126-1848e525f629","arxiv_id":"2502.04696","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"A hierarchical controller using Bayesian optimization to tune drift equilibrium and look-ahead tracking parameters for an MPC drift controller improves clothoid path tracking in simulation.","lead":"This paper proposes a two-layer controller for autonomous drifting: a Bayesian optimizer learns three tuning parameters, and a model predictive controller keeps the car near learned drift equilibria to follow a curved path. The authors show in Carsim simulation that their tuned controller tracks a clothoid path with lower error than three baselines, even when the road friction used in the model is too high.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported drifting RMSE in Tables 4/5 compares each controller to its own planned drift states, and ALMPC's reference is generated online from tracking error, so lower drifting RMSE does not yet establish better drift control.","rationale":"In good faith, the paper's architecture is coherent and the tracking-error comparison is meaningful because the reference path is common across controllers; the 8-shaped maneuver is genuine out-of-sample evidence, though qualitative. The most load-bearing soft spot is not the BO in-sample issue per se, but the comparability of the drifting RMSE metric. The paper states that each simulation has a unique planned series of drift states, so the 'RMSE-Drifting' values are errors to controller-specific references. For ALMPC, the reference is partly produced by feedback from the tracking error itself, making the drifting metric self-referential. A controller can reduce its own drifting RMSE simply by planning references that track the realized trajectory. This directly undermines the drifting half of the central claim, independent of whether the clothoid test path is the same as the BO training path. The reader's in-sample concern is valid but secondary: the paper claims a strategy for a given clothoid path, and offline BO on that path is the intended mode of operation; the missing held-out quantitative test mainly limits generality claims. The proposed common-reference computation is a single, decisive check. If ALMPC still achieves the lowest drifting RMSE under a common reference, the claim stands; if not, the verdict should remain conditional and the paper should either correct the metric or narrow the claim to tracking performance only.","tokens_in":18535,"tokens_out":18475,"duration_ms":190564,"concrete_test":"Re-run the four controllers from Section 4.2 and recompute the RMSE-Drifting columns using one common reference for all controllers: for example, the nominal DEP (delta_eq = -0.52 rad with the corresponding V_eq, beta_eq, r_eq from the true mu) or the MPC-PPT reference at each time step. If ALMPC's advantage in the drifting columns shrinks or disappears, the claimed drifting improvement is an artifact of each controller being measured against its own online-generated reference rather than a common standard.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim includes reaching lower drifting RMSE than MPC-PPT, MPC-APT, and MPC-DEP. The RMSE-Drifting columns in Tables 4 and 5 are defined as the error between actual vehicle states and 'the planned drift states at each given time step.' The paper explicitly notes: 'Since each simulation features a unique planned series of drift states along the path...' Thus every controller is scored against a different, controller-specific reference. This matters especially for ALMPC: its reference is generated online by the APT control law. The desired drift radius follows Eq. (16), and the steering equilibrium is modified by the feedback law \\hat{\\delta}_{eq} = \\delta_{eq} + k e_{la} in Eq. (17). Consequently the desired V_eq, beta_eq, r_eq, and delta_eq used in the MPC cost (Eq. 10) vary with the current tracking error. A controller that lets its own reference chase the actual trajectory can trivially report a small error with respect to that reference; this measures internal command tracking, not the quality of the drift equilibrium or the drifting motion. The comparison is therefore self-referential for exactly the columns intended to support the 'drifting performance' half of the headline claim. The tracking RMSE (e, Delta-psi) is common to all controllers and is not affected by this problem, but the drifting RMSE advantage is not yet evidence for the claimed drifting superiority.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes an Adaptive Learning-based Model Predictive Control (ALMPC) strategy for autonomous drift vehicles. The architecture consists of an upper-level Bayesian optimization (BO) supervisor that learns three parameters—the drift equilibrium steering angle δ_eq, and the weights w_r, w_e of an adaptive path tracking (APT) law—and a lower-level MPC that tracks the resulting drift equilibrium states. The APT law adjusts the desired drift radius and steering angle based on look-ahead error. The method is tested in Matlab-Carsim simulations on a clothoid-based reference path under both precise and misidentified road friction parameters, with ablations MPC-PPT, MPC-APT, and MPC-DEP, plus an 8-shaped maneuver and comparisons with DRL, genetic algorithms, and particle swarm optimization. The central claim is that ALMPC achieves lower tracking and drifting RMSE than the baselines, and remains stable under parameter misidentification.","tokens_in":18855,"tokens_out":3932,"duration_ms":40655,"significance":"The problem of controlling a vehicle in sustained drift while tracking a general path is timely and practically relevant for autonomous racing and extreme-condition driving. The proposed hierarchical decomposition—separating path tracking (upper-level learning) from drift stabilization (lower-level MPC)—is a sensible design idea, and the use of BO to directly tune control-relevant parameters rather than refine a model aligns with the performance-driven identification philosophy. The paper is generally well organized, and the simulation platform (Matlab-Carsim) is appropriate for a first evaluation. The ablations that isolate APT learning and DEP learning are a useful contribution. However, the current evidence is weakened by two load-bearing issues: the headline tracking metrics are computed on the same path used to fit the BO objective, and the drifting RMSE is evaluated against controller-specific, online-generated references. If the authors address these with out-of-sample evaluations and a controller-independent drift metric, the work could become a solid contribution to learning-based control for autonomous drifting.","major_comments":[{"comment":"The BO objective J(θ) in Eq. (18) is evaluated over the entire clothoid path (N_k = 184 samples, T = 18.4 s) using exactly the same tracking metrics |e_k| and |Δψ_k| that are later reported as RMSE-Tracking in Tables 4 and 5. Consequently, the reported tracking improvements of ALMPC over the baselines are in-sample fits to the test path, not out-of-sample predictions. The only out-of-sample evidence, the 8-shaped maneuver in §4.3.2, is presented qualitatively with no quantitative error metrics, making it insufficient to support the generalization claim. I request a held-out path evaluation (e.g., a different clothoid or a curved track) with the BO-fixed parameters, along with reporting of tracking RMSE on that held-out path for all controllers.","section":"§4.2, Tables 4–5, Eq. (18)"},{"comment":"The RMSE-Drifting columns in Tables 4 and 5 measure the deviation of actual states from 'the planned drift states at each given time step,' and the text explicitly acknowledges that 'each simulation features a unique planned series of drift states along the path.' For ALMPC, the planned drift states are generated online by the APT law: the desired radius R_eq depends on the look-ahead error through Eq. (16), and the steering equilibrium δ̂_eq is modified by the feedback term k·e_la in Eq. (17). Thus the reference for the drifting RMSE changes with the tracking error, and a controller that lets its reference chase the actual trajectory can trivially reduce this RMSE. This makes the drifting-performance half of the headline claim self-referential. Please report drifting error against a fixed, controller-independent reference (e.g., a nominal DEP trajectory computed from the system model) or use absolute drift-quality indicators such as sustained large sideslip angles and consistent yaw rate, which do not depend on the controller's own reference.","section":"Tables 4–5, §4.2.1, Eqs. (16)–(17)"},{"comment":"The comparison with deep reinforcement learning (DRL) is not quantitatively substantive. The text claims a '10% less RMSE for the steering angle' but does not report tracking RMSE, drift-state RMSE, or any measures of variability. The training-time comparison (0h37min for BO-ALMPC vs 12h17min for DRL-ALMPC) is potentially confounded by different network architectures, hyperparameters, and hardware, and no details are given to assess the fairness of the comparison. This subsection should either be removed or expanded with a proper experimental setup (same vehicle model, same path, multiple seeds, and a table of performance metrics).","section":"§4.3.3, Fig. 14, Table 6"},{"comment":"All reported results come from single simulation runs. BO is stochastic (random initial samples, GP posterior sampling), and the RMSE differences in Tables 4 and 5 are small in some states (e.g., Δψ RMSE of 0.015 vs 0.018 rad in Table 4, and F_xr RMSE of 492.4 vs 489.1 N). Without repeated trials or confidence intervals, it is not possible to assess whether the observed improvements are statistically significant. Please report mean ± standard deviation over several independent BO runs (or at least over several initializations) for each controller and each case.","section":"§4.2, all simulation results"}],"minor_comments":[{"comment":"The entry 'δ0 −0.52 rad/s' appears to be a typo: δ0 is a steering angle, not a steering rate. It should be '−0.52 rad'.","section":"Table 2"},{"comment":"The parameters e_max and λ appear in the BO objective (Eq. (18)) and the barrier function (Eq. (19)) but are not listed in Table 2 or defined in the text. Their numerical values should be reported.","section":"Eqs. (18)–(20)"},{"comment":"The 'traditional MPC' used as a baseline in the 8-shaped maneuver is not defined. Please specify its cost function, prediction horizon, and parameter values, or refer explicitly to one of the earlier baselines (MPC-PPT, MPC-APT, or MPC-DEP).","section":"§4.3.2"},{"comment":"The text states that the 'best parameters are acquired in the 171 iterations' and Fig. 9 shows the performance cost. Please clarify whether the RMSE values in Tables 4 and 5 for ALMPC are obtained by re-simulating the final closed loop with the learned parameters from the best iteration, or are taken directly from that iteration's evaluation. This distinction is important for reproducibility.","section":"§4.2.1 and Figs. 7–8"},{"comment":"The DRL curve in Fig. 14 appears to be a single trajectory. Adding a shaded region over multiple seeds (or at least reporting the mean and variance of the lateral error over episodes) would make the comparison more convincing.","section":"§4.3.3, Fig. 14"},{"comment":"There are several typographical and wording issues: 'maximize lateral deviations' should be 'maximum lateral deviations'; 'undetermined disturbances' is likely 'unmodeled disturbances'; 'promise effective learning' should be 'ensure effective learning'. A careful language pass is recommended.","section":"Throughout"},{"comment":"Figure 6(b) is captioned only 'Drift Vehicle'; it would be clearer to describe what is being shown (e.g., the steering angle orientation relative to the path) in the caption or in the text.","section":"§4.2, Fig. 6"}],"recommendation":"major_revision","confidential_remarks":"The paper fits the scope of a robotics/control journal and does not show signs of misconduct. The core concern is that the principal quantitative claims are based on in-sample evaluation and a self-referential drifting metric; these are fixable with additional experiments rather than being fundamental flaws in the method. The paper would benefit from a stronger evaluation protocol, including held-out paths, repeated runs, and a controller-independent drift metric. If these issues are addressed, the work could be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the ALMPC combination is genuinely new as far as I know — BO learning the APT radius weights and the DEP steering offset, feeding a lower-level MPC — and the ablations are sensible. Worth a serious referee.\n\nWhat it does well: the hierarchy cleanly separates path tracking (upper APT) from drift-state regulation (lower MPC). The baseline comparisons (MPC-PPT, MPC-APT, MPC-DEP) are the right ones to isolate each learned component. The 8-shaped maneuver and the DRL comparison add independent, if qualitative, evidence. The internal logic is coherent and the paper is honest about its limitations in the discussion.\n\nWhere it's soft: the main quantitative claim rests on RMSEs on the same clothoid path that the BO objective was optimized on. That makes the tracking RMSE improvement partly a fitted result. I don't see repeated BO runs, seeds, or error bars, so I can't judge variance. Also, the drifting RMSE columns are computed against each controller's own planned drift states, which the paper says explicitly. Since ALMPC's reference is generated online from the tracking error, a controller that chases its own reference can report low drifting error without being better at drifting. That weakens the drifting half of the headline. The tracking RMSE (e and Δψ) is common to all controllers and not affected by this problem, so the tracking improvement is more credible.\n\nThe missing hyperparameters (λ, e_max, GP kernel length scale, etc.) make reproduction harder, but not impossible.\n\nWho it's for: readers working on autonomous drifting, aggressive driving, and performance-driven MPC tuning. It'll be useful as a reference for the architecture. Not a methods breakthrough, but a solid engineering contribution.\n\nRecommendation: send it to review. Require held-out path evaluation, repeated BO runs with variance reporting, and a fix to the drifting-RMSE comparison (e.g., report drifting error against a common reference or against the actual DEP). If those are handled, the paper is publishable.","headline":"Useful hierarchical drift-control paper with a genuinely novel combination and sensible ablations, but the headline RMSE comparisons are partly in-sample and the drifting-RMSE metric is self-referential.","tokens_in":19376,"tokens_out":1360,"would_cite":false,"duration_ms":13983,"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 learning-based MPC supervisor can keep a drifting vehicle on a curved path even when the friction model is wrong.","keywords":["learning-based control","autonomous drifting","path tracking","Bayesian optimization","model predictive control","drift equilibrium points","clothoid path"],"falsifier":"Run the learned parameter set (e.g., $\\delta_{eq}=-0.471$ rad, $w_r=1.903$, $w_e=1.032$ for the misidentified-friction case) on a different reference path, such as an 8-shaped course or a clothoid with reversed curvature, without re-running BO, and compare lateral-error RMSE against the MPC-PPT baseline; if ALMPC does not maintain a comparable margin there, the generalization claim is not supported.","tokens_in":18326,"feed_emoji":"🏎️","tokens_out":5776,"duration_ms":53030,"temperature":0.7,"pith_summary":"This paper claims that autonomous drift control can be made both path-accurate and model-robust by giving a lower-level MPC controller a drifting objective and letting an upper-level Bayesian optimizer tune a small set of path-tracking parameters. The optimizer learns three numbers: the drift-equilibrium steering angle and two weights in an adaptive path-tracking law that adjusts the desired drift radius and the equilibrium steering angle from a look-ahead error. Because the upper layer optimizes the parameters directly against closed-loop tracking cost, it can compensate for a misidentified road-friction coefficient without refining the vehicle model. In Matlab-Carsim simulations on a clothoid-based reference path, the resulting ALMPC controller reports lower RMSE for lateral tracking, course error, and most drift states than three MPC baselines, in both the precise-parameter and the 10% friction-error cases. The central payoff would be a drift controller whose tracking performance survives model mismatch, at the cost of an offline Bayesian-optimization session over the reference path.","feed_headline":"Bayesian tuning keeps drift cars glued to a curved path","feed_subtitle":"Simulations show lower tracking and drift error than three baselines, even when road friction is misidentified.","key_machinery":"The load-bearing object is the triple $\\theta = [\\delta_{eq}, w_r, w_e]$ learned by Bayesian optimization: $\\delta_{eq}$ is the drift-equilibrium steering angle that shifts all other equilibrium states computed from the linearized single-track model, and $w_r, w_e$ are the two gains of the adaptive path-tracking law $R_{eq} = w_r R_r + w_e e_{la}$. A feedback term $\\hat{\\delta}_{eq} = \\delta_{eq} + k e_{la}$ further adjusts the equilibrium steering angle from the look-ahead error. The upper-level BO treats closed-loop tracking performance over the entire clothoid path, measured by $J(\\theta)$ in Eq. 18, as an expensive black-box function, models it with a Gaussian process (Matérn 5/2 kernel), and selects new parameter sets by expected improvement. These learned parameters are then frozen and passed to the lower-level MPC, whose quadratic program keeps the vehicle states near the resulting drift equilibrium while respecting input and input-rate limits. The mechanism does the work of replacing both the usual predictive path-tracking loop and the requirement of an accurately identified vehicle model.","core_discovery":"On the paper's own terms, the discovery is that the conflicting goals of keeping a drift vehicle near its unstable drift equilibrium and following a general curved path can be separated in a hierarchy: a lower-level MPC tracks the current drift equilibrium point, while an upper-level BO supervisor selects the equilibrium steering angle $\\delta_{eq}$ and the weights $(w_r, w_e)$ of an adaptive path-tracking (APT) law. The APT law sets the desired drift radius as $R_{eq} = w_r R_r + w_e e_{la}$, where $R_r$ is the curvature radius of the reference path and $e_{la}$ is a look-ahead lateral error, and shifts the equilibrium steering angle by $\\hat{\\delta}_{eq} = \\delta_{eq} + k e_{la}$. BO minimizes a closed-loop tracking cost $J(\\theta)$ built from lateral error, course error, a soft barrier on large deviations, and the increment of lateral error, using a Gaussian-process surrogate with a Matérn-5/2 kernel and expected-improvement acquisition. The paper reports that with both APT and DEP learning (ALMPC), the vehicle follows the clothoid path with lower tracking RMSE and steadier drift states than the baselines MPC-PPT, MPC-APT, and MPC-DEP, and that the advantage grows when the road friction in Carsim is set 10% lower than the value used to derive the model.","pith_inferences":["The reported RMSE comparisons are computed on the same clothoid path used to evaluate the BO objective, so the headline improvements are in-sample fits; a direct test would be re-running the learned parameters on a different path (for example, the 8-shaped maneuver) and reporting numerical tracking errors there.","One could plausibly extend the scheme to online or periodic re-tuning by sliding the BO objective window along the path, which would make the controller adaptive to time-varying friction instead of only to a single misidentified value.","The BO search space only includes $\\delta_{eq}, w_r, w_e$; learning additional equilibrium states, such as $\\beta_{eq}$ or $V_{eq}$, could either improve performance or destabilize the lower-level MPC, and the paper's choice to keep them model-derived is a safety-motivated restriction that future work could test.","The comparison with DRL suggests that sample-efficient BO may be the more practical route for expensive closed-loop drift tuning, but that comparison is single-scenario and would need repeated runs to confirm."],"forward_implications":["If the central claim is right, drift-vehicle path tracking can be implemented with a one-time BO tuning session instead of continuously solving a prediction-based path-tracking optimization at every control step.","The same performance-driven identification should compensate for modeling errors in other sensitive parameters, such as tire stiffness or mass distribution, not only road friction, as long as they shift the DEP.","The reported ALMPC gains imply that tuning the controller reference (the DEP) and the tracking law jointly is more effective than tuning either one alone, which is the comparison the paper makes with MPC-APT and MPC-DEP.","Because the learned parameters are re-usable across the path used for tuning, the approach is compatible with low-computational hardware at runtime; the paper reports an online execution time of 9.32 s for the BO-based controller."],"supporting_citations":[{"why":"Supplies the friction-circle rear tire force model of Eq. 7, which couples longitudinal and lateral rear forces in the drift vehicle dynamics.","marker":"[32]"},{"why":"Establishes drift equilibrium points through equilibrium analysis of the drifting vehicle model, providing the DEP concept the lower MPC tracks.","marker":"[5]"},{"why":"Provides the steady-state drifting stabilization analysis that underlies the existence and use of DEP for drift control.","marker":"[4]"},{"why":"Presents a prior combined drifting-state and trajectory-tracking MPC, representing the prediction-based approach whose computational burden ALMPC aims to release.","marker":"[13]"},{"why":"Introduces the prediction-based path tracking (PPT) control method using MPC, which serves as the baseline that APT replaces.","marker":"[35]"},{"why":"Motivates performance-oriented model learning, the rationale for identifying best-performing parameters directly rather than refining the vehicle model.","marker":"[22]"},{"why":"Supplies the Bayesian optimization foundation, including exploration-exploitation balance and the GP/EI machinery used by the upper-level supervisor.","marker":"[26]"},{"why":"Provides the efficient global optimization framework and the expected-improvement acquisition function formula used in Eq. 23.","marker":"[38]"}],"fun_headline_variants":["Hierarchical MPC learns drift to follow curved paths","Self-tuning drift control with Bayesian optimization","Adaptive learning MPC for stable drifting on curves","BO-tuned MPC keeps drift cars on track"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Bayesian optimizer's cost function, evaluated over the full 18.4-second clothoid path, is identical to the path and metric used in the reported RMSE tables, so the headline gains are in-sample fits; if the learned settings are transferred to a new path, the reported margins may shrink or disappear.","fun_headline_variants_meta":{"raw":{"variants":["Hierarchical MPC learns drift to follow curved paths","Self-tuning drift control with Bayesian optimization","Adaptive learning MPC for stable drifting on curves","BO-tuned MPC keeps drift cars on track"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000283,"raw_usage":{"total_tokens":1726,"prompt_tokens":1055,"completion_tokens":671,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":671,"completion_tokens_details":{"reasoning_tokens":614}},"tokens_in":671,"tokens_out":671,"duration_ms":6788,"temperature":1.0,"reasoning_tokens":614,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-08T21:49:19.744262+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the learned parameter set (e.g., $\\delta_{eq}=-0.471$ rad, $w_r=1.903$, $w_e=1.032$ for the misidentified-friction case) on a different reference path, such as an 8-shaped course or a clothoid with reversed curvature, without re-running BO, and compare lateral-error RMSE against the MPC-PPT baseline; if ALMPC does not maintain a comparable margin there, the generalization claim is not supported.","supporting_citations":[{"cited_title":"Goh, Tushar Goel, and J","cited_arxiv_id":null,"evidence_quote":"Supplies the friction-circle rear tire force model of Eq. 7, which couples longitudinal and lateral rear forces in the drift vehicle dynamics."},{"cited_title":"Hindiyeh and J","cited_arxiv_id":null,"evidence_quote":"Establishes drift equilibrium points through equilibrium analysis of the drifting vehicle model, providing the DEP concept the lower MPC tracks."},{"cited_title":"Steady-state drifting stabilization of rwd vehicles","cited_arxiv_id":null,"evidence_quote":"Provides the steady-state drifting stabilization analysis that underlies the existence and use of DEP for drift control."},{"cited_title":"Combined fast control of drifting state and tra- jectory tracking for autonomous vehicles based on mpc controller","cited_arxiv_id":null,"evidence_quote":"Presents a prior combined drifting-state and trajectory-tracking MPC, representing the prediction-based approach whose computational burden ALMPC aims to release."},{"cited_title":"An mpc-based controller framework for agile maneu- vering of autonomous vehicles","cited_arxiv_id":null,"evidence_quote":"Introduces the prediction-based path tracking (PPT) control method using MPC, which serves as the baseline that APT replaces."},{"cited_title":"Performance-oriented model learning for data-driven mpc design","cited_arxiv_id":null,"evidence_quote":"Motivates performance-oriented model learning, the rationale for identifying best-performing parameters directly rather than refining the vehicle model."},{"cited_title":"Adams, and Nando de Freitas","cited_arxiv_id":null,"evidence_quote":"Supplies the Bayesian optimization foundation, including exploration-exploitation balance and the GP/EI machinery used by the upper-level supervisor."},{"cited_title":"Efficient global optimization of expensive black-box functions","cited_arxiv_id":null,"evidence_quote":"Provides the efficient global optimization framework and the expected-improvement acquisition function formula used in Eq. 23."}],"review_version":1}