{"id":"3ad339da-893a-423c-b44b-aa2e6cec04f4","arxiv_id":"2411.11762","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A hybrid TD3-MPC control framework enabled a full-scale consumer electric vehicle to autonomously execute drift U-turns and right-angle turns.","lead":"The authors train a TD3 reinforcement learning controller for drifting through corners in simulation, then deploy it on a consumer electric vehicle using a model predictive controller as a tracking and safety layer. Real-vehicle tests show the car can complete drift U-turns and right-angle turns with sustained high side-slip angles.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Trained RL policy's simulation times exactly match the hand-crafted Bézier pre-trajectory times (2.81 s and 4.52 s), indicating the RL contributes no time optimization; this undermines the central 'RL-based transient drift cornering' claim.","rationale":"The reader's weakest_assumption concerned the optimality of the preview trajectory and the ability of the linear MPC to correct in the nonlinear drift regime. Our stress-test finds a sharper, more concrete instance of the same problem: the paper's own tables show that the TD3 policy in simulation attains cornering times identical to the Bézier pre-trajectory times for both reported scenarios. This exact equality (2.81 s for the 90° turn, 4.52 s for the U-turn) strongly suggests the RL controller did not optimize beyond the hand-crafted path, so the 'preview trajectory' is effectively the Bézier trajectory. That undercuts the central claim that the method performs RL-based trajectory optimization for minimal cornering time. It also makes the real-vehicle demonstration less surprising: the vehicle is tracking a pre-planned line with MPC correction, with the RL policy acting as a feedforward generator that reproduces the planned maneuver. We do not accuse the authors of any misconduct; the concern is internal to the presented numbers and can be settled by additional experiments or by releasing the trajectory data. The reader's CONDITIONAL verdict remains appropriate: the paper should demonstrate that the RL policy improves over the Bézier baseline or temper the claim accordingly. We did not find a separate overriding flaw in the MPC linearity argument because the real-vehicle success, if reproducible, would support the fusion concept; however, the apparent absence of genuine RL optimization is the more load-bearing issue for the paper's stated contribution. A secondary concern, the unsupported 'first deployment' claim relative to reference [19], also needs reconciliation but is secondary to the technical point above.","tokens_in":15357,"tokens_out":8392,"duration_ms":86517,"concrete_test":"Re-run the TD3 training for the 135° turn (present in Table I but absent from Tables III and IV) and report the final simulated cornering time. If it equals 4.09 s (the Bézier time from Table I), the pattern is confirmed. Additionally, release the preview trajectory T_p for the U-turn and compute its maximum lateral deviation from the Bézier pre-trajectory; if the deviation is below 0.2 m and the speed profile matches within 1 m/s, the RL policy is simply tracking the hand-crafted path rather than performing independent trajectory optimization.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that the TD3-based controller performs transient drift cornering aimed at optimizing cornering time. Yet the paper's own data show that the trained RL policy in simulation achieves exactly the same cornering time as the hand-crafted Bézier pre-trajectory: Table I lists 2.81 s for the 90° turn and 4.52 s for the U-turn under Bézier optimization; Tables III and IV list the RL policy in simulation at exactly 2.81 s and 4.52 s for the same scenarios. This equality to two decimal places indicates that the RL optimization did not find any faster line than the pre-optimized path; the policy appears to have collapsed to tracking the Bézier trajectory at the planned speed. Consequently, the preview trajectory T_p generated by simulating the policy (Eq. 18) is not an independently optimized drift trajectory but a re-issuance of the Bézier path. If so, the claimed contribution—RL-based transient drift cornering that optimizes cornering time—is not demonstrated; the real-vehicle success reduces to tracking a pre-planned path with a fused feedforward/corrective controller, and the novelty claim ('first deployment of scenario-to-action RL-based transient drift cornering') is substantially weakened. The paper needs to show that the RL policy produces a trajectory strictly faster (or otherwise better) than the Bézier baseline, and provide reward weights and trajectory data to rule out a reward-shaping artifact.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper presents a control framework for high-speed drift cornering of autonomous electric vehicles, combining a TD3 reinforcement-learning (RL) controller trained in simulation with a two-step linear-tire MPC that provides corrective inputs during real-vehicle deployment. A Bézier-based pre-trajectory optimizer shapes the RL reward and generates an initial speed plan. In deployment, the trained policy is simulated on a matched virtual curve to produce a preview trajectory, which the MPC then tracks while the RL actions serve as feedforward inputs. The approach is validated on a consumer-grade electric vehicle performing a drift U-turn and a right-angle turn on wet road, with maximum speeds of 10.4 m/s and 10.6 m/s and maximum side-slip angles of 63.7° and 46.1°, respectively. The authors claim this is the first deployment of a scenario-to-action RL-based transient drift cornering algorithm on a consumer electric vehicle.","tokens_in":15645,"tokens_out":4082,"duration_ms":39742,"significance":"If the central claims hold, the paper makes a notable empirical contribution: it demonstrates that an RL-trained drift controller can be transferred to a full-scale consumer EV with the aid of an MPC wrapper, and it provides real-vehicle data and video evidence. The comparison against RL-alone and MPC-alone baselines is a useful design study. However, the significance is weakened by two issues: (i) the trained RL policy's simulation times exactly match the Bézier pre-trajectory times, suggesting that the RL does not in fact optimize cornering time as claimed; and (ii) the preview trajectory is generated by the policy itself, making the deployment evaluation to some extent self-referential. The paper would be strengthened by showing that RL produces a strictly faster or qualitatively different trajectory than the pre-optimized one, and by providing repeated trials and standard statistical measures.","major_comments":[{"comment":"The reported simulation cornering times of the trained TD3 policy (2.81 s for the 90° turn in Table IV, 4.52 s for the U-turn in Table III) are exactly equal, to two decimal places, to the Bézier pre-trajectory times in Table I. This equality strongly suggests that the RL policy has collapsed to tracking the pre-optimized Bézier path at the planned speed, rather than discovering a time-optimal transient drift line. If this is the case, the central contribution \"TD3-based controller ... capable of performing transient drift cornering aimed at optimizing cornering time\" is not demonstrated, and the preview trajectory T_p in Eq. (18) is not an independent optimum but a re-issuance of the Bézier trajectory. Please provide the reward weights (k_pl, k_pv, k_s, k_s1, k_t1, k_t2), training reward curves, and a side-by-side comparison of the RL policy's trajectory and speed profile against the Bézier pre-trajectory. At minimum, show that the RL policy achieves a strictly lower cornering time or a materially different (e.g., higher side-slip) trajectory in simulation.","section":"§III.A and Tables I, III, IV"},{"comment":"The preview trajectory T_p is generated by simulating the trained TD3 policy on the matched virtual curve (Eq. 18), and the MPC then tracks this trajectory. This makes the reference trajectory endogenous: the real-vehicle success is evaluated against the controller's own simulated output, not against an independently derived optimal line or a ground-truth minimum-time path. The claim that the vehicle tracks the \"optimal trajectory\" is therefore circular in the absence of external validation. Please compare against an independent time-optimal trajectory (e.g., from a separate optimization method) or revise the language to \"policy-generated trajectory\" and provide evidence that this trajectory is near-optimal with respect to cornering time in simulation.","section":"§IV.A"},{"comment":"Each evaluation in Tables III and IV reports a single run per policy. Drift maneuvers are highly sensitive to initial conditions, tire temperature, surface variations, and sensor noise; without repeated runs and error bars (or at least a demonstration of trial-to-trial consistency), the comparison between the proposed policy and the baselines is not statistically grounded. In addition, the baselines are not fully specified: the MPC-only policy is cited to [30] but the exact weights, horizons, and linear tire parameters used here are not given, and the RL-only policy's network architecture and exploration noise during the real-vehicle test are not described. Please add repetition counts, variance or range statistics, and complete baseline parameter settings to make the comparison reproducible.","section":"§V.B and V.C, Tables III and IV"},{"comment":"The robustness analysis in Table V varies the adhesion coefficient in the RL training environment, not in the real vehicle or in the MPC's model. This tests the sensitivity of the simulated policy to friction mismatch, but it does not directly support the claim that the real-vehicle controller is robust to real-world variations, since the MPC correction loop in deployment is not part of this experiment. Please clarify the scope of this robustness test and supplement it with real-world runs on different surfaces (the wet right-angle test already provides some evidence) or explicitly discuss how the MPC correction is expected to handle friction deviations beyond the training envelope.","section":"§V.D"}],"minor_comments":[{"comment":"The description of the constants k_t1 and k_t2 is inconsistent: the text first says \"The constant k_t1 is a positive value\" and then later says \"The constant k_t1, on the other hand, is a positive value that incentivizes achieving the shortest possible extreme cornering time.\" Please clarify the distinct roles of k_t1 and k_t2 in the terminal reward.","section":"Eq. (12) and surrounding text"},{"comment":"The matrices A_t and B_t in Eq. (25) are not defined; the statement that their determination is \"well-established and will not be detailed here\" leaves the reader unable to reproduce the MPC. Please provide the explicit matrices or a reference that includes them.","section":"Eq. (25) and following"},{"comment":"Equations (27) and (28) appear to have formatting issues: the state vector definitions and the integration limits are unclear (e.g., the term \"260\" in the matrix exponential). Please revise these equations for readability and correctness.","section":"Eqs. (27) and (28)"},{"comment":"The phrase \"We believes that this integration enhances\" should read \"We believe that this integration enhances.\"","section":"§IV"},{"comment":"In the abstract, there is a double comma in \"fusion mechanism, , where\" that should be corrected to \"fusion mechanism, where.\"","section":"Abstract"},{"comment":"The paper uses \"Task Completion (deg)\" as an evaluation metric but does not define how this angle is measured (e.g., accumulated heading change, reached checkpoint angle, or final vehicle heading relative to the exit). Please define the metric precisely in the text.","section":"Tables III and IV"},{"comment":"The real-vehicle state plots would be more informative if annotated with the start and end of the corner, the period of the safety fallback (if any), and the side-slip threshold used for fallback activation.","section":"Figures 9 and 11"}],"recommendation":"major_revision","confidential_remarks":"The real-vehicle deployment is genuinely interesting and within the scope of IEEE TIE. My main concern is the exact equality between the RL policy's simulation times and the Bézier pre-trajectory times (2.81 s and 4.52 s), which, if not resolved, effectively removes the RL's claimed role as a cornering-time optimizer. I would need to see training curves, reward weights, and trajectory comparisons before I could recommend acceptance. I would also ask for repeated real-vehicle trials or at least a statement about variability. The paper does not currently mention code or data availability, which would help reproducibility given the many unspecified hyperparameters."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe paper is worth a look for the real-vehicle footage alone: a full-scale consumer EV doing a drift U-turn at 10.4 m/s with 63.7° side-slip, plus a wet-road right-angle drift, with a video link. That is a genuine engineering data point, and the RL-MPC fusion scheme is a sensible way to get a neural policy onto real hardware: use the RL actions as feedforward, let a two-step linear MPC correct the tracking.\n\nBut the central claim that the RL policy optimizes cornering time is not supported by the paper's own numbers. The Bézier pre-trajectory times in Table I are exactly 2.81 s (90° turn) and 4.52 s (U-turn). The trained RL policy in simulation achieves exactly 2.81 s and 4.52 s in Tables III and IV. Two decimal places, two scenarios. That is not a coincidence; it tells me the RL policy collapsed into tracking the Bézier path and did not find a faster line. So the 'preview trajectory' T_p generated by simulating the policy is just the Bézier path in disguise, and the 'RL-based transient drift cornering' novelty is substantially weakened. To salvage the claim, the authors need to show learning curves or a comparison where the RL policy is faster than the Bézier baseline. If they cannot, the contribution should be reframed as a trajectory-tracking fusion controller, which is still useful.\n\nOther soft spots: the 'first deployment' claim sits next to their own reference [19], which already did sim-to-real RL drifting on a real vehicle. The real-vehicle tests are single trials with no error bars, no repeated runs, and no code or hyperparameters, so the comparison with the RL-only and MPC-only baselines is hard to assess. Several equations in the preprint are garbled or incomplete (Eq. 1, 3, 18, 25), which may be a formatting issue but needs fixing.\n\nProportionately: the engineering demonstration is plausible and the fusion idea is reasonable. The paper deserves a serious referee, but only with a major-revision request. The authors should be asked to (1) provide learning curves and trajectory comparisons showing whether RL improves on the Bézier baseline, (2) run repeated trials and report variance, (3) reconcile the \"first\" claim with [19], and (4) release code or parameters. If the time-equality issue is not resolved, the paper should be revised to drop the 'RL optimizes cornering time' claim.\n\nMy take: send it to review, but expect heavy revision.","headline":"Real-vehicle drift demo is real, but the RL contribution collapses to Bézier tracking on the paper's own numbers.","tokens_in":16199,"tokens_out":3903,"would_cite":false,"duration_ms":35199,"reading_group":"maybe","serious_thinker":"unclear","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A simulation-trained RL policy can drive a consumer electric vehicle through real drift U-turns and wet right-angle turns when a two-step linear MPC handles the trajectory tracking.","keywords":["autonomous driving","drift cornering","reinforcement learning","TD3","model predictive control","sim-to-real transfer","trajectory optimization","electric vehicle dynamics"],"falsifier":"Run the same U-turn test on a surface with measured peak friction near 0.55: the paper's own parameter sweep predicts only partial completion (127 of 180 degrees), so a full completion would contradict the claimed robustness boundary, while tracking failure with the MPC correction saturated would show that the linear correction cannot cover large model mismatches.","tokens_in":15153,"feed_emoji":"🏎️","tokens_out":7675,"duration_ms":70825,"temperature":0.7,"pith_summary":"This paper tackles the gap between simulation-trained reinforcement-learning controllers and real vehicles, a gap that has kept drift cornering research mostly in simulation. It proposes an RL-MPC fusion in which a TD3 policy trained in a simulated environment produces the primary control actions, while a two-step model predictive controller with a linear tire model generates corrective inputs so the vehicle tracks the trajectory the TD3 policy produces. The paper reports the first deployment of a scenario-to-action RL-based transient drift cornering algorithm on consumer-grade electric vehicles. In real-vehicle tests the fused controller completed a U-turn at a maximum speed of 10.4 m/s with a maximum side-slip angle of 63.7 degrees, and a right-angle drift turn on wet asphalt. If these results hold, simulation-trained extreme-maneuver controllers can be moved to real vehicles without retraining, shortening the path to time-optimal autonomous cornering.","feed_headline":"Consumer EV completes drift U-turn with RL-MPC fusion","feed_subtitle":"Simulation-trained TD3 supplies the maneuver; a two-step MPC corrects it in real time on wet asphalt.","key_machinery":"The load-bearing mechanism is the fusion of three components. First, a Bézier-based pre-trajectory optimizer solves, in Frenet coordinates, for a minimum-curvature path $l(s)$ and a curvature-limited speed plan; that pre-optimized trajectory defines the immediate reward that makes TD3 training feasible despite a sparse terminal time reward. Second, the trained TD3 policy, run offline on the virtual curve matched to the real test site, produces the preview trajectory $T_p$ in Cartesian coordinates; this trajectory carries the optimal-maneuver information. Third, a two-step model predictive controller built on a linearized three-degree-of-freedom vehicle model generates corrective inputs $\\Delta u_{MPC}$ that are added to the RL actions, with a safety fallback that applies moderate braking when the side-slip angle exceeds a threshold. The MPC's role is not to plan the drift but to hold the vehicle on the RL-produced trajectory, which is why the authors can use a linear tire model in a regime where the dynamics are strongly nonlinear.","core_discovery":"The paper's central claim is that a simulated RL policy can be made to control a real vehicle in a highly nonlinear drift regime by splitting its role in two. Offline, the trained TD3 policy is run on a matched virtual curve to produce a preview trajectory; this trajectory is the reference, and it is deliberately different from the Bézier pre-optimized path used during training. Online, the policy's actions are still sent to the vehicle as primary inputs, but a two-step MPC with a linear tire model measures tracking error against the preview trajectory and adds corrective inputs in steering and longitudinal acceleration, converted to torque and brake pressure. The fused input is $u_t = A_{RL} + \\Delta u_{MPC}$. On a consumer-grade rear-wheel-drive electric vehicle, this controller completed a 180-degree drift U-turn in 4.8 s with a maximum speed of 10.4 m/s and a maximum center-of-mass side-slip angle of 63.7 degrees, and completed a 90-degree right-angle drift turn on a wet road in 3.19 s, while the RL-only policy failed the U-turn and the MPC-only policy could not stabilize it.","pith_inferences":["Because the TD3 policy is reduced to a reference generator, the same RL-MPC wrapper could in principle deploy other simulation-trained maneuver policies, such as obstacle-avoidance or recovery maneuvers, without retraining the wrapper.","The virtual-curve matching step is currently offline and scenario-specific; a natural extension is online curve identification, which would let the preview trajectory be regenerated as the vehicle approaches an unseen corner.","The robustness sweep varies friction inside the training environment, not on the real road; directly measuring real surface friction and watching whether the MPC correction grows in proportion would quantify how much of the claimed gap-bridging is due to the fusion.","The failure at 0.55 suggests that a friction-estimation module, rather than a fixed corrective MPC, would be the next step if the controller is to operate safely near the edge of the traction envelope."],"forward_implications":["Sim-to-real transfer for transient drift control can be achieved without domain randomization or on-vehicle retraining; the MPC correction absorbs a meaningful amount of model mismatch.","Consumer-grade electric vehicles, not specialized research platforms, become capable of executing high-side-slip cornering maneuvers, changing what production stability control could attempt.","The U-turn and right-angle turn were completed in 4.81 s and 3.19 s respectively, close to the simulated RL baselines of 4.52 s and 2.81 s, so the fusion preserves most of the speed advantage of the learned policy.","The controller degrades gracefully as the adhesion coefficient in the training environment drops from 0.95 to 0.65, with cornering time increasing but the task still completed; at 0.55 the maneuver is only partially completed, defining a practical operating envelope."],"supporting_citations":[{"why":"Supplies the deep-RL drifting baseline that motivated training a scenario-to-action controller in simulation.","marker":"[8]"},{"why":"Demonstrates real-vehicle steady-state drifting and figure-eight maneuvers with feedforward-feedback and NMPC, the real-vehicle deployment direction this work extends to transient RL-based cornering.","marker":"[15,16]"},{"why":"Plans reference drift trajectories with high sideslip for RC-scale cornering, the rule-based approach the paper contrasts with learned maneuvers.","marker":"[17]"},{"why":"Motivates Bézier-curve trajectory generation used for the pre-optimized path that shapes RL rewards and speed planning.","marker":"[25]"},{"why":"Introduces TD3, the actor-critic algorithm whose twin critics and delayed updates the paper adopts for the RL policy.","marker":"[27]"},{"why":"Provides the linear tire model and MPC formulation that the corrective-input controller builds on.","marker":"[29]"},{"why":"Is the MPC-only trajectory-tracking baseline that the paper compares against in the real-vehicle U-turn and right-angle tests.","marker":"[30]"}],"fun_headline_variants":["RL-MPC fusion drifts consumer EV through U-turn","First RL-based drift cornering on consumer EV","Hybrid RL-MPC controller master EV drift turns","Sim-trained TD3 and MPC execute real-vehicle drift","Drift U-turn achieved on consumer EV via RL-MPC"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing assumption is that simulating the trained TD3 policy on a matched virtual curve yields a preview trajectory close enough to the real optimum, and that the two-step linear-tire MPC can correct the real vehicle's deviations well enough without invoking the full nonlinear dynamics.","fun_headline_variants_meta":{"raw":{"variants":["RL-MPC fusion drifts consumer EV through U-turn","First RL-based drift cornering on consumer EV","Hybrid RL-MPC controller master EV drift turns","Sim-trained TD3 and MPC execute real-vehicle drift","Drift U-turn achieved on consumer EV via RL-MPC"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00035,"raw_usage":{"total_tokens":1953,"prompt_tokens":1027,"completion_tokens":926,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":643,"completion_tokens_details":{"reasoning_tokens":847}},"tokens_in":643,"tokens_out":926,"duration_ms":9071,"temperature":1.0,"reasoning_tokens":847,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T18:08:59.643837+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same U-turn test on a surface with measured peak friction near 0.55: the paper's own parameter sweep predicts only partial completion (127 of 180 degrees), so a full completion would contradict the claimed robustness boundary, while tracking failure with the MPC correction saturated would show that the linear correction cannot cover large model mismatches.","supporting_citations":[{"cited_title":"High -speed autonomous drifting with deep reinforcement learning,","cited_arxiv_id":null,"evidence_quote":"Supplies the deep-RL drifting baseline that motivated training a scenario-to-action controller in simulation."},{"cited_title":"Drift control for cornering maneuver of autonomous vehicles,","cited_arxiv_id":null,"evidence_quote":"Plans reference drift trajectories with high sideslip for RC-scale cornering, the rule-based approach the paper contrasts with learned maneuvers."},{"cited_title":"Trajectory and Velocity Planning Method of Emergency Rescue Vehicle Based on Segmented Three - Dimensional Quartic Bezier Curve,","cited_arxiv_id":null,"evidence_quote":"Motivates Bézier-curve trajectory generation used for the pre-optimized path that shapes RL rewards and speed planning."},{"cited_title":"Twin -delayed ddpg: A deep reinforcement learning technique to model a continuous movement of an intelligent robot agent,","cited_arxiv_id":null,"evidence_quote":"Introduces TD3, the actor-critic algorithm whose twin critics and delayed updates the paper adopts for the RL policy."},{"cited_title":"A Three -Dimensional Dynamics Control Framework of Vehicle Lateral Stability and Rollover Prevention via Active Braking With MPC,","cited_arxiv_id":null,"evidence_quote":"Provides the linear tire model and MPC formulation that the corrective-input controller builds on."},{"cited_title":"Trajectory Tracking of Autonomous Vehicle Based on Model Predictive Control With PID Feedback ,","cited_arxiv_id":null,"evidence_quote":"Is the MPC-only trajectory-tracking baseline that the paper compares against in the real-vehicle U-turn and right-angle tests."}],"review_version":1}