Pith. sign in

REVIEW 4 major objections 4 minor 45 references

This paper claims that a minimum-lap-time trajectory computed by optimal control can replace expert demonstrations and guide a staged reinforcement-learning controller to execute autonomous drifting faster than a human driver in simulation.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · deepseek-v4-flash

2026-08-04 01:15 UTC pith:QX5EOW76

load-bearing objection A promising integration of MLT planning and curriculum RL for drifting, but internal inconsistencies and an untested planner-to-simulator gap leave the central claim unsupported as written. the 4 major comments →

arxiv 2608.00113 v1 pith:QX5EOW76 submitted 2026-07-31 cs.RO

Track-Guided Hierarchical Reinforcement Learning for Autonomous Vehicle Drifting with Minimum-Lap-Time Planning

Sheng Zhao , Bolin Zhao , Xiaodong Wu , Chen Lv This is my paper
classification cs.RO
keywords drift controlreinforcement learningminimum-lap-time planningcurriculum learningoptimal controlautonomous racingsideslip anglevehicle dynamics
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The paper proposes a two-part framework for autonomous drifting that minimizes lap time. First, an optimal-control planner computes a minimum-lap-time drift trajectory from a simplified vehicle-and-tire model, including speed, sideslip, and yaw profiles. Second, that trajectory is used as prior training data for a curriculum-based reinforcement-learning controller that learns to drift in three stages: basic drift control on a straight track, drift cornering on hairpins, then full-lap racing. The authors claim the resulting controller outperforms both a previous RL drifting method and a human driver on three simulated tracks, approaching the planner's theoretical lower bound. The core point is that model-based trajectory optimization can substitute for costly expert demonstrations in teaching extreme vehicle maneuvers.

Core claim

The central claim is that minimum-lap-time drift planning (MLTDP) provides a good enough reference that a track-guided RL controller (TgRL) can learn to stabilize large-sideslip drifting and reduce lap time. The paper formulates drifting as an optimal control problem in Frenet coordinates, minimizing travel time over arc length subject to vehicle dynamics, tire limits, road boundaries, and state and control constraints. The resulting reference trajectory supplies speed, sideslip angle, and yaw-rate targets. An entropy-regularized actor-critic agent is then trained through a three-stage curriculum—basic drift control, drift cornering, drift racing—with a reward that combines tracking errors a

What carries the argument

The load-bearing objects are the minimum-lap-time reference trajectory and the three-stage curriculum. The planner solves an optimal control problem over arc length, minimizing the integral of (1-l·κ)/(v·cos(Δψ+β)) under a 3-DoF bicycle model with Magic Formula tire forces, producing positions, speeds, sideslip angles, and yaw rates that serve as prior data. The controller then uses that trajectory in stagewise training: first learning to hold a drift on a straight track, then applying it through hairpins, then combining both into a full-lap race policy. The reward's terminal term is scaled by the gap between achieved lap time and the planned optimal lap time, which is what connects the RL o

Load-bearing premise

The plan's value hinges on the assumption that the simplified vehicle-and-tire model, tuned for a low-friction road, produces a trajectory the simulated car can actually follow and that is close to the fastest possible trajectory for that car.

What would settle it

Retrain the controller with the planner's reference replaced by a randomly perturbed or centerline trajectory while keeping the reward structure fixed; if lap times stay similar, the minimum-lap-time prior is not the cause of the speed-up. Alternatively, measure tracking of the planned sideslip profile at the two corners where the paper reports the largest deviations; if the agent cannot follow the drift posture there, the lap-time gain would have to come from something other than executing the MLT plan.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • An RL drift controller can be trained without human expert demonstrations, using an optimal-control trajectory as the prior.
  • A three-stage curriculum (straight-line drift, corner drift, full-lap race) makes learning feasible where a plain actor-critic agent gets trapped in negative reward.
  • The trained policy generalizes across three different track layouts and consistently beats a prior RL drifting baseline, with the kart-track average lap time dropping from 51.2 s to 45.4 s.
  • Executed drifts reach average sideslip angles around 16–18 degrees, well beyond the steady-state cornering envelope, while keeping yaw error lower than the baseline.
  • The framework approaches the idealized planner's lap time within about 12–17 percent on the tested tracks.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Our inference: if the planner-to-simulator model gap is the main source of the remaining 12–17 percent, updating the plan online with a learned dynamics model could close much of that gap without changing the curriculum.
  • Our inference: the action space restricts throttle to 0.6–1.0 and omits braking, so the controller cannot reproduce planner behavior that requires lifting off the throttle or braking; part of the measured gap may stem from this restriction rather than from policy failure.
  • Our inference: the same recipe—optimal-control trajectory as a curriculum prior plus a lap-time-scaled terminal reward—should transfer to other extreme maneuvers, such as drift parking or emergency obstacle avoidance, where a planner can supply the reference and RL absorbs model error.
  • Our inference: a sharper test of the mechanism would corrupt only the planner's sideslip profile while keeping the geometric path fixed; if lap times barely change, the benefit comes from the racing line itself rather than from learning the drift posture.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. This paper proposes MLTDP-TgRL, a hierarchical planning/RL framework for autonomous drifting. MLTDP formulates a minimum-lap-time optimal control problem in Frenet coordinates using a 3-DoF bicycle model with a Magic Formula tire model (μ=0.6). The resulting trajectory is used as a reference prior for a SAC-based controller trained in CARLA via a three-stage curriculum: basic drift control, corner drifting, and full race policy. The reward combines instant tracking terms with an end reward based on the planner's lap time. Evaluations on Lkarting, Willow Springs, and Nelson Ledges compare against a published drift RL baseline (Cai_DRIFT) and a human driver; ablations vary the reference (MLT vs centerline), RL variant, end reward, and action smoothing. The paper claims superior lap times and larger sideslip angles while maintaining tracking accuracy.

Significance. If the claims hold, the framework is a useful step toward replacing expert demonstrations with model-based trajectory priors for drift control, and the three-stage curriculum plus action smoothing are sensible engineering contributions. The comparison to a published drift RL baseline and human data on three tracks, together with the ablation studies, is a substantial evaluation. However, the current manuscript does not yet substantiate the central lap-time claim because of numerical inconsistencies in key results, missing uncertainty quantification, and incomplete validation of planner feasibility. These issues are fixable but require a revision.

major comments (4)
  1. [§5.4, Table 4] The headline Lkarting result is internally inconsistent: the text says MLT+TgRL "completes laps in an average of 55.4 s," but Table 4 gives Avg = 45.4 s for MLT+TgRL and 51.2 s for Cai_DRIFT. The stated "approximately 10%" gain is consistent with 45.4 vs 51.2, so 55.4 appears to be a typo, but it is central and must be corrected. Similarly, Sec. 5.3 reports TgRL-Control lap time "about 36.5 s," while Table 4 lists Min = 37.2 s. These numbers need to be reconciled before the results can be assessed.
  2. [§4.4.2, Table 3, §5.3] The end reward is claimed to be derived from the MLTDP optimum, but the illustrative example uses Ttotal = 46 s and k6 = 5e5, while Sec. 5.3 and Table 4 report MLT Planning ≈33 s/33.1 s and Table 3 lists k6 = 1e5. With Ttotal = 33.1 s and k6 = 1e5, a 50 s lap would give an end reward of about 3.4e3, not 2.25e5. Because the end reward is one of the key mechanisms for transferring the MLT objective into RL, the actual Ttotal used for Lkarting (and per track, if varied) must be reported consistently; otherwise the claim that the reward encodes the minimum-lap-time objective is not verifiable.
  3. [§5.3, §5.4] The MLTDP trajectory is computed from a simplified 3-DoF bicycle/Magic Formula model (μ=0.6) but evaluated with CARLA's built-in vehicle model. The paper acknowledges "differences in the vehicle models used for planning and control" and reports a large sideslip gap (~30° planned vs ~18° executed). Yet contribution (2) calls these "dynamically feasible" trajectories and Sec. 5.4 treats MLT Planning as a "theoretical lower bound." No experiment shows that the planner's solution is dynamically feasible for the CARLA vehicle. Since the main claim is that the MLT prior, not just the reward shaping/curriculum, produces the lap-time gains, the paper should either (i) report closed-loop tracking of the MLT trajectory (e.g., with a well-tuned MPC or an oracle tracker), or (ii) soften the lower-bound language and state explicitly that the planner is a heuristic prior. Figure 12(a)'s MLT-vs-centerl
  4. [Table 4, §5.4, Fig. 12] There is no uncertainty quantification. The averages in Table 4 are over ten laps, but no standard deviations, confidence intervals, number of training seeds, or statistical comparisons are reported. RL training is stochastic, and the differences against Cai_DRIFT on Willow Springs (Min 181.1 vs 193.1; Avg 197.5 vs 197.7) are small enough that seed-to-seed variation could change the conclusion. Similarly, Fig. 12 reports single numbers per method. Add per-seed results, error bars, and an explicit statement of how many seeds and laps were used; otherwise the claim of "consistently" superior performance is not supported.
minor comments (4)
  1. [Fig. 12(e), §4.3] The text calls AAS an "Adaptive Angle Shaping reward," but in Sec. 4.3 AAS is Adaptive Action Smoothing, not a reward. Correct the label and terminology.
  2. [Table 3, Eq. (23)] The ASS speed range is listed as [16.67, 20] km/h, but vehicle speeds in the paper are in m/s (e.g., 0–30 m/s in Eq. 16). If the intended values are m/s, the range is implausibly low for the stated purpose; if km/h, convert to m/s and use consistent units.
  3. [§5.4] The sentence that the Lkarting average sideslip angle of 15.9° is "close to the optimal planning reference of 29.8°" is misleading; 15.9° is roughly half of 29.8°. Clarify that 29.8° is a per-lap maximum, not a time-average, or rephrase.
  4. [Table 4] The column header "Sideslip Angle Avg (Laps)" is confusing, and the MLT Planning rows present a single run with no indication of repeatability. Define the metric precisely and use consistent formatting.

Circularity Check

0 steps flagged

No significant circularity: lap-time results are measured against external baselines; MLTDP trajectory is an independent OCP output used only as reward guidance.

full rationale

The paper's derivation chain is not circular in the sense that matters here. The MLTDP trajectory is produced by solving the optimal-control problem in Eqs. (7)-(15) from the 3-DoF bicycle/Magic Formula model, with no dependence on the RL controller or on the reported lap times. The RL agent is then trained with a reward (Eqs. 24-27) that tracks this trajectory and uses T_total from MLTDP as an end-reward anchor. This is intentional guidance, not a fitted parameter renamed as a prediction: the final lap times in Table 4 are measured in CARLA against external baselines (Cai_DRIFT and a human driver), so the claimed lap-time improvement is empirical and falsifiable. The ablation study (Fig. 12) also compares against centerline-tracking agents, providing independent evidence that the MLT reference, not merely the reward shaping, contributes. The Sec. 5.3 admission of vehicle-model mismatch ('This discrepancy is likely due to differences in the vehicle models used for planning and control') is a feasibility/correctness concern, not a circular step: the planner is not defined in terms of the RL outcome. The inconsistency between the Sec. 4.4.2 example T_total=46 s and the Sec. 5.3/Table 4 MLT Planning time of ~33 s weakens the claim that the end reward is faithfully tied to the planner's optimum, but it does not make the derivation reduce to its inputs. No load-bearing self-citations, uniqueness theorems, or ansatz-smuggling are present; the self-citations are background. Therefore the circularity score is 0.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 0 invented entities

The paper rests on standard vehicle-dynamics modeling choices and simulator fidelity assumptions. The most consequential are the transfer of the simplified planning model to the CARLA vehicle and the hand-tuned reward design. No new physical entities are introduced.

free parameters (8)
  • Magic Formula tire coefficients B,C,D,E = not provided
    Empirical tire-model constants used in Eq (6); required by MLTDP but absent from Table 1, so the planned trajectories cannot be regenerated.
  • Reward function kernel widths k1,k2,k3 = 8, 3200, 5
    Hand-set Gaussian kernel widths in Eq (24) that shape tracking reward; not justified by data.
  • Instant reward weights k_ey,k_eψ,k_eβ,k_ev = 20, 40, 20, 40
    Hand-tuned weights in Eq (25) balancing path, heading, sideslip, and speed errors; directly influence learned policy.
  • Velocity discount K and threshold v_thre = 0.5, 5 m/s
    Eq (26) penalizes slow driving; threshold chosen by hand.
  • End reward coefficients k4,k5,k6,k7 = 1e5, -0.05, 1e5, -0.2 (Table 3; Sec 4.4.2 example uses k6=5e5)
    Eq (27) terminal rewards; k7 defines the optimal lap-time target sensitivity; k6 inconsistent between table and text.
  • AAS smoothing parameters λτ, λδ,min, λδ,max, [v_min,v_max] = λτ=0.2; [16.67,20] km/h; λδ range not given
    Eq (22)-(23) adaptive action smoothing; λδ endpoints are unspecified, so exact controller behavior is not reproducible.
  • Lookahead horizon N = 10
    State space size (Sec 4.2); chosen by the authors.
  • Curriculum stage episode counts M,P = not specified
    Stage boundaries in Fig 6 and Sec 4.6; without values the curriculum is not reproducible.
axioms (6)
  • domain assumption The 3-DoF bicycle model (Eq 1) and Magic Formula combined-slip tire model (Eq 5-6) adequately describe vehicle drift dynamics for planning.
    Sec 2 uses these models for MLTDP; if inaccurate, the planned 'minimum-lap-time' trajectory is not dynamically feasible.
  • domain assumption The road friction coefficient is constant at μ=0.6 on all three tracks (Table 1).
    Used in both planning and CARLA setup; low-friction surfaces vary in reality.
  • domain assumption The CARLA simulator's built-in vehicle model is a valid testbed that captures the nonlinearities of drifting.
    Sec 5.1: the built-in CARLA model is adopted; no evidence of fidelity to a real vehicle is provided.
  • domain assumption SNOPT converges to a sufficiently accurate local solution of the MLT OCP.
    Sec 5.1 uses SNOPT; no grid-independence or optimality gap is reported.
  • ad hoc to paper The hand-designed reward (Eq 24-28) correctly encodes the minimum-lap-time objective.
    The reward combines tracking errors, speed term, and terminal lap-time term; its weights are not derived from data.
  • ad hoc to paper The three-stage curriculum transfers to full-track racing without simulator-to-simulator gaps beyond the reported transfer shocks.
    Sec 4.6 assumes skills learned on straight and hairpin tracks compose into a complete race policy; no guarantee is given.

pith-pipeline@v1.3.0-alltime-deepseek · 19212 in / 20307 out tokens · 193252 ms · 2026-08-04T01:15:25.793902+00:00 · methodology

0 comments
read the original abstract

In Formula 1, drivers optimize racing lines within tire grip limits to minimize lap times; however, in rally racing, drivers intentionally break traction to drift on loose surfaces. This maneuver rapidly aligns the vehicle for corner exits, ultimately reducing lap time. Autonomously executing such maneuvers formulates a complex dual-objective control problem: stabilizing highly nonlinear drift dynamics while strictly minimizing lap time. Addressing this challenge motivates the development of advanced Minimum-Lap-Time (MLT) drift control architectures. This paper proposes a planning-control framework specifically designed for MLT drifting scenario. First, we formulate an optimal control problem to generate a MLT drift planning trajectory, which is used as prior data to train a deep reinforcement learning drift controller. Given that drifting involves extremely large sideslip angles and is therefore challenging to learn directly, a Track-guided Reinforcement Learning (TgRL) drift control method is proposed to enable progressive training in a step-by-step manner, from drift control policy, to drift corner policy, and finally to a comprehensive drift race policy. The reward function incorporates both an instant reward term and an end reward term derived from the Minimum-Lap-Time objective. Simulation results demonstrate that the proposed framework enables the agent to learn a drift racing policy that not only ensures vehicle motion control performance but also effectively reduces lap time.

Figures

Figures reproduced from arXiv: 2608.00113 by Bolin Zhao, Chen Lv, Sheng Zhao, Xiaodong Wu.

Figure 1
Figure 1. Figure 1: Essence of drift strategy: under varying road con￾ditions (corner curvature, surface friction, lane width) and different vehicle speeds, the agent leverages the RL policy network to infer the corresponding optimal drift posture and control method. Championship (WRC) [20, 33, 34]. In extreme avoidance cases, drivers typically prioritize braking to prevent rear-end collisions (e.g., at red-light intersection… view at source ↗
Figure 2
Figure 2. Figure 2: The equations of motion are then ⎧ ⎪ ⎪ ⎪ ⎨ ⎪ ⎪ ⎪ ⎩ ̇𝑣 = −𝐹𝑌𝑓 sin(𝛿 − 𝛽) + 𝐹𝑋𝑟 cos(𝛽) + 𝐹𝑌𝑟 sin(𝛽) 𝑚 𝛽̇ = 𝐹𝑌𝑓 cos(𝛿 − 𝛽) − 𝐹𝑋𝑟 sin(𝛽) + 𝐹𝑌𝑟 cos(𝛽) 𝑚𝑣 − 𝑟 ̇𝑟 = 𝑎𝐹𝑌𝑓 cos(𝛿) − 𝑏𝐹𝑌𝑟 𝐼𝑧 (1) [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: These data serve as crucial prior information for the drift tracking control discussed in Section 4. 3.2. Objective function The problem of planning the MLT trajectory through a track for a vehicle can be formulated as an OCP. In this formulation, the performance index of the optimal control problem is the minimum time to traverse the curve: min 𝐽 = ∫ 𝑡𝑓 𝑡0 𝑑𝑡 (12) where, 𝑡0 is known, 𝑡𝑓 is unknown, so it … view at source ↗
Figure 4
Figure 4. Figure 4: Drivable boundary constraints definition. 3.3.2. Control variable constraints Eq (17) defines the bounds on the control variables. Based on engineering experience, the front wheel steering angle 𝛿𝑓 is typically limited to within ±30◦ . The slip ratio of the driving wheels is required to be greater than −1. { −𝜋∕6 ≤ 𝛿 ≤ 𝜋∕6 −1 ≤ 𝜆𝑟 (17) 3.3.3. Drivable boundary constraints [PITH_FULL_IMAGE:figures/full_fig… view at source ↗
Figure 5
Figure 5. Figure 5: The proposed MLTDP-TgRL vehicle autonomous drifting framework. Therefore, the dimension of the set  is determined by the parameter 𝑁, which typically takes the value 𝑁 = 10. To accommodate termination conditions during training, a terminal state indicator, EndFlag, is introduced. A train￾ing episode is terminated once the vehicle experiences a collision, deviates by 180ř from the reference heading (i.e., … view at source ↗
Figure 6
Figure 6. Figure 6: TgRL training method. The TgRL training method is designed as a progressive learning framework in which the blue segments of the track represent normal steady-state driving and the orange segments indicate drifting states. The reinforcement learning (RL) agent is trained in a step-by-step manner across three key stages. During the basic drift control policy learning stage (episodes 0 to M), the vehicle pri… view at source ↗
Figure 8
Figure 8. Figure 8: A human driver operating a Logitech G29 simulator setup to perform tests on different tracks [PITH_FULL_IMAGE:figures/full_fig_p010_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Drift test on Lkarting racetrack. (a) Trajectories; (b) Velocity; (c) Sideslip angle; (d) Steer; (e) Throttle. 50 100 150 200 250 300 350 400 450 500 Episode -100,000 -50,000 0 50,000 100,000 150,000 200,000 250,000 300,000 Cumulative Episodic Reward Stage 1: Basic Policy Stage 2: Corner Policy Stage 3: Race Policy TgRL (ours) SAC Baseline [PITH_FULL_IMAGE:figures/full_fig_p011_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Training convergence of the proposed TgRL com￾pared to the SAC baseline. The training process of TgRL is divided into three progressive curriculum stages. The inset plot details the transfer shocks (yellow triangles) that occur at the boundaries of tasks with increasing complexity (Episodes 100 and 200). the standard SAC agent struggles with early-stage explo￾ration, remaining trapped in a suboptimal nega… view at source ↗
Figure 11
Figure 11. Figure 11: Three-dimensional phase diagram of vehicle dynamics states. 5.4. Different Racetrack Result Analysis To further validate the effectiveness and generalization capability of the proposed drift control framework, addi￾tional experiments were conducted on two well-known cir￾cuits, Willow Springs and Nelson Ledges, in addition to the previously tested Lkarting track. The layouts of these race￾tracks are shown … view at source ↗
Figure 12
Figure 12. Figure 12: The ablation experiment of our methods: (a) Lap time; (b) Sideslip angle (max); (c) Lateral offset to centerline (avg). (d) Impact of End reward to the lap time and L.O; (e) Impact of AAS to the lap time and sideslip angle. of whether TgRL or standard RL is used, consistently out￾perform those trained with centerline references, highlight￾ing the critical role of model-based trajectory optimization in sha… view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

45 extracted references · 1 linked inside Pith

  1. [1]

    R. Y. Hindiyeh, J. Christian Gerdes, A controller framework for autonomous drifting: Design, stability, and experimental validation, J. of Dyn. Syst., Meas., and Control 136 (2014) 051015

    2014

  2. [2]

    Z. Shan, J. Zhao, B. Zhu, C. Lv, Y. Zhao, L. Ge, S. Zhong, Safe and efficient trajectory planning considering longitudinal and lateral coupled limits, IEEE Trans. Veh. Technol. 73 (2024) 10714–10719

    2024

  3. [3]

    J. Ni, J. Hu, C. Xiang, Envelope control for four-wheel independently actuated autonomous ground vehicle through afs/dyc integrated con- trol, IEEE Trans. Veh. Technol. 66 (2017) 9712–9726

    2017

  4. [4]

    Cheng, B.-B

    S. Cheng, B.-B. Hu, H.-L. Wei, L. Li, C. Lv, Deep learning-based hybrid dynamic modeling and improved handling stability assessment for autonomous vehicles at driving limits, IEEE Trans. Veh. Technol. 74 (2025) 5582–5593

    2025

  5. [5]

    N. A. Spielberg, M. Brown, N. R. Kapania, J. C. Kegelman, J. C. Gerdes, Neural network vehicle models for high-performance auto- mated driving, Sci. Robot. 4 (2019) eaaw1975

    2019

  6. [6]

    N. D. Broadbent, T. Weber, D. Mori, J. C. Gerdes, Neural network tire force modeling forăautomated drifting, in: 16th Int. Symp. on Adv. Vehicle Control, Springer Nature Switzerland, Cham, 2024, pp. 378–384

    2024

  7. [7]

    T. P. Weber, J. C. Gerdes, Modeling and control for dynamic drifting trajectories, IEEE Trans. Intell. Veh. 9 (2024) 3731–3741

    2024

  8. [8]

    T. P. Weber, R. K. Aggarwal, J. C. Gerdes, Human-inspired au- tonomous racing in low friction environments, IEEE Trans. Intell. Veh. (2024) 1–14

    2024

  9. [9]

    Q. Ma, X. Yin, X. Zhang, X. Xu, X. Yao, Game-theoretic receding- horizon reinforcement learning for lateral control of autonomous vehicles, IEEE Trans. Veh. Technol. 73 (2024) 14547–14562

    2024

  10. [10]

    X. Wu, J. Li, C. Su, J. Fan, M. Xu, A deep reinforcement learning based hierarchical eco-driving strategy for connected and automated hevs, IEEE Trans. Veh. Technol. 72 (2023) 13901–13916

    2023

  11. [11]

    S. Zhao, J. Zhang, N. Masoud, Y. Jiang, H. Huang, T. Liu, Drift cornering control and real-vehicle deployment for electric vehicles, IEEE Transactions on Industrial Electronics 72 (2025) 13509–13520

    2025

  12. [12]

    P. R. Wurman, et al., Outracing champion gran turismo drivers with deep reinforcement learning, Nature 602 (2022) 223–228

    2022

  13. [13]

    P. Cai, X. Mei, L. Tai, Y. Sun, M. Liu, High-speed autonomous drifting with deep reinforcement learning, IEEE Robot. and Automat. Lett. 5 (2020) 1247–1254

    2020

  14. [14]

    Jiang, X

    Y. Jiang, X. Xu, X. Zhang, J. Huang, S. Gao, Reinforcement learning-based drift control for autonomous vehicles, in: 2021 China Automat. Congr. (CAC), 2021, pp. 4508–4513. doi:10.1109/CAC53003. 2021.9727514

    arXiv 2021

  15. [15]

    Y. Yin, S. E. Li, K. Li, J. Yang, F. Ma, Self-learning drift control of automated vehicles beyond handling limit after rear-end collision, Transp. Saf. and Environ. 2 (2020) 97–105

    2020

  16. [16]

    S. Zhao, J. Zhang, C. He, X. Hou, H. Huang, Adaptive drift control of autonomous electric vehicles after brake system failures, IEEE Trans. Ind. Electron. 71 (2024) 6041–6052

    2024

  17. [17]

    S. H. Tóth, Ádám Bárdos, Z. J. Viharos, Tabular Q-learning based reinforcement learning agent for autonomous vehicle drift initiation and stabilization, IFAC-PapersOnLine 56 (2023) 4896–4903. 22nd IFAC World Congress

    2023

  18. [18]

    S. H. Tóth, Z. J. Viharos, . Bárdos, Z. Szalay, Sim-to-real application of reinforcement learning agents for autonomous, real vehicle drift- ing, Vehicles 6 (2024) 781–798

    2024

  19. [19]

    Y. Wang, X. Yuan, C. Sun, Learning autonomous race driving with action mapping reinforcement learning, ISA Trans. 150 (2024) 1–14. Sheng Zhao et al.: Preprint submitted to Elsevier Page 14 of 15 Track-Guided Hierarchical Reinforcement Learning for Autonomous Vehicle Drifting with Minimum-Lap-Time Planning

    2024

  20. [20]

    X. Hou, J. Zhang, C. He, Y. Ji, J. Zhang, J. Han, Autonomous driving at the handling limit using residual reinforcement learning, Adv. Eng. Inform. 54 (2022) 101754

    2022

  21. [21]

    Domberg, C

    F. Domberg, C. C. Wembers, H. Patel, G. Schildbach, Deep drifting: Autonomous drifting of arbitrary trajectories using deep reinforce- ment learning, in: 2022 IEEE Int. Conf. on Robot. and Automat. (ICRA), 2022, pp. 7753–7759. doi: 10.1109/ICRA46639.2022.9812249

    arXiv 2022

  22. [22]

    Domberg, B

    F. Domberg, B. Barkow, G. Schildbach, Vision-based autonomous trajectory drifting using deep reinforcement learning, in: AmEC 2024 Automot. meets Electro. and Control; 14. GMM Symp., 2024, pp. 47– 52

    2024

  23. [23]

    Djeumou, M

    F. Djeumou, M. Thompson, M. Suminaka, J. Subosits, Reference-free formula drift with reinforcement learning: From driving data to tire energy-inspired, real-world policies, ArXiv abs/2410.20990 (2024)

    Pith/arXiv arXiv 2024

  24. [24]

    Orgován, T

    L. Orgován, T. Bécsi, S. Aradi, Autonomous drifting using rein- forcement learning, Periodica Polytechnica Transp. Eng. 49 (2021) 292300

    2021

  25. [25]

    B. Leng, Y. Yu, M. Liu, et al., Deep reinforcement learning-based drift parking control of automated vehicles, Sci. China Technological Sciences 66 (2023) 1152–1165

    2023

  26. [26]

    Bhattacharjee, D

    S. Bhattacharjee, D. Schnieders, Autonomous drifting RC car with reinforcement learning, interim report (2018)

    2018

  27. [27]

    Hoshino, J

    H. Hoshino, J. Li, A. Menon, J. M. Dolan, Y. Nakahira, Autonomous drifting based on maximal safety probability learning, in: 2024 IEEE 27th Int. Conf. on Intell. Transp. Syst. (ITSC 2024), 2024, pp. 3930–

    2024

  28. [28]

    J. Li, X. Wu, M. Xu, Y. Liu, Deep reinforcement learning and reward shaping based eco-driving control for automated hevs among signalized intersections, Energy 251 (2022). Cited by: 82

    2022

  29. [29]

    J. Li, A. Fotouhi, W. Pan, Y. Liu, Y. Zhang, Z. Chen, Deep rein- forcement learning-based eco-driving control for connected electric vehicles at signalized intersections considering traffic uncertainties, Energy 279 (2023). Cited by: 42; All Open Access, Green Open Access

    2023

  30. [30]

    D. Li, J. Zhang, S. Lin, Planning and control of drifting-based collision avoidance strategy under emergency driving conditions, Control Eng. Pract. 139 (2023) 105625

    2023

  31. [31]

    Bertipaglia, D

    A. Bertipaglia, D. Tavernini, U. Montanaro, M. Alirezaei, R. Happee, A. Sorniotti, B. Shyrokau, Model predictive contouring control for vehicle obstacle avoidance at the limit of handling using torque vectoring*, in: 2024 IEEE International Conference on Advanced Intelligent Mechatronics (AIM), 2024, pp. 1468–1475. doi: 10.1109/ AIM55361.2024.10637113

    arXiv 2024

  32. [32]

    X. Zhao, G. Chen, Z. Gao, J. Yao, Z. Gao, M. Hua, Autonomous obstacle avoidance for distributed drive electric vehicles via dynamic drifting, IEEE Trans. on Transp. Electrific. 10 (2024) 8893–8906

    2024

  33. [33]

    Kabzan, L

    J. Kabzan, L. Hewing, A. Liniger, M. N. Zeilinger, Learning-based model predictive control for autonomous racing, IEEE Robot. and Automat. Lett. 4 (2019) 3363–3370

    2019

  34. [34]

    Djeumou, T

    F. Djeumou, T. J. Lew, N. Ding, M. Thompson, M. Suminaka, M. Greiff, J. Subosits, One model to drift them all: Physics-informed conditional diffusion model for driving at the limits, in: 8th Annual Conf. on Robot Learn., 2024

    2024

  35. [35]

    Huang, Y

    Y. Huang, Y. Chen, Vehicle lateral stability control based on shiftable stability regions and dynamic margins, IEEE Trans. Veh. Technol. 69 (2020) 14727–14738

    2020

  36. [36]

    H. Lu, X. Wu, S. Zhao, L. Yan, J. Lu, Controlling nonlinear vehicular motions by exploiting linearized feedback law under delay-tolerance: stability, gain-scheduling, and validation, Meccanica (2025)

    2025

  37. [37]

    Ajanovi, E

    Z. Ajanovi, E. Regolin, B. Shyrokau, H. ati, M. Horn, A. Ferrara, Search-based task and motion planning for hybrid systems: Agile autonomous vehicles, Eng. Appl. of Artif. Intell. 121 (2023) 105893

    2023

  38. [38]

    J. Li, X. Wu, X. Bai, Y. Liu, M. Xu, Intelligent eco-driving control for urban cavs using a model-based controller assisted deep reinforce- ment learning, IEEE Trans. Intell. Transp. Syst. 26 (2025) 7624–7639

    2025

  39. [39]

    J. Li, X. Wu, J. Fan, Y. Liu, M. Xu, Overcoming driving challenges in complex urban traffic: A multi-objective eco-driving strategy via safety model based reinforcement learning, Energy 284 (2023). Cited by: 23

    2023

  40. [40]

    J. Li, X. Wu, M. Xu, Y. Liu, Multiobjective eco-driving strategy for connected and automated electric vehicles considering complex urban traffic influence factors, IEEE Transactions on Transportation Electrification 10 (2024) 10043 10058. Cited by: 5

    2024

  41. [41]

    Haarnoja, A

    T. Haarnoja, A. Zhou, P. Abbeel, S. Levine, Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochas- tic actor, in: J. Dy, A. Krause (Eds.), Proc. of the 35th Int. Conf. on Mach. Learn., volume 80 of Proc. of Mach. Learn. Res., PMLR, 2018, pp. 1861–1870

    2018

  42. [42]

    Huang, H

    W. Huang, H. Liu, Z. Huang, C. Lv, Safety-aware human-in-the-loop reinforcement learning with shared control for autonomous driving, IEEE Trans. Intell. Transp. Syst. 25 (2024) 16181–16192

    2024

  43. [43]

    P. E. Gill, W. Murray, M. A. Saunders, Snopt: An sqp algorithm for large-scale constrained optimization, SIAM Rev. 47 (2005) 99–131

    2005

  44. [44]

    Dosovitskiy, G

    A. Dosovitskiy, G. Ros, F. Codevilla, A. Lopez, V . Koltun, CARLA: An open urban driving simulator, in: S. Levine, V . Vanhoucke, K. Goldberg (Eds.), Proc.s of the 1st Annu. Conf. on Robot Learn., volume 78 of Proceedings of Mach. Learn. Res. , PMLR, 2017, pp. 1–16. Sheng Zhao et al.: Preprint submitted to Elsevier Page 15 of 15

    2017

  45. [3935]

    doi:10.1109/ITSC58415.2024.10919509

    arXiv 2024