Pith. sign in

REVIEW 3 major objections 5 minor 34 references

Learning-based Autonomous Oversteer Control and Collision Avoidance

T0 review · 3 major / 5 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read This paper claims that a hybrid imitation-reinforcement algorithm can train a car to recover from an oversteer skid and dodge obstacles using only imperfect human demonstrations.

desk verdict A useful oversteer benchmark and a genuine 81.8% empirical result, but the 'near-optimal 99.0%' headline rests on an unsupported post-hoc exclusion of failures. read the letter →

arxiv 2505.15275 v1 pith:LPL74FWA submitted 2025-05-21 cs.RO cs.AIcs.LG

Seokjun Lee , Seung-Hyun Kong This is my paper · ORCID
classification cs.ROcs.AIcs.LG
keywords autonomousdrivingoversteercontrolcollisionavoidancesoftactor-critichybridlearningimitationreinforcementend-to-end
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

The paper tries to establish that a car can be taught, end to end, to handle a sudden oversteer skid and simultaneously steer around obstacles, using only imperfect human demonstrations instead of expert trajectories. It proposes QC-SAC, a hybrid imitation-and-reinforcement learning algorithm that keeps only the demonstration actions judged better than the current policy by a learned Q-function, trains that Q-function on both experience and demonstrations, and continually adds its own successful episodes to the demonstration set. On a kick-plate benchmark with random obstacles, QC-SAC reaches an 81.8% raw success rate; excluding failures the authors count as physically impossible, it reaches 99.0%, which they describe as near-optimal. The point of caring: oversteer is a common cause of fatal accidents, and prior autonomous methods either track predefined paths or assume empty roads.

What carries the argument

The core mechanism is QC-SAC, a hybrid of Soft Actor-Critic and behavior cloning. Its Q-Compared Objective computes a weighting factor $C(s_d, a_d) = \max(Q^-(s_d, a_d) - Q(s_d, \pi_\phi(s_d)), 0)$ that multiplies the behavior-cloning loss for each demonstration: a demonstration action is imitated only when its estimated Q-value exceeds that of the policy's own action in the same state, and the weight grows with the quality gap. Q-Network from Demonstration (QNfD) augments every Q-network update with a batch of demonstration transitions, so the critic that drives that comparison is well trained even when successful interactions are rare. Selective Demonstration Data Update (SDDU) appends any new episode whose total reward beats the running average of the demonstration set, letting the agent expand its own demonstration library with fresh successes and avoid overfitting to a fixed small dataset. Together these pieces let the policy learn from suboptimal human data while still improving through reinforcement.

What would settle it

Re-run the 87 failure episodes the paper classifies as physically impossible with an oracle planner that has full knowledge of obstacle positions and the same 700 degrees-per-second steering-rate limit; if any one is completed without collision, the 99.0% near-optimal claim is falsified. A formal reachable-set computation over the benchmark's obstacle placements would settle the question without rerunning learning.

Watch

Extended reading notes

Core claim

The authors claim that QC-SAC is the first end-to-end driving policy that both stabilizes an oversteering vehicle and avoids obstacles, and that it achieves this from immature demonstrations. Concretely, in 500 randomized test episodes with up to two obstacles on a three-lane road after a kick-plate disturbance, the QC-SAC policy succeeds 409 times (81.8%), compared with 34.6% for BC-SAC, 19.0% for SAC, and 0.0% for behavior cloning. Of the 91 failures, the authors classify 87 as collisions that were physically impossible to avoid because obstacles blocked the skid direction, leaving 99.0% success among avoidable episodes; they take this as evidence of near-optimal performance. The ablation study attributes the gain to all three components of QC-SAC, with the Q-network-from-demonstration component the most critical.

Load-bearing premise

The paper's near-optimal claim rests on the assumption that the 87 collision episodes it labels physically impossible truly have no collision-free path within the action limits, and that label was assigned after the fact rather than proved by reachability analysis.

Editorial extensions

If this is right

  • An end-to-end policy can stabilize a skidding vehicle and plan an evasive maneuver at the same time, so the vehicle does not need a precomputed dynamically feasible trajectory.
  • Imperfect human demonstrations, of the kind available for rare safety-critical events, can be enough to train a strong policy if the learning rule filters them by estimated action quality.
  • Enforcing a steering-rate limit through the action representation prevents physically impossible steering commands without breaking the Markov property.
  • The benchmark itself gives future work a standardized way to measure oversteer control with collision avoidance, including a human success rate as a reference.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • If the Q-comparison mechanism is sound, it should transfer to other hybrid IL/RL settings where demonstrations are noisy but not useless, such as teleoperated manipulation, provided a reliable Q-estimate can be learned.
  • The 99.0% figure is not a directly measured rate but an estimate that depends on classifying 87 failure episodes as unavoidable; a formal reachability analysis over the benchmark's obstacle placements would convert it into a verified bound.
  • Because the human demonstrators succeeded only about 15% of the time in the same task, QC-SAC appears to be doing more than imitating humans; comparing it to a model-predictive controller with the same action limits would show how much of the gain comes from the learning rule versus task geometry.
  • The benchmark design, with obstacles placed 30 to 70 meters ahead and one fewer obstacle than lanes, creates a distribution with an inherent collision floor; reporting success conditioned on the existence of a feasible path would make future comparisons cleaner.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 5 minor

Summary. The paper proposes QC-SAC, a hybrid imitation/reinforcement learning algorithm for end-to-end autonomous control of a vehicle during oversteer with obstacle avoidance. The authors introduce a simulator benchmark based on a virtual kick plate, collect suboptimal human demonstrations, and compare QC-SAC against BC, SAC, and BC-SAC. They report a raw success rate of 81.8% (409/500 episodes) and, after excluding 87 of 91 failures as 'physically impossible', claim a 99.0% near-optimal success rate. The paper also provides an ablation study of the three main components: QCO, QNfD, and SDDU.

Significance. If the near-optimal claim were properly supported, this would be a substantive advance for safety-critical autonomous driving at the handling limits. The benchmark itself is a useful contribution, the comparison to human performance (approximately 15% success) is informative, and the raw 81.8% success rate represents a large improvement over the best baseline (BC-SAC at 34.6%). The ablation study cleanly demonstrates that all three proposed components contribute to the training reward. However, the central '99.0% near-optimal' claim is not substantiated by the evidence presented, and the theoretical justification of SDDU conflates learned and true Q-functions. With the current presentation, the paper's strongest defensible result is the raw success rate, not the headline near-optimality claim.

major comments (3)
  1. [Section V, Table I] The 99.0% success rate and 'near-optimal' claim depend entirely on excluding 87 of the 91 failure episodes as 'physically impossible' without any formal reachability or controllability analysis. The manuscript provides only a qualitative description — obstacles 'completely blocking the path in the direction of the vehicle's skid' — and example videos. There is no definition of 'blocking the path,' no search over admissible action sequences (pedal in [-1,1], steering rate bounded at 700°/s), no forward simulation of the excluded episodes, and no reachable-set computation. If even a small fraction of those 87 episodes are actually avoidable, the success rate on avoidable episodes is overstated. The raw 409/500 (81.8%) is a valid result, but it does not by itself justify the 99.0% figure. Please either add a rigorous infeasibility analysis (e.g., reachable sets or an exhaustive search over a discretized action space) or remove the 99.0% claim and report the raw success rate with appropriate uncertainty.
  2. [Section III-C, Eq. (10) and following] The SDDU theoretical justification assumes that the learned Q-function equals the true Q-function. The argument 'from Q(s,a)=E[Σγ^k r], it follows that E_{(s,a)∼D'}[Q(s,a)] > E_{(s,a)∼D}[Q(s,a)]' is valid only for the true Q; the Q-networks used in Eq. (7) are learned approximators and need not preserve this inequality. Furthermore, SDDU selects episodes with total episode reward higher than the average episode reward of the demonstration set, but Eq. (10) and the subsequent argument are about per-state-action reward expectations. A longer episode with lower per-step reward can have a higher total reward, so the claimed monotonicity of E[C(s,a)] does not follow. The mechanism may be a useful heuristic, but the current derivation is not a proof. Please revise the theoretical claim to a heuristic motivation or state conditions under which the learned Q preserves the needed ordering.
  3. [Section V] No confidence intervals or statistical significance tests are reported for the success rates in Table I. With 500 episodes, the 95% Wilson interval for the raw 81.8% is approximately [78.2%, 85.0%]; the 99.0% figure is conditional on the post-hoc exclusion and has even larger uncertainty. The paper also does not state whether the 500 test episodes are aggregated over the five training seeds shown in Fig. 6 or a single seed; the training-curve standard deviations suggest substantial seed-to-seed variance. Please report per-seed test success rates and confidence intervals, or at least explicitly state the number of seeds used for the reported success rates.
minor comments (5)
  1. [Section IV-A and footnote 2] The main text says the obstacle distances are 'closer than the legally recommended safe distance,' but the footnote says the corresponding TTC values (1.54 s to 3.60 s) are 'within the legally defined safe distances.' These statements are contradictory; please clarify the intended comparison.
  2. [Section IV-C] The statement that 'the current state s_t is influenced only by the previous state s_{t-1} and the previous action a_{t-1}' is not the formal Markov property used in RL, which concerns the transition probability p(s'|s,a). The argument for including the steering angle in the state to constrain steering rate is sound, but it should be phrased in terms of the transition dynamics.
  3. [Section III-A, Eq. (7)] In Eq. (7), the target Q-function Q^- is used for the demonstration action while the current Q is used for the policy action. The text does not explain this asymmetric use of target versus current Q; please clarify the rationale.
  4. [Section V, human driver comparison] The human driver success rate of about 15% is mentioned without details of the test protocol or number of runs. Please include it in Table I or add a footnote describing the human test condition, including the number of trials and the driver's familiarity with the task.
  5. [Throughout] There are minor typographical errors, e.g., 'To our knowledege' in Section II-A and 'overstreer' in Section IV-D. Also, the abbreviations 'QCO,' 'QNfD,' and 'SDDU' are not expanded in the abstract; consider mentioning them there or in the introduction.

Circularity Check

2 steps flagged · score 6.0 of 10

The 'near-optimal' 99.0% success rate is an artifact of a post hoc exclusion of 87 failures as 'physically impossible' without reachability analysis; the SDDU derivation also assumes the learned Q equals the true Q.

  1. fitted input called prediction [Section V, Performance Evaluation, final paragraph]
    "among the 91 episodes where QC-SAC fails to avoid collision, 95.6% (87 episodes) involve obstacles completely blocking the path in the direction of the vehicle’s skid caused by the kick plate, making collision avoidance physically impossible within the given action space. Therefore, excluding unavoidable collisions, the driving policy developed via QC-SAC achieves a 99.0% success rate, demonstrating an almost optimal (i.e., near-optimal) performance."

    The 99.0% 'near-optimal' success rate is computed by removing 87 of the 91 failures, a set classified post hoc as 'physically impossible.' The paper provides no reachability or controllability analysis, no exhaustive action search, and no formal definition of 'completely blocking'; only example videos are cited. Because the classification is applied after observing the failures, the excluded set is a post hoc selection that mechanically forces the success rate to 409/413 = 99.0%. In the absence of an independent, a priori criterion, the 'near-optimal' claim is equivalent to the assumption that those 87 episodes are unavoidable.

  2. other [Section III-C, Selective Demonstration Data Update (SDDU), after Eq. (10)]
    "After the iterative process of SDDU, we obtain the updated dataset D′, which satisfies E(s,a)∼D′[r(s, a)] > E(s,a)∼D[r(s, a)]. Then, from Q(s, a) = E[P∞ k=0 γkrt+k|st = s, at = a], it follows that E(s,a)∼D′[Q(s, a)] > E(s,a)∼D[Q(s, a)]."

    The SDDU proof invokes the definition of the true action-value function to conclude that the updated demonstration set D′ has higher Q values. However, the Q used in Eq. (7) and throughout the algorithm is a learned Q-network, trained via QNfD on the demonstration set D (which SDDU subsequently updates). The step assumes the learned Q equals the true Q, an equality that is not established and is precisely what QNfD attempts to approximate using those demonstrations. Thus the justification of SDDU reduces to an assumption about the very Q-function whose training depends on the demonstrations SDDU selects. This self-referential loop makes the theoretical derivation circular, even though the empirical ablation study provides independent evidence for SDDU's value.

full rationale

The central 'near-optimal' claim (99.0% success rate) rests on removing 87 of 91 failure episodes as 'physically impossible' based on a post hoc visual classification, with no formal reachability analysis. This makes the headline result a direct arithmetic consequence of the exclusion rule, rather than an independently measured optimality property; the exclusion is equivalent to assuming the very thing it purports to establish. A second, more methodological circularity appears in the SDDU derivation: the paper uses the identity of the true Q-function to prove that the self-updated demonstration set has higher Q, while the actual Q-network is trained on that same demonstration set, so the proof presupposes the accuracy of the learned Q that QNfD is supposed to improve. The paper's raw 81.8% success rate and the comparison with BC, SAC, and BC-SAC baselines are independent empirical results, and the ablation study supports the components, so the circularity is partial rather than total. The self-citation to FER [33] is not load-bearing for the main claims. Overall score 6 reflects that one central claim reduces by construction to a post hoc fit, while other evaluation results retain independent content.

Assumptions & free parameters 4 free parameters · 4 assumptions · 0 invented entities

The central claim rests on a simulated environment, a learned Q-function used to gate demonstrations, and a post-hoc classification of failure cases. Reward shape weights and success thresholds are hand-chosen. No new physical entities are introduced.

free parameters (4)
  • Reward weights λ1, λ2, λ3 = 0.8, 0.2, 0.2
    Hand-selected weights for safe distance, progress, and auxiliary rewards; they shape the learned policy and the SDDU episode-selection threshold.
  • Reward requirement values = d=3.5m, a=2.943m/s^2, β=20°, δdot=3000°/s
    Hand-set anchors in the reward function (12); they determine what counts as good progress and stability.
  • Success/failure thresholds for episode termination = β<1° for 100 steps (success); β>37° spin or collision (failure)
    Hand-chosen terminal conditions that define the reported success rate.
  • Network training hyperparameters = not reported
    Learning rates, target update rate, entropy temperature, FER sampling distribution, and network architecture are not given, though they affect results.
assumptions (4)
  • domain assumption The IPG CarMaker simulator accurately reproduces the vehicle dynamics of oversteer, the kick-plate disturbance, and the slippery road friction reduction.
    All training and evaluation occur in this proprietary simulator; no hardware validation is provided.
  • domain assumption The learned Q-function is accurate enough that the Q-Compared Objective weighting in Eq. (7) identifies genuinely better demonstration actions.
    QCO and the SDDU justification in Section III-C rely on Q-values matching true returns.
  • ad hoc to paper The 87 excluded failure cases are truly unavoidable within the action space.
    The paper asserts these are 'physically impossible' based on obstacle positions, but provides no reachability analysis; this assumption is introduced specifically to support the near-optimal claim.
  • domain assumption The 200 human demonstration episodes, though suboptimal, cover the states needed to bootstrap the policy.
    The method depends on demonstrations being informative enough for initial Q and policy learning.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Learning-based Autonomous Oversteer Control and Collision Avoidance." pith.science (2026). https://pith.science/paper/LPL74FWA

@misc{pith2026250515275,
  author       = {Pith},
  title        = {Pith review of: Learning-based Autonomous Oversteer Control and Collision Avoidance},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/LPL74FWA}},
  note         = {Machine review of arXiv:2505.15275}
}
read the original abstract

Oversteer, wherein a vehicle's rear tires lose traction and induce unintentional excessive yaw, poses critical safety challenges. Failing to control oversteer often leads to severe traffic accidents. Although recent autonomous driving efforts have attempted to handle oversteer through stabilizing maneuvers, the majority rely on expert-defined trajectories or assume obstacle-free environments, limiting real-world applicability. This paper introduces a novel end-to-end (E2E) autonomous driving approach that tackles oversteer control and collision avoidance simultaneously. Existing E2E techniques, including Imitation Learning (IL), Reinforcement Learning (RL), and Hybrid Learning (HL), generally require near-optimal demonstrations or extensive experience. Yet even skilled human drivers struggle to provide perfect demonstrations under oversteer, and high transition variance hinders accumulating sufficient data. Hence, we present Q-Compared Soft Actor-Critic (QC-SAC), a new HL algorithm that effectively learns from suboptimal demonstration data and adapts rapidly to new conditions. To evaluate QC-SAC, we introduce a benchmark inspired by real-world driver training: a vehicle encounters sudden oversteer on a slippery surface and must avoid randomly placed obstacles ahead. Experimental results show QC-SAC attains near-optimal driving policies, significantly surpassing state-of-the-art IL, RL, and HL baselines. Our method demonstrates the world's first safe autonomous oversteer control with obstacle avoidance.

Figures

Figures reproduced from arXiv: 2505.15275 by the authors.

Figure 1
Figure 1. Vehicle oversteer and understeer. (a) Oversteer: the rear tires lose [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Research goals. (a) Ego vehicle in brown must control the oversteer [PITH_FULL_IMAGE:figures/full_fig_p001_2.png] view at source ↗
Figure 3
Figure 3. Concept diagram. Impact of the quality of demonstration data on [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figures from the paper (4 more)
Figure 4
Figure 4. Figure 4: Real-world driver training process and oversteer control and collision avoidance benchmark in a virtual environment. a, Kick plate inducing oversteer, [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Experimental setup. (a) Definition of d and ψ in vehicle state. (b) Representation of surrounding state. (c) An example of the reward function. (x¯ = 1). set by multiplying the output value by 700°/s, ensuring that the maximum steering angular velocity is 700°/s. This …
Figure 6
Figure 6. Figure 6: Training curve. The solid line represents the average reward of five instances of each training technique initialized with random seeds, while the [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]
Figure 7
Figure 7. Figure 7: Captured scenes during the test runs. Videos of the test runs using [PITH_FULL_IMAGE:figures/full_fig_p009_7.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 22 canonical work pages

  1. [1]

    Morton,How to drift: The art of oversteer

    P. Morton,How to drift: The art of oversteer. CarTech Inc, 2006

    work page 2006

  2. [2]

    Effects of electronic stability control (esc) on accidents: A review of empirical evidence,

    A. Erke, “Effects of electronic stability control (esc) on accidents: A review of empirical evidence,”Accident Analysis & Prevention, vol. 40, no. 1, pp. 167–173, 2008

    work page 2008

  3. [3]

    A controller for automated drifting along complex trajectories,

    J. Y . Goh, T. Goel, and J. C. Gerdes, “A controller for automated drifting along complex trajectories,” in14th International Symposium on Advanced Vehicle Control (AVEC 2018), vol. 7, 2018, pp. 1–6

    work page 2018

  4. [4]

    Steady-state drifting stabilization of rwd vehicles,

    E. Velenis, D. Katzourakis, E. Frazzoli, P. Tsiotras, and R. Happee, “Steady-state drifting stabilization of rwd vehicles,”Control Engineering Practice, vol. 19, no. 11, pp. 1363–1376, 2011

    2011

  5. [5]

    Drift control for cornering maneuver of autonomous vehicles,

    F. Zhang, J. Gonzales, S. E. Li, F. Borrelli, and K. Li, “Drift control for cornering maneuver of autonomous vehicles,”Mechatronics, vol. 54, pp. 167–174, 2018

    work page 2018

  6. [6]

    Autonomous drifting control in 3d car racing simulator,

    I. Zubov, I. Afanasyev, A. Gabdullin, R. Mustafin, and I. Shimchik, “Autonomous drifting control in 3d car racing simulator,” in2018 International Conference on Intelligent Systems (IS). IEEE, 2018, pp. 235–241

    work page 2018

  7. [7]

    Teaching a vehicle to autonomously drift: A data-based approach using neural networks,

    M. Acosta and S. Kanarachos, “Teaching a vehicle to autonomously drift: A data-based approach using neural networks,”Knowledge-Based Systems, vol. 153, pp. 12–28, 2018

    work page 2018

  8. [8]

    High-speed autonomous drifting with deep reinforcement learning,

    P. Cai, X. Mei, L. Tai, Y . Sun, and M. Liu, “High-speed autonomous drifting with deep reinforcement learning,”IEEE Robotics and Automa- tion Letters, vol. 5, no. 2, pp. 1247–1254, 2020

    work page 2020

Show all 34 references
  1. [9]

    Autonomous drifting using simulation-aided reinforcement learning,

    M. Cutler and J. P. How, “Autonomous drifting using simulation-aided reinforcement learning,” in2016 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2016, pp. 5442–5448

    2016

  2. [10]

    Recent advancements in end-to-end au- tonomous driving using deep learning: A survey,

    P. S. Chib and P. Singh, “Recent advancements in end-to-end au- tonomous driving using deep learning: A survey,”IEEE Transactions on Intelligent Vehicles, 2023

    2023

  3. [11]

    A survey on imitation learning techniques for end-to-end autonomous vehicles,

    L. Le Mero, D. Yi, M. Dianati, and A. Mouzakitis, “A survey on imitation learning techniques for end-to-end autonomous vehicles,” IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 9, pp. 14 128–14 147, 2022

    2022

  4. [12]

    A survey of deep rl and il for autonomous driv- ing policy learning,

    Z. Zhu and H. Zhao, “A survey of deep rl and il for autonomous driv- ing policy learning,”IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 9, pp. 14 043–14 065, 2021

    2021

  5. [13]

    Deep reinforcement learning for autonomous driving: A survey,

    B. R. Kiran, I. Sobh, V . Talpaert, P. Mannion, A. A. Al Sallab, S. Yo- gamani, and P. P ´erez, “Deep reinforcement learning for autonomous driving: A survey,”IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 6, pp. 4909–4926, 2021

    2021

  6. [14]

    Outracing champion gran turismo drivers with deep reinforcement learning,

    P. R. Wurman, S. Barrett, K. Kawamoto, J. MacGlashan, K. Subrama- nian, T. J. Walsh, R. Capobianco, A. Devlic, F. Eckert, F. Fuchset al., “Outracing champion gran turismo drivers with deep reinforcement learning,”Nature, vol. 602, no. 7896, pp. 223–228, 2022

    2022

  7. [15]

    Champion-level drone racing using deep reinforcement learning,

    E. Kaufmann, L. Bauersfeld, A. Loquercio, M. M ¨uller, V . Koltun, and D. Scaramuzza, “Champion-level drone racing using deep reinforcement learning,”Nature, vol. 620, no. 7976, pp. 982–987, 2023

    2023

  8. [16]

    A novel policy based on action confidence limit to improve exploration efficiency in reinforcement learning,

    F. Huang, X. Deng, Y . He, and W. Jiang, “A novel policy based on action confidence limit to improve exploration efficiency in reinforcement learning,”Information Sciences, vol. 640, p. 119011, 2023

    2023

  9. [17]

    An empirical study of model-free exploration for deep reinforcement learning,

    X. Zhao, “An empirical study of model-free exploration for deep reinforcement learning,” Ph.D. dissertation, University of Alberta, 2021

    2021

  10. [18]

    Deep q-learning from demonstrations,

    T. Hester, M. Vecerik, O. Pietquin, M. Lanctot, T. Schaul, B. Piot, D. Horgan, J. Quan, A. Sendonaris, I. Osbandet al., “Deep q-learning from demonstrations,” inProceedings of the AAAI conference on artifi- cial intelligence, vol. 32, 2018

    2018

  11. [19]

    Learning complex dexterous manipulation with deep reinforcement learning and demonstrations,

    A. Rajeswaran, V . Kumar, A. Gupta, G. Vezzani, J. Schulman, E. Todorov, and S. Levine, “Learning complex dexterous manipulation with deep reinforcement learning and demonstrations,”arXiv preprint arXiv:1709.10087, 2017

    2017 arXiv

  12. [20]

    Residual reinforcement learning from demonstrations,

    M. Alakuijala, G. Dulac-Arnold, J. Mairal, J. Ponce, and C. Schmid, “Residual reinforcement learning from demonstrations,”arXiv preprint arXiv:2106.08050, 2021

    2021 arXiv

  13. [21]

    Learning to drive like human beings: A method based on deep reinforcement learning,

    Y . Tian, X. Cao, K. Huang, C. Fei, Z. Zheng, and X. Ji, “Learning to drive like human beings: A method based on deep reinforcement learning,”IEEE Transactions on Intelligent Transportation Systems, vol. 23, no. 7, pp. 6357–6367, 2021

    2021

  14. [22]

    Imitation is not enough: Ro- bustifying imitation with reinforcement learning for challenging driving scenarios,

    Y . Lu, J. Fu, G. Tucker, X. Pan, E. Bronstein, R. Roelofs, B. Sapp, B. White, A. Faust, S. Whitesonet al., “Imitation is not enough: Ro- bustifying imitation with reinforcement learning for challenging driving scenarios,” in2023 IEEE/RSJ International Conference on Intelligen...

    2023

  15. [23]

    Re- inforcement learning from imperfect demonstrations,

    Y . Gao, H. Xu, J. Lin, F. Yu, S. Levine, and T. Darrell, “Re- inforcement learning from imperfect demonstrations,”arXiv preprint arXiv:1802.05313, 2018

    2018 arXiv

  16. [24]

    An aggressive cornering framework for autonomous vehicles combining trajectory planning and drift control,

    W. Weng, C. Hu, Z. Li, H. Su, and L. Xie, “An aggressive cornering framework for autonomous vehicles combining trajectory planning and drift control,” in2024 IEEE Intelligent Vehicles Symposium (IV). IEEE, 2024, pp. 2749–2755

    2024

  17. [25]

    The magic formula tyre model,

    H. B. Pacejka and E. Bakker, “The magic formula tyre model,”Vehicle system dynamics, vol. 21, no. S1, pp. 1–18, 1992

    1992

  18. [26]

    Environmental effects on pacejka’s scaling factors,

    F. Braghin, F. Cheli, and E. Sabbioni, “Environmental effects on pacejka’s scaling factors,”Vehicle System Dynamics, vol. 44, no. 7, pp. 547–568, 2006

    2006

  19. [27]

    Vision-based au- tonomous car racing using deep imitative reinforcement learning,

    P. Cai, H. Wang, H. Huang, Y . Liu, and M. Liu, “Vision-based au- tonomous car racing using deep imitative reinforcement learning,”IEEE Robotics and Automation Letters, vol. 6, no. 4, pp. 7262–7269, 2021

    2021

  20. [28]

    Improved reinforcement learning through imitation learning pretraining towards image-based autonomous driv- ing,

    T. Wang and D. E. Chang, “Improved reinforcement learning through imitation learning pretraining towards image-based autonomous driv- ing,” in2019 19th international conference on control, automation and systems (ICCAS). IEEE, 2019, pp. 1306–1310

    2019

  21. [29]

    Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,

    T. Haarnoja, A. Zhou, P. Abbeel, and S. Levine, “Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,” inInternational conference on machine learning. PMLR, 2018, pp. 1861–1870

    2018

  22. [30]

    Soft actor- critic algorithms and applications,

    T. Haarnoja, A. Zhou, K. Hartikainen, G. Tucker, S. Ha, J. Tan, V . Kumar, H. Zhu, A. Gupta, P. Abbeel, and S. Levine, “Soft actor- critic algorithms and applications,” 2019

    2019

  23. [31]

    Efficient reinforcement learning for autonomous driving with parameterized skills and priors,

    L. Wang, J. Liu, H. Shao, W. Wang, R. Chen, Y . Liu, and S. L. Waslander, “Efficient reinforcement learning for autonomous driving with parameterized skills and priors,”arXiv preprint arXiv:2305.04412, 2023

    2023 arXiv

  24. [32]

    Double q-learning,

    H. Hasselt, “Double q-learning,”Advances in neural information pro- cessing systems, vol. 23, 2010

    2010

  25. [33]

    Enhanced off-policy reinforcement learning with focused experience replay,

    S.-H. Kong, I. M. A. Nahrendra, and D.-H. Paek, “Enhanced off-policy reinforcement learning with focused experience replay,”IEEE Access, vol. 9, pp. 93 152–93 164, 2021

    2021

  26. [34]

    Modification of vehicle handling characteristics via steer-by-wire,

    P. Yih and J. C. Gerdes, “Modification of vehicle handling characteristics via steer-by-wire,”IEEE transactions on control systems technology, vol. 13, no. 6, pp. 965–976, 2005. LEEet al.: LEARNING-BASED AUTONOMOUS OVERSTEER CONTROL AND COLLISION A VOIDANCE 11 Seokjun LeeSeokj...

    2005

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.