Pith. sign in

REVIEW 3 major objections 31 references

Expert racing lines and vehicle-dynamics safety envelopes let reinforcement learning race faster while staying stable.

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 · grok-4.5

2026-07-15 14:13 UTC pith:J7Y77Y75

load-bearing objection Solid engineering combo of MCRL grid guidance, dual CBF soft costs, and curriculum that beats TAL/PPO/DDPG on Tempelhof sim; the linear-tire envelope is the softest link, not a collapse of the sim claim. the 3 major comments →

arxiv 2603.05842 v2 pith:J7Y77Y75 submitted 2026-03-06 cs.RO

Expert Knowledge-driven Reinforcement Learning for Autonomous Racing via Trajectory Guidance and Dynamics Constraints

Bo Leng , Weiqi Zhang , Zhuoren Li , Lu Xiong , Guizhe Jin , Ran Yu , Chen Lv This is my paper
classification cs.RO
keywords Autonomous RacingReinforcement LearningRacing Line GuidanceDynamics ConstraintsCurriculum LearningControl Barrier Functions
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.

Pure reinforcement learning is a poor fit for high-speed racing: exploration is inefficient, rewards are sparse, and trial-and-error can produce unsafe yaw and sideslip. TraD-RL injects expert knowledge three ways: a precomputed minimum-curvature racing line augments the agent’s observation and shapes dense tracking rewards; control-barrier functions built from yaw-rate and sideslip limits act as soft physics constraints via adaptive Lagrange multipliers; and a two-stage curriculum first teaches expert-speed following then lifts the speed cap so the policy can exceed the expert baseline. On a high-fidelity Tempelhof Airport Street Circuit simulation the method posts higher average lap speed and lower lap time than DDPG, PPO, and a trajectory-aided baseline while reducing time-averaged stability violations. A reader who wants autonomous racing that is both competitive and physically plausible cares because the approach shows performance and safety need not be traded off when domain priors are embedded correctly.

Core claim

TraD-RL shows that embedding a minimum-curvature racing line into state and reward, enforcing CBF soft constraints on yaw rate and sideslip, and training with a guidance-to-exploration curriculum produces a racing policy that simultaneously raises lap speed and keeps the vehicle inside a safe operating envelope, outperforming pure RL and trajectory-aided baselines on the Tempelhof circuit.

What carries the argument

TraD-RL: an RL loop that (1) augments the ego-centric occupancy grid and reward with a minimum-curvature racing line, (2) regularizes the policy with CBF costs on yaw rate and sideslip via dual ascent on Lagrange multipliers, and (3) switches curriculum from expert-speed tracking to maximum-speed exploration.

Load-bearing premise

The safe envelope built from a linear bicycle model and simplified sideslip limits is accurate enough to keep the car stable when tires actually run deep in the nonlinear friction regime.

What would settle it

Deploy the trained policy on a nonlinear tire model (or a real chassis) and check whether yaw rate and sideslip still stay inside the claimed envelope while lap times remain competitive; large unmodeled violations or a forced collapse in speed would falsify the claim.

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

If this is right

  • Early training reaches reliable full-lap completion once trajectory guidance and constraints are active.
  • The learned policy can beat the expert racing line’s lap time without abandoning the safe envelope.
  • Sideslip and yaw-rate violations fall relative to unconstrained high-speed baselines without forcing an overly conservative speed trade-off.
  • Ablations show both the racing-line module and the dynamics-constraint module are required for the joint speed-and-safety gain.

Where Pith is reading between the lines

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

  • The same guidance-plus-CBF pattern could transfer to other limit-handling tasks such as high-speed emergency evasion on public roads.
  • If the linear envelope underestimates tire saturation, real-vehicle use would need online envelope adaptation or a nonlinear prior.
  • Multi-car racing would require extending the grid observation and the barrier set to other agents, not only track geometry.

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

3 major / 0 minor

Summary. The paper proposes TraD-RL, a soft-constrained RL controller for autonomous racing that embeds expert priors in three ways: (i) a Minimum Curvature Racing Line (MCRL) used both as an occupancy-grid observation channel and for dense tracking/heading/speed rewards; (ii) CBF-style soft constraints on yaw rate and sideslip, derived from a linear bicycle model and a β–ω envelope, enforced via dual cost critics and adaptive Lagrangian multipliers with ReLU dead-zones; and (iii) a two-stage curriculum that first tracks MCRL reference speed then switches to a maximum-speed objective so the agent can exceed the prior. On a Tempelhof Airport Street Circuit simulation the method reports higher lap average speed and lower lap time than DDPG, PPO, and TAL, with fewer time-averaged ω/β envelope violations than DDPG and TAL, and ablations attribute the gains to the trajectory-guidance and dynamics-constraint modules.

Significance. If the results hold under broader validation, the work is a useful engineering contribution to safe high-speed RL: it shows a practical recipe for combining geometric racing-line priors, soft CBF costs, and curriculum to improve both lap time and envelope adherence relative to strong baselines (including trajectory-aided TAL). Strengths include clear ablations (w/o TG, w/o DC), multi-metric training curves, a continuous-corner case study, and an explicit algorithm listing. The contribution is incremental rather than foundational—soft Lagrangian CBF costs and racing-line guidance are known ingredients—but the integrated package and the reported performance–safety trade-off on a realistic street circuit are of interest to the autonomous-racing community.

major comments (3)
  1. The safety interpretation is overstated relative to the model used to build the prior. Section 2.2 and Eqs. (4)–(6) adopt a Dynamic Bicycle Model with linear cornering stiffness (Table 1: C_αf = C_αr = 64 kN/rad); the CBF envelopes in Eqs. (17)–(23) and Fig. 7 then use |ω| ≤ μg/u and a Pacejka-derived α_r,peak under that linearization. At the reported speeds (entry ~70 m/s, mid-corner ~30 m/s) tires operate deep in the nonlinear friction regime the linear model excludes. Soft penalties on a potentially mis-specified h neither guarantee plant stability nor correctly trade speed for safety—the regime where Table 3 credits the DC module. The sim claim can still hold inside the authors’ plant, but the paper should either (a) restate claims as “envelope adherence under the linear prior” rather than physics-informed limit-handling safety, or (b) re-derive/validate the envelope against a nonlin
  2. Generalization evidence is thin for the central claim of synergistic performance and safety. All results (Figs. 10–13, Tables 2–3) use a single track (Tempelhof) and a single vehicle parameter set (Table 1), in simulation only. Curriculum switch T_switch, cost thresholds d_ω/d_β, CBF gain α, and reward band edges are free parameters with no sensitivity study. At minimum the authors should report multi-seed statistics (mean ± std over ≥3 seeds) for Table 2, and ideally one additional track or a friction/mass perturbation, so that the claimed gains over TAL are not confounded by track-specific tuning of the MCRL prior and the dual multipliers.
  3. The constrained objective and cost evaluation leave important implementation details unspecified, which affects reproducibility of the safety results. Eqs. (22)–(23) define instantaneous CBF costs, then a sliding-window average is mentioned without window length or how ḡ is obtained from the discrete plant; Eq. (28) uses ReLU(max_j Q_Ck − d_k) with dual critics, but d_ω, d_β, α, and the window size never appear numerically. Without these values (or a sensitivity plot), it is hard to judge whether the reported reduction in time-averaged ω/β violations (Table 2) is robust or an artifact of a particular dead-zone width. Please list the numerical safety hyperparameters and clarify discrete-time CBF evaluation.

Circularity Check

1 steps flagged

Empirical RL paper with independent external benchmarks; only minor self-reference is that DC safety metrics count the same envelope the Lagrangian already penalizes.

specific steps
  1. self definitional [Sec. 3.3 Eqs. (20)–(23) and Sec. 4.3.2 safety metrics / Table 2–3]
    "we define the instantaneous calculation formulas for the yaw rate cost c_ω and the sideslip angle cost c_β as follows: c_ω = max(−(ḣ_ω + α h_ω), 0) … c_β = max(− min(ḣ_β1 + α h_β1, ḣ_β2 + α h_β2), 0). … Time-averaged Lap ω-unsafe Times: The number of times the vehicle violates the yaw rate dynamic boundaries (i.e., boundaries 1 and 3 in Fig. 7) … Time-averaged Lap β-unsafe Times: The number of times the vehicle violates the center-of-mass sideslip angle dynamic boundaries (i.e., boundaries 2 and 4 in Fig. 7)"

    The DC module’s training objective is soft Lagrangian penalization of CBF violations of the β–ω envelope. The reported safety metrics count crossings of that same envelope. Therefore the ablation result that w/ DC has fewer time-averaged ω/β-unsafe times than w/o DC is partly true by construction of the objective, not an independent discovery of stability. (Lap time/speed and baseline comparisons remain non-circular.)

full rationale

TraD-RL is an engineering/empirical RL method paper, not a first-principles derivation. The load-bearing claims are measured lap time, average speed, and constraint-violation counts against external baselines (PPO, DDPG, TAL) and ablations on a fixed Tempelhof sim. MCRL is an explicit expert input used for early guidance and reward shaping; the two-stage curriculum deliberately replaces target-speed tracking with a max-velocity objective so the policy can beat that prior—so expert performance is not the claimed optimum by construction. CBF costs and Lagrangian multipliers are soft training regularizers, not uniqueness theorems or fitted parameters renamed as predictions. No self-citation uniqueness chain, no ansatz smuggled as external math, and no renaming of a known empirical law. The only mild circularity is that the paper’s primary safety metrics (time-averaged ω/β-unsafe times) are defined as crossings of the same β–ω envelope used to build the CBF costs; thus the ablation claim “adding DC reduces unsafe times” is partly forced by the training objective. That does not force the independent racing-performance gains or the comparisons to baselines that never saw those costs. Score 2 is proportionate.

Axiom & Free-Parameter Ledger

5 free parameters · 6 axioms · 2 invented entities

The central empirical claim rests on standard MDP/RL and vehicle-dynamics background, plus several modeling and training choices that are free or paper-specific: linear bicycle tires for the plant, a hand-designed safe envelope and cost thresholds, a fixed curriculum switch step, and a particular dense reward mix. No new physical entity is postulated; the 'invented' pieces are algorithmic constructs (TraD-RL modules).

free parameters (5)
  • Curriculum switch step T_switch = 200000 steps
    Hard-coded at 200000 of 250000 training steps; stage transition is not learned and directly shapes whether the agent can beat MCRL speeds.
  • Safety cost thresholds d_ω, d_β
    Allowable expected CBF violation levels in the constrained objective (Eq. 24); values are design choices that set the speed–safety trade-off.
  • CBF class-K gain α and sliding-window cost averaging
    Linear α and windowed averaging of c_ω, c_β are chosen for training stability; they scale how strictly the soft constraints bite.
  • Reward component weights / normalization and speed band edges (u_h1, u_h2, u_l1, u_l2)
    Multi-term reward (track, high/low speed, lap, MCRL tracking, heading, target speed) is hand-shaped and normalized to [-1,1]; band edges are not derived from first principles.
  • Lagrangian and network learning rates, entropy coefficient schedule = lr_actor/critic=1e-4; lr_λ=5e-5
    Actor/critic lr 1e-4, λ lr 5e-5, γ=0.99, buffer 200k, CNN/MLP widths—standard but claim-sensitive hyperparameters.
axioms (6)
  • domain assumption Racing control can be cast as a discounted MDP with continuous actions (a, δ) and ego-centric grid observations.
    Section 2.1 and 3.1; standard RL modeling choice that ignores continuous-time optimality certificates.
  • domain assumption Lateral/yaw dynamics are adequately captured by a bicycle model with small-angle kinematics and linear tire forces F=C_α α inside the training plant.
    Section 2.2 Eqs. (4)–(6); used despite the paper’s own emphasis on nonlinear limit handling.
  • domain assumption Yaw-rate and sideslip bounds (Eqs. 17–19) define a meaningful safe operating envelope enforceable via CBF inequalities (20)–(23).
    Section 3.3 citing [25–27]; soft Lagrangian enforcement does not yield hard safety guarantees.
  • domain assumption Minimum-curvature path plus GGV-based speed profile is a useful expert prior for observation augmentation and early rewards.
    Section 3.2; MCRL optimality is geometric, not proven time-optimal for the learned closed-loop dynamics.
  • ad hoc to paper Soft constrained policy optimization with dual cost critics and ReLU dead-zones (Eqs. 25–29) sufficiently enforces dynamics priors during exploration.
    Section 3.3–3.4; algorithmic design choice specific to TraD-RL’s training loop.
  • domain assumption Tempelhof high-fidelity simulation results are informative about racing performance and stability of the method.
    Section 4; no real-vehicle or multi-sim transfer evidence.
invented entities (2)
  • TraD-RL framework (MCRL-augmented grid + dual CBF cost critics + two-stage curriculum) no independent evidence
    purpose: Package expert trajectory guidance and dynamics soft constraints into one racing RL agent.
    Named method of the paper; algorithmic composition rather than a new physical object. Independent evidence is only the reported sim metrics.
  • Reference-augmented occupancy observation o_MCRL no independent evidence
    purpose: Inject racing-line geometry into the CNN policy input as a binary grid channel.
    Paper-specific state design (Fig. 6); no external validation beyond this study’s training curves.

pith-pipeline@v1.1.0-grok45 · 22509 in / 4038 out tokens · 41981 ms · 2026-07-15T14:13:21.712730+00:00 · methodology

0 comments
read the original abstract

Reinforcement learning has demonstrated significant potential in the field of autonomous driving. However, it suffers from defects such as training instability and unsafe action outputs when faced with autonomous racing environments characterized by high dynamics and strong nonlinearities. To this end, this paper proposes a trajectory guidance and dynamics constraints Reinforcement Learning (TraD-RL) method for autonomous racing. The key features of this method are as follows: 1) leveraging the prior expert racing line to construct an augmented state representation and facilitate reward shaping, thereby integrating domain knowledge to stabilize early-stage policy learning; 2) embedding explicit vehicle dynamic priors into a safe operating envelope formulated via control barrier functions to enable safety-constrained learning; and 3) adopting a multi-stage curriculum learning strategy that shifts from expert-guided learning to autonomous exploration, allowing the learned policy to surpass expert-level performance. The proposed method is evaluated in a high-fidelity simulation environment modeled after the Tempelhof Airport Street Circuit. Experimental results demonstrate that TraD-RL effectively improves both lap speed and driving stability of the autonomous racing vehicle, achieving a synergistic optimization of racing performance and safety.

Figures

Figures reproduced from arXiv: 2603.05842 by Bo Leng, Chen Lv, Guizhe Jin, Lu Xiong, Ran Yu, Weiqi Zhang, Zhuoren Li.

Figure 1
Figure 1. Figure 1: The reinforcement learning decision-making and control framework driven by expert prior knowledge.. 2. Preliminaries 2.1. Reinforcement Learning Reinforcement learning is typically mathematically for￾malized as a Markov Decision Process (MDP), defined by the tuple  = ⟨, , , , 𝛾⟩. Here,  denotes the state space,  represents the action space,  ∶ ×× → [0, 1] is the state transition probability func… view at source ↗
Figure 3
Figure 3. Figure 3: Schematic diagram of the ego-centric grid observa￾tion space construction. 3.1. RL Racing Framework Observation Space This paper adopts an ego-centric occupancy grid map as the fundamental representation of the observation space in order to describe the unstructured road features of the racetrack. The observation vector consists of two parts: environ￾mental geometric features and vehicle kinematic states, … view at source ↗
Figure 4
Figure 4. Figure 4: Parameterization of the optimal racing line relative to the track centerline. Low-Speed Reward: When the speed falls within the low-speed range [𝑢𝑙1 , 𝑢𝑙2 ], a penalty inversely proportional to speed is applied. 𝑟𝑙𝑠 = 1 − 𝑢 𝑢 2 𝑙2 (9) Lap Completion Reward: To incentivize the agent to successfully complete a full lap, a discrete terminal reward 𝑟lap is granted whenever the vehicle crosses the finish line. … view at source ↗
Figure 6
Figure 6. Figure 6: Schematic illustration of mapping the racing line prior into the occupancy grid. grid. Specifically, grid cells covering the racing line way￾points take the value 1, while the remaining areas are 0. The incorporation of this expert prior information significantly enhances training efficiency. It not only assists the agent in rapidly comprehending the trajectory guidance mechanism but also effectively compr… view at source ↗
Figure 8
Figure 8. Figure 8: Aerial view of the Berlin Tempelhof Airport Street Circuit (Formula E Berlin E-Prix). (a) The Actor network architecture for action distribution output. (b) The Critic (and Cost-Critic) network architecture for value estimation [PITH_FULL_IMAGE:figures/full_fig_p008_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Schematic diagram of the proposed network archi￾tectures. and fed into an MLP to estimate the Q-value of the state￾action pair. The detailed experimental parameters are as follows: • Training Scale: The total number of training steps is set to 𝑇max = 250, 000. The agent begins the learning process after an initial exploration phase of 𝑇start = 20, 000 steps. The two-stage training step threshold, 𝑇𝑠𝑤𝑖𝑡𝑐ℎ =… view at source ↗
Figure 10
Figure 10. Figure 10: Learning curves of multi-dimensional performance metrics for different algorithms during the training process: (a) Time-averaged lap reward; (b) Lap time; (c) Lap average speed; (d) Lap progress. times, constitutes a core indicator for evaluating the vehi￾cle’s dynamic stability at handling limits. • Lap Progress: The percentage of the track completed before the vehicle collides or runs off the track; if … view at source ↗
Figure 11
Figure 11. Figure 11: Statistical distribution of vehicle dynamics states for different algorithms during testing: (a) Yaw rate; (b) Sideslip angle. efficiently along the boundaries of the safe envelope. In con￾trast, while the PPO algorithm achieves the lowest violation counts, this is entirely attributed to its overly conservative driving policy that sacrifices racing performance to avoid instability. However, such a strateg… view at source ↗
Figure 12
Figure 12. Figure 12: Experimental results comparison in a continuous corner (S-curve) section of the Berlin Tempelhof Airport Street Circuit. Left: Trajectory and speed heat map distributions of TAL and the proposed method (Ours), with the red solid line representing the MCRL. Right: Time-series comparison of vehicle speed, yaw rate, and sideslip angle during the cornering process. speed control commands, demonstrating a more… view at source ↗
Figure 13
Figure 13. Figure 13: Performance and safety comparison of different ablation algorithms during the testing phase: (a) Box plot of lap time; (b) Box plot of average lap speed; (c) Statistical distribution of yaw rate; (d) Statistical distribution of sideslip angle. proposed method exhibits the most compact distribution for both lap time and lap speed. This implies that the dynamics constraints module not only guarantees absolu… 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

31 extracted references · 2 linked inside Pith

  1. [1]

    Muraleedharan, I

    A. Muraleedharan, I. Simões, G. Pau, Randomized model predictive control for autonomous racing, in: 2025 IEEE Intelligent Vehicles Symposium (IV), IEEE, 2025, pp. 1078–1085

    2025

  2. [2]

    Kalaria, C

    D. Kalaria, C. Maheshwari, S. Sastry,{∖alpha}-racer: Real-time algorithm for game-theoretic motion planning and control in au- tonomous racing using near-potential function, arXiv preprint arXiv:2412.08855 (2024)

    Pith/arXiv arXiv 2024

  3. [3]

    Langmann, L

    A. Langmann, L. Ögretmen, F. Werner, J. Betz, Online veloc- ity profile generation and tracking for sampling-based local plan- ning algorithms in autonomous racing environments, arXiv preprint arXiv:2505.05157 (2025)

    Pith/arXiv arXiv 2025

  4. [4]

    J. Betz, T. Betz, F. Fent, M. Geisslinger, A. Heilmeier, L. Hermans- dorfer, T. Herrmann, S. Huch, P. Karle, M. Lienkamp, et al., Tum autonomousmotorsport:Anautonomousracingsoftwarefortheindy autonomous challenge, Journal of Field Robotics 40 (4) (2023) 783– 809

    2023

  5. [5]

    Wischnewski, M

    A. Wischnewski, M. Geisslinger, J. Betz, T. Betz, F. Fent, A. Heilmeier, L. Hermansdorfer, T. Herrmann, S. Huch, P. Karle, et al., Indy autonomous challenge-autonomous race cars at the han- dlinglimits,in:12thInternationalMunichChassisSymposium2021: chassis. tech plus, Springer, 2022, pp. 163–182

    2022

  6. [6]

    A2RL - Abu Dhabi Autonomous Racing League, Autonomous car race,https://a2rl.io/autonomous-car-race, accessed: 2024-05- 20 (2024)

    2024

  7. [7]

    Demeter, M

    Z. Demeter, M. Hell, G. Hajgató, Lessons learned from an au- tonomous race car competition, Engineering Proceedings 79 (1) (2024) 25

    2024

  8. [8]

    B. Leng, R. Yu, C. Tu, L. Xiong, A. Eichberger, Z. Li, Seamless overtakingmaneuversforautomateddriving:Integratedmotionplan- ningbasedonhybridmodelpredictivecontrol,IEEETransactionson Industrial Electronics (2026)

    2026

  9. [9]

    2288–2295

    J.F.Bongard,V.L.Krieger,B.Lohmann,Dynamicconstrainttighten- ingfornonlinearmpcforautonomousracingviacontractionanalysis, in:2025IEEEIntelligentVehiclesSymposium(IV),IEEE,2025,pp. 2288–2295

    2025

  10. [10]

    Brown, J

    M. Brown, J. C. Gerdes, Coordinating tire forces to avoid obstacles using nonlinear model predictive control, IEEE Transactions on In- telligent Vehicles 5 (1) (2019) 21–31

    2019

  11. [11]

    Mangharam, Autonomous vehicles on the edge: A survey on autonomous vehicle racing, IEEE Open Journal of Intelligent Trans- portation Systems 3 (2022) 458–488

    J.Betz,H.Zheng,A.Liniger,U.Rosolia,P.Karle,M.Behl,V.Krovi, R. Mangharam, Autonomous vehicles on the edge: A survey on autonomous vehicle racing, IEEE Open Journal of Intelligent Trans- portation Systems 3 (2022) 458–488

    2022

  12. [12]

    9403–9409

    Y.Song,H.Lin,E.Kaufmann,P.Dürr,D.Scaramuzza,Autonomous overtaking in gran turismo sport using curriculum reinforcement learning, in: 2021 IEEE international conference on robotics and automation (ICRA), IEEE, 2021, pp. 9403–9409

    2021

  13. [13]

    Kaufmann, L

    E. Kaufmann, L. Bauersfeld, A. Loquercio, M. Müller, V. Koltun, D. Scaramuzza, Champion-level drone racing using deep reinforce- ment learning, Nature 620 (7976) (2023) 982–987

    2023

  14. [14]

    Kabzan, L

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

    2019

  15. [15]

    13250–13256

    H.Xue,E.L.Zhu,J.M.Dolan,F.Borrelli,Learningmodelpredictive control with error dynamics regression for autonomous racing, in: 2024 IEEE International Conference on Robotics and Automation (ICRA), IEEE, 2024, pp. 13250–13256

    2024

  16. [16]

    Weaver, C

    C. Weaver, C. Tang, C. Hao, K. Kawamoto, M. Tomizuka, W. Zhan, Betail: Behavior transformer adversarial imitation learning from hu- man racing gameplay, IEEE Robotics and Automation Letters 9 (8) (2024) 7302–7309

    2024

  17. [17]

    B. D. Evans, H. W. Jordaan, H. A. Engelbrecht, Safe reinforcement learning for high-speed autonomous racing, Cognitive Robotics 3 (2023) 107–126

    2023

  18. [18]

    Z. Li, B. Leng, L. Xiong, A. Eichberger, C. Huang, J. Hu, Safety- enhanced deep reinforcement learning for autonomous driving: Dare to make mistakes to learn better and faster, IEEE Transactions on Intelligent Transportation Systems (2026)

    2026

  19. [19]

    Fuchs, Y

    F. Fuchs, Y. Song, E. Kaufmann, D. Scaramuzza, P. Dürr, Super- human performance in gran turismo sport using deep reinforcement learning, IEEE Robotics and Automation Letters 6 (3) (2021) 4257– 4264

    2021

  20. [20]

    P.R.Wurman,S.Barrett,K.Kawamoto,J.MacGlashan,K.Subrama- nian,T.J.Walsh,R.Capobianco,A.Devlic,F.Eckert,F.Fuchs,etal., Outracing champion gran turismo drivers with deep reinforcement learning, Nature 602 (7896) (2022) 223–228

    2022

  21. [21]

    C. Tang, B. Abbatematteo, J. Hu, R. Chandra, R. Martín-Martín, P. Stone, Deep reinforcement learning for robotics: A survey of real-world successes, Annual Review of Control, Robotics, and Au- tonomous Systems 8 (1) (2025) 153–188

    2025

  22. [22]

    Ladosz, L

    P. Ladosz, L. Weng, M. Kim, H. Oh, Exploration in deep reinforce- ment learning: A survey, Information Fusion 85 (2022) 1–22

    2022

  23. [23]

    Zhang, Q

    Z. Zhang, Q. Liu, Y. Li, K. Lin, L. Li, Safe reinforcement learning in autonomous driving with epistemic uncertainty estimation, IEEE Transactions on Intelligent Transportation Systems 25 (10) (2024) 13653–13666

    2024

  24. [24]

    Heilmeier, A

    A. Heilmeier, A. Wischnewski, L. Hermansdorfer, J. Betz, M. Lienkamp, B. Lohmann, Minimum curvature trajectory planning and control for an autonomous race car, Vehicle System Dynamics (2020)

    2020

  25. [25]

    Z. Zhu, X. Tang, Y. Qin, Y. Huang, E. Hashemi, A survey of lateral stability criterion and control application for autonomous vehicles, IEEE Transactions on Intelligent Transportation Systems 24 (10) (2023) 10382–10399

    2023

  26. [26]

    J.-H. Ryu, D. Ogay, S. Bulavintsev, H. Kim, J.-S. Park, Development and experiences of an autonomous vehicle for high-speed navigation and obstacle avoidance, in: Frontiers of Intelligent Autonomous Sys- tems, Springer, 2013, pp. 105–116

    2013

  27. [27]

    A. D. Ames, S. Coogan, M. Egerstedt, G. Notomista, K. Sreenath, P. Tabuada, Control barrier functions: Theory and applications, in: 201918thEuropeancontrolconference(ECC),Ieee,2019,pp.3420– 3431

    2019

  28. [28]

    Accessed: 2015-07-13 (2015)

    FIA Formula E, Berlin circuit, fiaformulae.com, archived from the original on July 13, 2015. Accessed: 2015-07-13 (2015). URLhttp://www.fiaformulae.com/en/calendar/berlin-eprix/ berlin-circuit.aspx

    2015

  29. [29]

    B. D. Evans, H. A. Engelbrecht, H. W. Jordaan, High-speed au- tonomous racing using trajectory-aided deep reinforcement learning, IEEE Robotics and Automation Letters 8 (9) (2023) 5353–5359. Bo Leng et al.:Preprint submitted to ElsevierPage 13 of 14 Expert Knowledge-driven Reinforcement Learning for Autonomous Racing via Trajectory Guidance and Dynamics C...

    2023

  30. [30]

    Guizhe Jinreceived the bachelor’s degree in mechanical engineering from Harbin Institute of Technology, Weihai, China, in 2019

    He was a recipient of the National Science Fund for Distinguished Young Scholars. Guizhe Jinreceived the bachelor’s degree in mechanical engineering from Harbin Institute of Technology, Weihai, China, in 2019. He is cur- rently pursuing the master’s degree with the College of Automotive and Energy Engineering, Tongji University, Shanghai, China. His resea...

    2019

  31. [31]

    He joined NTU, as a Nanyang Assistant Professor and has founded the Automated Driving andHuman-MachineSystem(AutoMan)Research Laboratory in June 2018. His research focuses on advanced vehicle control and intelligence, where he has contributed four books, over 200 papers, and received 12 granted patents and five patent applications. Bo Leng et al.:Preprint...

    2018