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

Bo Leng; Chen Lv; Guizhe Jin; Lu Xiong; Ran Yu; Weiqi Zhang; Zhuoren Li

REVIEW 4 major objections 6 minor 31 references

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

T0 review · 4 major / 6 minor · reviewed 2026-08-04 · deepseek-v4-flash

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

Pith's one-line read TraD-RL embeds a racing line and dynamic stability limits into reinforcement learning to make autonomous race cars both faster and safer.

desk verdict A coherent RL-for-racing recipe, but the reported numbers contradict each other and the ablation undercuts the synergy claim. read the letter →

arxiv 2603.05842 v3 pith:J7Y77Y75 submitted 2026-03-06 cs.RO

Bo Leng , Weiqi Zhang , Zhuoren Li , Lu Xiong , Guizhe Jin , Ran Yu , Chen Lv This is my paper · ORCID

classification cs.RO

keywords autonomousracingreinforcementlearningminimumcurvaturelinecontrolbarrierfunctionsdynamicsconstraintscurriculumvehiclestabilityenvelopesideslipangle

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 contends that the two standard weaknesses of reinforcement learning in autonomous racing—inefficient exploration and unsafe actions near the physical limits of handling—can be corrected by injecting expert vehicle knowledge into the learning process. Its TraD-RL framework guides the agent with a prepackaged Minimum Curvature Racing Line and constrains yaw rate and sideslip angle through a control-barrier-function safety envelope enforced by adaptive Lagrangian penalties. A two-stage curriculum first teaches the policy to follow the expert line, then releases it to explore the vehicle's limits. On a simulated Tempelhof circuit, the method reports a 58.83 s lap (versus 61.31 s for the strongest baseline) with lower sideslip-boundary violations, and ablations indicate both injected modules contribute. The central claim is that racing performance and dynamic stability improve together, not as a trade-off.

What carries the argument

The load-bearing mechanism is the combination of a geometric path prior and a physical-state safety prior. The Minimum Curvature Racing Line (MCRL) is the path that minimizes the sum of squared curvatures over lateral offsets from the track centerline, subject to staying within the drivable width; it provides a global reference encoded as an extra binary grid in the observation vector and as three dense reward terms (distance, speed, heading). The safety prior is a control-barrier-function (CBF) envelope over yaw rate ω and sideslip angle β, whose boundaries come from the friction limit and the rear-tire peak slip angle; violation costs are fed through cost critics and adaptive Lagrangian mu

What would settle it

Run TraD-RL with the same training budget on a real instrumented race car (or an independently validated simulator) and compare normalized yaw-rate and sideslip violations with an unconstrained baseline; if the violations do not drop, or if the lap-time gap vanishes when the envelope is recomputed from measured tire/vehicle data, the central claim fails. A simpler computational check: perturb the cornering stiffness and friction values in Eqs. (17)–(19) by a few percent and see whether the safety gains persist; if small envelope errors flip the results, the safety improvement is not robust.

Watch

Extended reading notes

Core claim

The central discovery is that a reinforcement-learning racing policy can be made both faster and safer by structuring the agent's information and constraints around verifiable expert priors. The authors generate an MCRL by minimizing the sum of squared curvatures of a laterally displaced path within the track bounds, then encode it as an additional occupancy-grid channel and use it for trajectory, speed, and heading rewards. Separately, they define a safe operating envelope in the sideslip-angle/yaw-rate phase plane, with yaw-rate limit |ω| ≤ μg/u and sideslip limit derived from the rear-tire peak slip angle, and enforce it as control-barrier-function inequalities via Lagrangian relaxation w

Load-bearing premise

The safety envelope in Eqs. (17)–(19) and the “high-fidelity” simulation are treated as accurate models of the real vehicle's stability limits and racing behavior; if the envelope is not the true physical limit or the simulator is not faithful to the real Tempelhof circuit, the claimed stability improvements are artifacts of the simulator's own state variables.

Editorial extensions

If this is right

A racing RL policy can converge to high performance without exhaustive exploration if the action search is structured by an expert line; the 83% average-speed gap between w/o TG and full TraD-RL illustrates the cost of missing this prior.
Explicit yaw-rate and sideslip constraints do not have to slow the car; TraD-RL reports the fastest lap average speed among compared methods while also reducing boundary violations relative to DDPG and TAL.
The two-stage curriculum implies that safe exploration can be scheduled: first stay near a known-good trajectory, then relax to the physical limit, rather than attempting limit driving from the start.
Ablation results indicate that removing either module leads to a different failure pattern (conservative slow driving without guidance; fast but unstable driving without constraints), so the gains come from the combination, not either component alone.
The reported 100% lap progress after roughly 15k steps suggests the safety constraints stabilize training itself, not just the deployed policy.

Reading between the lines

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

The ReLU dead-zone penalty (costs penalized only above a threshold) is a general recipe for safety-constrained RL and could transfer to other domains with a definable safe envelope, such as highway merging or drone navigation.
The MCRL occupancy-grid augmentation is decoupled from the specific actor-critic architecture, so one could test whether it accelerates learning for other off-policy algorithms or in head-to-head multi-agent racing where optimal lines change.
A natural next step the authors do not take is to enforce the same CBF constraints as a hard safety filter at deployment time; if the envelope is correct, this would convert soft training penalties into a guarantee, at the cost of some performance.
The safety numbers are computed from the same envelope used in training, so an independent validation with measured tire data or a different tire model would be the real test of whether the stability improvement persists outside the simulator.

Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Here's my take. The paper's building blocks are standard—MCRL guidance, CBF-style Lagrangian penalties on yaw rate and sideslip, and a two-stage curriculum—but the specific combination is a legitimate engineering contribution worth talking about. The modeling is clearly written, the baselines include a close prior (TAL), and an ablation is attempted. That is more than many RL racing papers do.

The soft spots are real, and they are in the evidence. The lap-time/average-speed pairs don't resolve to the track length. For the same method, Ours, Table 2 reports 39.79 m/s over 58.83 s; Table 3 reports 39.39 m/s over the same 58.83 s. Those imply distances of about 2341 m and 2318 m. TAL implies 2371 m, and w/o TG implies 1764 m—71% of the track. If some of these are incomplete laps, then comparing lap times across rows is not valid, and the paper never says which are incomplete. That is a load-bearing internal inconsistency.

Then there is the ablation. Removing the dynamics constraints yields a faster lap (55.59 s vs. 58.83 s), yet the text claims the constraints do not simply trade speed for safety. That direct contradiction undermines the central message of synergistic optimization. Maybe the constraints reduce variance or improve safety in ways the numerics don't capture, but the paper doesn't show that. It just states the opposite of what its own table says.

The abstract promises two tracks; the body only reports Tempelhof. The safety metrics are violations of the very CBF costs used as training penalties, so the reported safety gains partly re-measure the objective itself. No error bars, no code, no data. These are fixable in revision, but they are not cosmetic.

If I were the editor, I would send this to peer review, because the method is potentially useful and the baseline comparison is meaningful. But I would require the authors to reconcile the distance/lap-time inconsistencies, report ablation results honestly, correct the abstract, and provide either code or error bars. As written, the quantitative claims are not supported.

Recommendation: serious referee, major revision. It's a maybe for reading group, and I wouldn't cite it in its current form.

Referee Report

4 major / 6 minor

Summary. The paper proposes TraD-RL, a reinforcement-learning framework for autonomous racing that combines three mechanisms: observation augmentation and reward shaping from a precomputed Minimum Curvature Racing Line, Lagrangian-relaxed control-barrier-function-style constraints on yaw rate and sideslip angle, and a two-stage curriculum from trajectory-following to high-speed exploration. The method is evaluated in a simulation of the Berlin Tempelhof circuit against PPO, DDPG, TAL, and two ablated variants. The central claim is that TraD-RL achieves a synergistic improvement in racing performance and driving stability without a speed-safety trade-off.

Significance. If the reported results were reliable, the paper would offer a useful empirical study of combining trajectory priors with dynamics-based safety constraints for RL racing, and the explicit formulation and algorithm pseudocode are valuable. However, the quantitative evidence is internally inconsistent: the same method is assigned two different average speeds for the same lap time, the implied completed distances do not match the stated track length, and the ablation table shows that removing the dynamics constraint gives a faster lap. As a result, the central 'synergistic optimization' claim is not currently supported. The paper also uses safety metrics that partly re-measure the training objective.

major comments (4)

[§4.4.2, Tables 2 and 3] The quantitative evidence is internally inconsistent. For the same method 'Ours' and the same lap time of 58.83 s, Table 2 reports a lap average speed of 39.79 m/s while Table 3 reports 39.39 m/s; these imply completed distances of 2,341 m and 2,318 m, respectively, neither equal to the 2,469 m Tempelhof circuit. Other rows also imply inconsistent distances: TAL implies 2,371 m, w/o DC 2,301 m, and w/o TG only 1,763 m (about 71% of the track). If all rows are completed laps, the numbers cannot all be correct; if some rows are incomplete laps, comparing lap times and normalized safety counts across rows is invalid. The central racing-performance claim is therefore not supported as written.
[§4.5, Table 3; Abstract; §4.4.2] The ablation results contradict the claimed 'synergistic optimization of racing performance and safety' and the statement that the dynamics constraints 'do not simply trade speed for safety.' Removing the dynamics constraints produces a faster lap (55.59 s vs. 58.83 s), while removing trajectory guidance produces fewer time-averaged unsafe counts (9.95/4.10 vs. 16.50/4.61). Thus, on the reported numbers, TraD-RL is not simultaneously faster and safer than both ablations; it is a specific trade-off point. The claim that each component is 'crucial' and that the method achieves synergy needs to be re-quantified or substantially revised.
[§4.3.2, Eqs. (22)-(23), Eq. (28)] The safety metrics are based on the same CBF costs that are penalized during training. The time-averaged lap ω-unsafe and β-unsafe times count violations of h_ω<0 and h_β<0, where the costs c_ω and c_β are exactly the quantities minimized through the Lagrangian penalty in Eq. (28). Reporting these as independent evidence of improved 'driving stability' therefore partially re-measures the optimized objective. The authors should provide independent behavioral indicators—off-track events, collisions, spin-outs, tire-saturation margins, or statistical tests on the full state distributions—before claiming that the constraints improve safety rather than merely reduce a training cost.
[§3.3, Eqs. (24)-(28)] The paper claims that safety constraints are 'strictly satisfying' and that the policy is confined to a safe operating envelope, but the actual optimization uses expected-cost thresholds d_ω and d_β and a ReLU dead-zone in Eq. (28), which permits constraint violations whenever the expected cost is below threshold and imposes no penalty inside the dead-zone. This is a soft constraint, not a strict CBF guarantee. The wording should be softened or, alternatively, the authors should provide pointwise or worst-case violation bounds to substantiate a strict-safety claim.

minor comments (6)

[Abstract and §4.1] The abstract states 'Experiments on two racetracks,' but the experimental section describes only the Berlin Tempelhof circuit. Please either add the second track or correct the abstract.
[§3.1 and §3.2.3] The total reward is never explicitly assembled. The relative weights or combination rule for r_track, r_hs, r_ls, r_lap, r_MCRL, r_TS, and r_H are not specified, which makes the method difficult to reproduce. Please provide the exact reward aggregation and any normalization constants.
[§4.3.2, §4.4.2] The safety metric name 'Time-averaged lap ω-unsafe Times' is ambiguous: it counts events but is said to be normalized by lap time, which would give units of events per second. Clarify the definition. Also, the text in §4.4.2 refers to 'Table 4.1' when it should refer to Table 2.
[§4.1] The environment is described as 'high-fidelity,' but the vehicle model presented in §2.2 is a linear-tire dynamic bicycle model, and the simulation software is not identified. Please specify the simulator and state whether the linear tire model is the actual simulation model; otherwise the 'high-fidelity' claim is not supported.
[Throughout] There are several typographical errors, e.g., 'Trajextory guidance anndDynamics constraints' in Section 1 and 'aSchool of Mechanical' in the affiliation block. A careful proofread is needed.
[Eq. (24) and Algorithm 1] The entropy regularization coefficient υ is used in Eq. (24) and updated in Algorithm 1, but it is not defined in the text. Please specify its role and update rule.

Circularity Check

1 steps flagged · score 4.0 of 10

Safety 'predictions' re-measure the CBF training objective; lap-time comparison remains independent.

self definitional [§4.3.2 (safety metrics) vs §3.3 Eqs. (20)-(23) and §3.4 Eq. (28)]
"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) within a single lap."

The reported safety metric counts violations of the same yaw-rate/sideslip envelope that the paper itself encodes as CBF costs c_ω (Eq. 22) and c_β (Eq. 23), and minimizes through the Lagrangian penalty in Eq. (28). Thus the claimed reduction in ω/β-unsafe times is not an independent behavioral prediction; it is a re-measurement of the training objective. This does not affect the lap-time comparisons, but it makes the 'stability' half of the claimed speed–safety synergy partly tautological: the policy is trained to suppress these exact violations and then evaluated by counting them.

full rationale

The trajectory-guidance component is not circular: the MCRL is generated from a curvature-minimization optimization (Eq. 13), not fitted to the reported lap times, and the reward-shaping terms (Eqs. 14-16) guide exploration without forcing the final lap-time outcome. The dynamics-constraint envelope (Eqs. 17-19) is imported from external vehicle-dynamics literature, and the CBF formulation cites external theory; no load-bearing uniqueness claim is made through self-citation. The principal circularity is localized to the safety evaluation: the 'Time-averaged lap ω/β-unsafe Times' are computed from the same boundaries used to define the CBF training costs, so the safety improvement is partially by construction rather than independently verified. The lap-time and average-speed comparisons against PPO, DDPG and TAL are external to this loop. Note also that Table 2 and Table 3 give inconsistent average speeds for the same method and lap time (39.79 vs 39.39 m/s for Ours, both with 58.83 s), and the ablation shows w/o DC with a faster lap time (55.59 s) and w/o TG with fewer reported violations; these are internal-consistency/correctness concerns, not circularity, and do not raise the circularity score further.

Assumptions & free parameters 5 free parameters · 5 assumptions · 0 invented entities

The central claims rest on five hand-set or unreported constants (reward weights, CBF coefficient, cost thresholds, speed bands, curriculum target speed) plus five domain assumptions about vehicle modeling, stability envelopes, and simulation fidelity. No invented entities are introduced. This is a high parameter burden for an empirical method paper.

free parameters (5)

Reward-shaping weights (r_MCRL, r_TS, r_H; basic rewards)
Section 3.2.3 and Section 3.1 define multiple reward terms, but no combination weights are reported; the balance between speed, tracking, and stability is hand-set and undocumented.
CBF coefficient alpha
Eqs. (20)-(23) define constraint costs with a class-K coefficient alpha; its value is never reported, yet it sets how aggressively violations are penalized.
Safety cost thresholds d_omega and d_beta
Eq. (24) constrains expected yaw-rate and sideslip costs, but the thresholds are not specified; they directly control the speed-vs-safety tradeoff.
Speed-reward band boundaries (u_h1, u_h2, u_l1, u_l2)
Eqs. (8)-(9) use four speed thresholds; values are not reported.
Curriculum stage switch T_switch and stage-2 maximum target speed = T_switch=200000; max target speed not reported
Section 4.1 reports 200,000 training steps; the 'maximum physically allowable speed' used in the high-speed exploration stage is not quantified.

assumptions (5)

standard math Bellman equations and policy-gradient updates are valid foundations for the RL objective (Eqs. 1-3).
Standard RL background used without proof.
domain assumption The dynamic bicycle model with linear tire forces and small-angle approximation (Eqs. 4-6) adequately represents racing behavior at the friction limit.
Introduced in Section 2.2; high-speed racing is exactly where linear tire models are least accurate.
domain assumption The yaw-rate and sideslip boundaries in Eqs. (17)-(19) define the true safe operating envelope.
Taken from survey [25] and Pacejka-based peak slip angle [26]; no validation against the actual vehicle or simulation is provided.
ad hoc to paper Soft Lagrangian penalties with a ReLU dead-zone (Eq. 28) and expected-cost thresholds (Eq. 24) are sufficient to strictly satisfy safety constraints.
CBF theory guarantees forward invariance only if the constraint holds at every state; the implementation only penalizes average violations, so hard safety is assumed, not proven.
domain assumption The minimum-curvature racing line computed in Eq. (13) is a high-value reference for time-optimal racing, and the simulation environment is a faithful Tempelhof model.
MCRL is a geometric heuristic, not an optimal-control solution; simulator fidelity is asserted in Section 4.1 but not demonstrated.

how reviews work

0 comments

Cite this review

Pith. "Pith review of Expert Knowledge-driven Reinforcement Learning for Autonomous Racing via Trajectory Guidance and Dynamics Constraints." pith.science (2026). https://pith.science/paper/J7Y77Y75

@misc{pith2026260305842,
  author       = {Pith},
  title        = {Pith review of: Expert Knowledge-driven Reinforcement Learning for Autonomous Racing via Trajectory Guidance and Dynamics Constraints},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J7Y77Y75}},
  note         = {Machine review of arXiv:2603.05842}
}

read the original abstract

Reinforcement learning has shown significant potential for autonomous racing, but it still faces challenges such as training instability, inefficient exploration, and unsafe action outputs in high-dynamic racing scenarios. This paper proposes a Trajectory guidance and Dynamics constraints Reinforcement Learning (TraD-RL) framework for autonomous racing. The proposed method incorporates expert prior knowledge into policy learning through Minimum Curvature Racing Line (MCRL) guidance, explicit vehicle dynamics constraints, and two-stage curriculum learning. MCRL provides global path and velocity references through observation augmentation and reward shaping, thereby improving exploration efficiency and racing performance. Yaw rate and sideslip angle constraints are introduced to characterize the vehicle dynamic safe operating envelope, and the corresponding stability costs are incorporated into policy optimization through Lagrangian relaxation. Moreover, the two-stage curriculum learning strategy enables a progressive transition from stable trajectory following to high-speed performance exploration. Experiments on two racetracks demonstrate that TraD-RL improves racing performance while maintaining a favorable balance between speed and dynamic stability. Further analyses of ablation, sensitivity, and robustness validate the effectiveness and stability of the proposed framework.

Figures

Figures reproduced from arXiv: 2603.05842 by the authors.

**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 formalized 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 function,  ∶ … view at source ↗

**Figure 3.** Schematic diagram of the ego-centric grid observation 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: environmental geometric features and vehicle kinematic states, as illustr… view at source ↗

**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. 𝑟lap = { 1… view at source ↗

Figures from the paper (7 more)

**Figure 6.** Figure 6: Schematic illustration of mapping the racing line prior into the occupancy grid. grid. Specifically, grid cells covering the racing line waypoints take the value 1, while the remaining areas are 0. The incorporation of this expert prior information significantly enhan…

**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]

**Figure 9.** Figure 9: Schematic diagram of the proposed network architectures. and fed into an MLP to estimate the Q-value of the stateaction pair. The detailed experimental parameters are as follows: • Training Scale: The total number of training steps is set to 𝑇max = 250, 000. The agen…

**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 vehicle’s dynami…

**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 contrast, while the PPO algorithm achieves the lowest violation counts, this is en…

**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…

**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…

Review history (2 revisions) →

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

31 extracted references · 2 linked inside Pith

[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]

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)

arXiv 2024
[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)

arXiv 2025
[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]

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]

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

2024
[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]

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

2026

Show all 31 references

[9]

2288–2295

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

2025
[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]

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]

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]

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]

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]

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]

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]

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]

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]

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]

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]

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]

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]

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]

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]

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]

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]

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]

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]

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 Rei...

2023
[2018]

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 r...

2018
[2022]

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 ...

2019

Pith tools

Reviewed August 4, 2026 · model on record in the stance chip above.

[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)

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)

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