Pith. sign in

REVIEW 6 major objections 8 minor 2 cited by

LeAD: The LLM Enhanced Planning System Converged with End-to-end Autonomous Driving

T0 review · 6 major / 8 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read A low-frequency LLM planner, activated only when the vehicle is blocked, lifts an end-to-end driving system to a CARLA driving score of 71.96.

desk verdict The architecture is a reasonable dual-rate LLM-fallback design, but the headline benchmark claim is not established because LeAD's score comes from a fixed-weather deterministic protocol while the baselines are official randomized leaderboard scores. read the letter →

arxiv 2507.05754 v1 pith:L7SECR6S submitted 2025-07-08 cs.RO cs.AI

Yuhang Zhang , Jiaqi Liu , Chengkai Xu , Peng Hang , Jian Sun This is my paper
classification cs.ROcs.AI
keywords autonomousdrivinglargelanguagemodelsend-to-endplanningdual-ratearchitectureedgecaseschain-of-thoughtreasoningCARLALeaderboardsafetycontroller
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

This paper claims that a large language model can serve as a low-frequency 'backup driver' for an end-to-end autonomous driving system, handling exactly the edge cases that immobilize the primary planner. The proposed system, LeAD, runs a fast neural-network planner normally and activates an LLM-based reasoning module only when the vehicle is stuck beyond a waiting threshold. In the CARLA simulator's Leaderboard V1 benchmark, LeAD reports a driving score of 71.96 and route completion of 93.43%, ahead of the compared end-to-end baselines. The result matters because it suggests a practical division of labor: real-time reactive control stays in the neural network, while rare semantic reasoning is offloaded to a slow but flexible LLM.

What carries the argument

The load-bearing mechanism is the dual-rate asynchronous coupling plus the safety-gated LLM decision pipeline. A high-frequency transformer-based E2E network (ResNet backbone, multimodal camera and LiDAR inputs) generates waypoints and perception outputs, while a rule-based safety controller normally enforces collision avoidance. When the E2E planner fails to resolve a blockage and a waiting threshold is exceeded, a low-frequency LLM module activates: a scenario encoder converts perception and HD-map data into structured natural language with intentionally attenuated numerical precision, the LLM runs chain-of-thought reasoning over predefined decision options, and a second safety controller - which relaxes the E2E safety redundancies while preserving fundamental collision-avoidance constraints - converts the chosen high-level decision into executable control commands. The key design choice is that the LLM never drives in real time; it only suggests discrete maneuvers that the safety layer filters.

What would settle it

Run LeAD in CARLA on a route where an oncoming vehicle approaches at an angle on the boundary of the 8-direction heading discretization while the ego vehicle is blocked; if the LLM's lane-change decision sends the vehicle into a collision that the E2E-only baseline avoids, the encoded perception is losing safety-critical information.

Watch

Extended reading notes

Core claim

On the paper's own terms, the central discovery is that coupling a high-frequency end-to-end planner with a low-frequency LLM planner through an asynchronous safety-triggered mechanism yields both real-time responsiveness and improved handling of unconventional traffic scenarios. When the ego vehicle remains blocked after a predefined wait, the LLM receives a deliberately coarse textual description of the scene - road topology from HD maps, traffic participants' relative positions, discretized headings, and types - and performs chain-of-thought reasoning to pick a high-level action such as 'lane change right.' That action is decoded into trajectory parameters and passed through a safety controller that relaxes the fast system's redundancy constraints at low speed. In CARLA Leaderboard V1, this design attains a 71.96 driving score with 93.43% route completion, and an ablation removing the LLM module shows lower route completion and longer traversal times on three routes.

Load-bearing premise

The system assumes that converting raw perception data into coarse, qualitative text (dropping precise distances, speeds, and headings) still leaves enough information for the LLM to pick a safe high-level action, and that the rule-based safety controller will catch any error introduced by that quantization.

Editorial extensions

If this is right

  • If LeAD's result holds, LLM-based reasoning can be layered on existing end-to-end planners without sacrificing control frequency, because the LLM is only invoked during immobilization events.
  • The 8-direction heading discretization and qualitative position encoding suggest that high-level driving decisions may tolerate coarse perception representations, as long as a safety controller validates execution.
  • The ablation's reported 18.4% route-completion gain and 23.7% traversal-time reduction on select routes imply that a substantial fraction of E2E failures stem from semantic misinterpretation rather than raw perception error.
  • The framework predicts that performance gains from the LLM module grow with route complexity, a trend the authors observe across Town01, Town02, and Town04.

Reading between the lines

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

  • A testable extension: if the safety controller's downgraded constraints are the real source of the gain, then replacing the LLM with a random or heuristic decision selector over the same discrete action set should reproduce much of the driving-score improvement; the paper's ablation removes the whole LLM module rather than ablating the reasoning itself.
  • The deliberately lossy encoding raises an information-theoretic question: how much numerical precision can be discarded before lane-change or gap-acceptance decisions become unsafe? The paper does not probe this boundary.
  • Because the LLM occasionally chooses 'Hold,' the system trades longer blockages for fewer collisions and deadlocks; real-world deployment would need a timeout that overrides the LLM's conservative choices.
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

6 major / 8 minor

Summary. LeAD couples a high-frequency end-to-end planner with a low-frequency LLM planner that activates after the vehicle is blocked, using natural-language encoding of perception, chain-of-thought reasoning, and a rule-based safety controller. The paper reports CARLA Leaderboard V1 results (driving score 71.96, route completion 93.43%) and an ablation study claiming that the LLM module improves route completion and traversal time. A single hand-picked two-lane case is presented as qualitative validation of the approach.

Significance. If the reported results held, the architecture would be a useful contribution: it directly addresses LLM latency by introducing a dual-rate activation mechanism and provides a concrete way to graft semantic reasoning onto an end-to-end driving pipeline. The use of an external benchmark and the inclusion of an ablation study are appropriate, and the activation logic is clearly described. However, the central empirical claim is not currently established because the benchmark protocol differs from the baselines, the aggregate ablation numbers are not derivable from Table II, and Table I is internally inconsistent with the stated metric formula. The paper also provides no error bars or code, and the single case study encodes the answer in the prompt.

major comments (6)
  1. [Section IV, Table I] The comparison against TCP, Transfuser, Transfuser(r), and TF++ is invalid as reported. Section IV states that LeAD was evaluated under a controlled protocol with fixed weather and deterministic geofenced event triggers, whereas the baselines are described as 'selected from the CARLA Leaderboard V1' and therefore used the official randomized protocol. Since the paper does not state that the baselines were re-evaluated under the LeAD protocol, the 71.96 driving score and the '1-st position' claim are not comparable to the published baseline scores. The authors should re-run all baselines under the identical controlled protocol or evaluate LeAD under the official Leaderboard protocol.
  2. [Section V.B, Table I] The driving-score formula and Table I are inconsistent. Section V.B defines Driving Score as the product of Route Completion and the Penalty Multiplier, but the LeAD row reports 93.43 × 0.76 = 71.01, not 71.96, and similar discrepancies appear for the baselines (e.g., 89.65 × 0.63 = 56.48, not 55.04). The authors should either correct the reported values or clarify whether 'Infraction Penalty' is a different quantity for which the score formula is not as stated.
  3. [Section V.D, Table II] The claimed 18.4% average route-completion improvement and 23.7% traversal-duration reduction are not supported by the reported data. Table II shows percentage-point route-completion gains of 13.25 (Town01), 27.76 (Town02), and 3.23 (Town04), whose average is 14.7 percentage points; the corresponding relative gains are 25.8%, 54.1%, and 3.8%, averaging 27.9%. Neither calculation yields 18.4%, and no traversal-duration numbers are provided anywhere. The text also claims gains become 'more pronounced as the route complexity increases,' but Town04 shows the smallest gain, contradicting that trend. Please state exactly how the averages were computed and report per-route traversal times.
  4. [Section V.E, Fig. 6] The case study does not demonstrate emergent reasoning because the prompt already encodes the answer. Fig. 6 states that the ego is blocked, that there is a drivable lane on the right and no drivable lane on the left, and that the original driving point is in front, which leaves 'change to the right lane' as the only viable option. The CoT output is therefore largely a restatement of the prompt. Please provide quantitative edge-case statistics and examples in which the LLM must reason from partial or ambiguous information.
  5. [Section IV, Section V] The empirical evaluation lacks repeated runs and error bars. CARLA Leaderboard scenarios involve stochastic traffic, and the LLM API output is sampled (the temperature is not reported), so single-run scores for LeAD and the baselines do not establish the small margins in Table I. Please report the number of runs, seeds, and standard deviations or confidence intervals for all headline metrics.
  6. [Section III-D, Section III-B] The design intentionally attenuates numerical precision in perception data, converting quantitative measurements into discrete attribute information, and relies on the rule-based safety controller to catch LLM errors. No analysis or experiment addresses the failure mode in which quantization removes a safety-critical distance or speed margin. Please add a safety analysis or targeted experiments (e.g., perturbing the quantized attributes) to support the claim that the safety controller reliably constrains any LLM output.
minor comments (8)
  1. [Section I] The word 'transformer artitechture' should be 'transformer architecture.'
  2. [Figures 2-4] Figures 2 through 4 contain untranslated Chinese labels ('技术路线' and '技术路线与工作进展'), which should be removed or translated for an English-language journal.
  3. [Section III-E, Fig. 4] The paper states that GPT-4o-mini is used in 'zero-shot mode' but also describes exemplar demonstrations and standardized CoT templates; this is at least few-shot prompting and the description should be made consistent.
  4. [General] No code or data availability statement is provided, which limits reproducibility of the CARLA experiments.
  5. [Section V.E] Fig. 5(f) is labeled 'LLM Choice Analysis' but the text does not explain what is shown in this visualization; please describe it.
  6. [Conclusion] The conclusion says route completion improved 'by up to 18.4%,' while Section V.D calls the same number an average; the wording should be made consistent.
  7. [Table II] Table II has the same metric inconsistency: 85.47 × 0.84 = 71.79, not the reported 71.27 for AD&Town04.
  8. [References] References [12] and [21] are the same Transfuser paper under different titles; please consolidate them.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: LeAD's central performance claims are grounded in external CARLA simulation and ablation, not in self-referential reasoning.

full rationale

LeAD's central claim—that coupling a low-frequency LLM planner with a high-frequency E2E planner improves edge-case handling—is supported by closed-loop CARLA driving scores and by an ablation that removes the LLM module while keeping everything else fixed. The driving score, route completion, and infraction metrics are computed by the CARLA Leaderboard metric system, an external simulator/benchmark, not by a quantity defined in terms of the model's own outputs. No parameter in the paper is fitted to the reported leaderboard numbers and then renamed a prediction; the E2E weights come from imitation learning and the LLM is used in zero-shot mode (GPT-4o-mini, Section IV). The self-citations ([4,5,13,14,19,24]) appear in related work and in game-theoretic/trajectory-planning context; none is invoked as the unique justification for LeAD's architecture or as a uniqueness theorem forbidding alternatives, so they are not load-bearing. The closest thing to a circular flavor is the case-study prompt (Fig. 6), which states 'Your original driving point is in front of you' and 'There is a drivable lane on your right side,' making the LLM's 'right lane change' output heavily primed by the input. This is a confound for the qualitative demonstration, but it is not a derivation of the quantitative results: the 71.96 driving score and the Table II ablation gains are measured in the external benchmark, not defined by the LLM's choices. The invalid-protocol concern raised for the baseline comparison is a correctness/experimental-design issue, not a circularity issue, and is therefore not scored here.

Assumptions & free parameters 3 free parameters · 4 assumptions · 0 invented entities

The central claim does not rest on fitted constants; the empirical system has hand-chosen design parameters (activation threshold, action-to-trajectory mappings, look-ahead distance) whose values are not reported. The assumptions are that the E2E perception is accurate enough for textual summarization, that CARLA transfers to open-road driving, and that the safety validator is reliable.

free parameters (3)
  • LLM activation waiting threshold
    Section III-B: "Upon exceeding a predefined waiting threshold, the LLM redundant system activates." The value is not reported and is hand-chosen; it determines how often the LLM intervenes and affects all results.
  • Action-to-trajectory parameters
    Section III-E-3: each permitted decision maps to preconfigured steering curvature, target speed, and execution duration; these are hand-set and not given.
  • Look-ahead distance X for steering
    Section III-C-3: steering uses a preview point at distance X; X is not specified, presumably tuned or borrowed from prior work [30].
assumptions (4)
  • domain assumption E2E perception outputs and HD-map lane localization are accurate enough to ground LLM decisions.
    The LLM module relies on perception outputs from the E2E network and lane-level localization from HD maps (Section III-D); no error analysis is provided, and perception failures would propagate into the text description.
  • domain assumption CARLA Leaderboard V1 is a meaningful proxy for the open-road deployment the paper motivates.
    The introduction frames open-road deployment as the goal; evaluation is entirely in CARLA (Section IV), so sim-to-real transfer is assumed rather than tested.
  • domain assumption Stopping the car and waiting is the main detectable failure mode that the LLM needs to resolve.
    The LLM is triggered only after a waiting threshold (Section III-B); other edge cases, such as perception errors at speed, are not addressed by the coupling mechanism.
  • ad hoc to paper The rule-based safety validator reliably constrains any LLM output.
    The safety controller is described as preserving fundamental collision avoidance constraints (Section III-B), but its logic and verification are not formalized, and no safety analysis is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of LeAD: The LLM Enhanced Planning System Converged with End-to-end Autonomous Driving." pith.science (2026). https://pith.science/paper/L7SECR6S

@misc{pith2026250705754,
  author       = {Pith},
  title        = {Pith review of: LeAD: The LLM Enhanced Planning System Converged with End-to-end Autonomous Driving},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/L7SECR6S}},
  note         = {Machine review of arXiv:2507.05754}
}
read the original abstract

A principal barrier to large-scale deployment of urban autonomous driving systems lies in the prevalence of complex scenarios and edge cases. Existing systems fail to effectively interpret semantic information within traffic contexts and discern intentions of other participants, consequently generating decisions misaligned with skilled drivers' reasoning patterns. We present LeAD, a dual-rate autonomous driving architecture integrating imitation learning-based end-to-end (E2E) frameworks with large language model (LLM) augmentation. The high-frequency E2E subsystem maintains real-time perception-planning-control cycles, while the low-frequency LLM module enhances scenario comprehension through multi-modal perception fusion with HD maps and derives optimal decisions via chain-of-thought (CoT) reasoning when baseline planners encounter capability limitations. Our experimental evaluation in the CARLA Simulator demonstrates LeAD's superior handling of unconventional scenarios, achieving 71 points on Leaderboard V1 benchmark, with a route completion of 93%.

Figures

Figures reproduced from arXiv: 2507.05754 by the authors.

Figure 1
Figure 1. LeAD’s Basic Architecture. The proposed architec [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The dual-system decision architecture comprises two [PITH_FULL_IMAGE:figures/full_fig_p002_2.png] view at source ↗
Figure 2
Figure 2. Our dual-rate decision architecture comprises a high-frequency primary system performing real-time object [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figures from the paper (4 more)
Figure 3
Figure 3. Figure 3: End-to-end system neural network architecture. Build [PITH_FULL_IMAGE:figures/full_fig_p004_3.png]
Figure 4
Figure 4. Figure 4: LLM Prompt Pipline in LeAD. The framework implements hierarchical structuring of scene components through tiered [PITH_FULL_IMAGE:figures/full_fig_p005_4.png]
Figure 5
Figure 5. Figure 5: A Sample Case. (a)-(d) present sequential video [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]
Figure 6
Figure 6. Figure 6: Description of the example case VI. CONCLUSIONS Enhancing scene semantic comprehension and reasoning capabilities remains a key challenge in autonomous driving. In this work, we propose a dual-rate-system autonomous driving architecture LeAD, that effectively combines …

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Bench2Drive-Robust: Benchmarking Closed-Loop Autonomous Driving under Deployment Perturbations

    cs.RO 2026-05 unverdicted novelty 7.0 of 10

    Bench2Drive-Robust is a new closed-loop benchmark that evaluates end-to-end autonomous driving models under deployment perturbations from camera failures, ego-state errors, and compute delays, showing substantial perf...

  2. Steins;Gate Drive: Semantic Safety Arbitration over Structured Futures for Latency-Decoupled LLM Planning

    cs.RO 2026-05 unverdicted novelty 5.0 of 10

    SteinsGateDrive decouples LLM inference latency from vehicle control by pre-selecting alpha, beta, and gamma worldline futures that a runtime validates against safety contracts until abort conditions trigger.

Reference graph

Works this paper leans on

44 extracted references · 25 canonical work pages · cited by 2 Pith papers

  1. [1]

    Traffic jam caused by self-driving cars in san francisco sparks outcry and safety concerns,

    M. Garfinkle, “Traffic jam caused by self-driving cars in san francisco sparks outcry and safety concerns,” 2023. [Online]. Available: https://www.entrepreneur.com/business-news/ self-driving-robotaxi-cars-trigger-traffic-jam-spark-outcry/457485

    work page 2023

  2. [2]

    Chauffeurnet: Learning to drive by imitating the best and synthesizing the worst,

    M. Bansal, A. Krizhevsky, and A. Ogale, “Chauffeurnet: Learning to drive by imitating the best and synthesizing the worst,” arXiv preprint arXiv:1812.03079, 2018

    arXiv 2018

  3. [3]

    Densetnt: End-to-end trajectory pre- diction from dense goal sets,

    J. Gu, C. Sun, and H. Zhao, “Densetnt: End-to-end trajectory pre- diction from dense goal sets,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 15 303– 15 312

    2021

  4. [4]

    Game theory-based interactive control for human–machine cooperative driving,

    Y . Zhou, C. Huang, and P. Hang, “Game theory-based interactive control for human–machine cooperative driving,” Applied Sciences , vol. 14, no. 6, p. 2441, 2024

    work page 2024

  5. [5]

    Cooperative driving of connected autonomous vehicles in heterogeneous mixed traffic: A game theoretic approach,

    S. Fang, P. Hang, C. Wei, Y . Xing, and J. Sun, “Cooperative driving of connected autonomous vehicles in heterogeneous mixed traffic: A game theoretic approach,” IEEE Transactions on Intelligent Vehicles , 2024

    work page 2024

  6. [6]

    Driver-automation cooperative approach for shared steering control under multiple system constraints: Design and experiments,

    A.-T. Nguyen, C. Sentouh, and J.-C. Popieul, “Driver-automation cooperative approach for shared steering control under multiple system constraints: Design and experiments,” IEEE Transactions on Industrial Electronics, vol. 64, no. 5, pp. 3819–3830, 2016

    work page 2016

  7. [7]

    V oxposer: Composable 3d value maps for robotic manipulation with language models,

    W. Huang, C. Wang, R. Zhang, Y . Li, J. Wu, and L. Fei-Fei, “V oxposer: Composable 3d value maps for robotic manipulation with language models,” arXiv preprint arXiv:2307.05973 , 2023

    arXiv 2023

  8. [8]

    Driving with llms: Fusing object- level vector modality for explainable autonomous driving,

    L. Chen, O. Sinavski, J. H ¨unermann, A. Karnsund, A. J. Willmott, D. Birch, D. Maund, and J. Shotton, “Driving with llms: Fusing object- level vector modality for explainable autonomous driving,” in 2024 IEEE International Conference on Robotics and Automation (ICRA) . IEEE, 2024, pp. 14 093–14 100

    work page 2024

Show all 44 references
  1. [9]

    Driving everywhere with large language model policy adaptation,

    B. Li, Y . Wang, J. Mao, B. Ivanovic, S. Veer, K. Leung, and M. Pavone, “Driving everywhere with large language model policy adaptation,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 14 948–14 957

    2024

  2. [10]

    Chain-of-thought prompting elicits reasoning in large language models,

    J. Wei, X. Wang, D. Schuurmans, M. Bosma, F. Xia, E. Chi, Q. V . Le, D. Zhou et al. , “Chain-of-thought prompting elicits reasoning in large language models,” Advances in neural information processing systems, vol. 35, pp. 24 824–24 837, 2022

    2022

  3. [11]

    Drivelm: Driving with graph visual question answering,

    C. Sima, K. Renz, K. Chitta, L. Chen, H. Zhang, C. Xie, P. Luo, A. Geiger, and H. Li, “Drivelm: Driving with graph visual question answering,” arXiv preprint arXiv:2312.14150 , 2023

    2023 arXiv

  4. [12]

    Transfuser: Imitation with transformer-based sensor fusion for au- tonomous driving,

    K. Chitta, A. Prakash, B. Jaeger, Z. Yu, K. Renz, and A. Geiger, “Transfuser: Imitation with transformer-based sensor fusion for au- tonomous driving,”IEEE transactions on pattern analysis and machine intelligence, vol. 45, no. 11, pp. 12 878–12 895, 2022

    2022

  5. [13]

    Cooperative decision- making for cavs at unsignalized intersections: A marl approach with attention and hierarchical game priors,

    J. Liu, P. Hang, X. Na, C. Huang, and J. Sun, “Cooperative decision- making for cavs at unsignalized intersections: A marl approach with attention and hierarchical game priors,” IEEE Transactions on Intelli- gent Transportation Systems, 2024

    2024

  6. [14]

    Enhancing social decision- making of autonomous vehicles: A mixed-strategy game approach with interaction orientation identification,

    J. Liu, X. Qi, P. Hang, and J. Sun, “Enhancing social decision- making of autonomous vehicles: A mixed-strategy game approach with interaction orientation identification,” IEEE Transactions on Vehicular Technology, 2024

    2024

  7. [15]

    Gri: General reinforced imitation and its application to vision-based au- tonomous driving,

    R. Chekroun, M. Toromanoff, S. Hornauer, and F. Moutarde, “Gri: General reinforced imitation and its application to vision-based au- tonomous driving,” Robotics, vol. 12, no. 5, p. 127, 2023

    2023

  8. [16]

    Learning to drive from a world on rails,

    D. Chen, V . Koltun, and P. Kr ¨ahenb¨uhl, “Learning to drive from a world on rails,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2021, pp. 15 590–15 599

    2021

  9. [17]

    End-to- end urban driving by imitating a reinforcement learning coach,

    Z. Zhang, A. Liniger, D. Dai, F. Yu, and L. Van Gool, “End-to- end urban driving by imitating a reinforcement learning coach,” in Proceedings of the IEEE/CVF international conference on computer vision, 2021, pp. 15 222–15 232

    2021

  10. [18]

    End-to-end model-free reinforcement learning for urban driving using implicit affordances,

    M. Toromanoff, E. Wirbel, and F. Moutarde, “End-to-end model-free reinforcement learning for urban driving using implicit affordances,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2020, pp. 7153–7162

    2020

  11. [19]

    Tell- drive: Enhancing autonomous driving with teacher llm-guided deep reinforcement learning,

    C. Xu, J. Liu, S. Fang, Y . Cui, D. Chen, P. Hang, and J. Sun, “Tell- drive: Enhancing autonomous driving with teacher llm-guided deep reinforcement learning,” arXiv preprint arXiv:2502.01387 , 2025

    2025 arXiv

  12. [20]

    Planning-oriented autonomous driving,

    Y . Hu, J. Yang, L. Chen, K. Li, C. Sima, X. Zhu, S. Chai, S. Du, T. Lin, W. Wang et al. , “Planning-oriented autonomous driving,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 17 853–17 862

    2023

  13. [21]

    Transfuser: Imitation with transformer-based sensor fusion for au- tonomous driving,

    K. Chitta, A. Prakash, B. Jaeger, Z. Yu, K. Renz, and A. Geiger, “Transfuser: Imitation with transformer-based sensor fusion for au- tonomous driving,” Pattern Analysis and Machine Intelligence (PAMI), 2023

    2023

  14. [22]

    Reasonnet: End-to-end driving with temporal and global reasoning,

    H. Shao, L. Wang, R. Chen, S. L. Waslander, H. Li, and Y . Liu, “Reasonnet: End-to-end driving with temporal and global reasoning,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 13 723–13 733

    2023

  15. [23]

    Optimal time allocation for quadrotor trajectory generation,

    F. Gao, W. Wu, J. Pan, B. Zhou, and S. Shen, “Optimal time allocation for quadrotor trajectory generation,” in 2018 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS) . IEEE, 2018, pp. 4715–4722

    2018

  16. [24]

    Optimization-based trajectory planning for autonomous parking with irregularly placed obstacles: A lightweight iterative framework,

    B. Li, T. Acarman, Y . Zhang, Y . Ouyang, C. Yaman, Q. Kong, X. Zhong, and X. Peng, “Optimization-based trajectory planning for autonomous parking with irregularly placed obstacles: A lightweight iterative framework,” IEEE Transactions on Intelligent Transportation Systems, vo...

    2021

  17. [25]

    Drivevlm: The convergence of au- tonomous driving and large vision-language models,

    X. Tian, J. Gu, B. Li, Y . Liu, Y . Wang, Z. Zhao, K. Zhan, P. Jia, X. Lang, and H. Zhao, “Drivevlm: The convergence of au- tonomous driving and large vision-language models,” arXiv preprint arXiv:2402.12289, 2024

    2024 arXiv

  18. [26]

    Gpt-driver: Learning to drive with gpt,

    J. Mao, Y . Qian, J. Ye, H. Zhao, and Y . Wang, “Gpt-driver: Learning to drive with gpt,” arXiv preprint arXiv:2310.01415 , 2023

    2023 arXiv

  19. [27]

    Drivegpt4: Interpretable end-to-end autonomous driving via large language model,

    Z. Xu, Y . Zhang, E. Xie, Z. Zhao, Y . Guo, K.-Y . K. Wong, Z. Li, and H. Zhao, “Drivegpt4: Interpretable end-to-end autonomous driving via large language model,” IEEE Robotics and Automation Letters , 2024

    2024

  20. [28]

    Languagempc: Large language models as decision makers for autonomous driving,

    H. Sha, Y . Mu, Y . Jiang, L. Chen, C. Xu, P. Luo, S. E. Li, M. Tomizuka, W. Zhan, and M. Ding, “Languagempc: Large language models as decision makers for autonomous driving,” 2025. [Online]. Available: https://arxiv.org/abs/2310.03026

    2025 arXiv

  21. [29]

    Lmdrive: Closed-loop end-to-end driving with large language models,

    H. Shao, Y . Hu, L. Wang, G. Song, S. L. Waslander, Y . Liu, and H. Li, “Lmdrive: Closed-loop end-to-end driving with large language models,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition , 2024, pp. 15 120–15 130

    2024

  22. [30]

    Safety-enhanced autonomous driving using interpretable sensor fusion transformer,

    H. Shao, L. Wang, R. Chen, H. Li, and Y . Liu, “Safety-enhanced autonomous driving using interpretable sensor fusion transformer,” in Conference on Robot Learning . PMLR, 2023, pp. 726–737

    2023

  23. [31]

    Multi-modal fusion transformer for end-to-end autonomous driving,

    A. Prakash, K. Chitta, and A. Geiger, “Multi-modal fusion transformer for end-to-end autonomous driving,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition , 2021, pp. 7077–7087

    2021

  24. [32]

    Fusionad: Multi-modality fusion for pre- diction and planning tasks of autonomous driving,

    T. Ye, W. Jing, C. Hu, S. Huang, L. Gao, F. Li, J. Wang, K. Guo, W. Xiao, W. Mao et al. , “Fusionad: Multi-modality fusion for pre- diction and planning tasks of autonomous driving,” arXiv preprint arXiv:2308.01006, 2023

    2023 arXiv

  25. [33]

    Bevfusion: Multi-task multi-sensor fusion with unified bird’s- eye view representation,

    Z. Liu, H. Tang, A. Amini, X. Yang, H. Mao, D. L. Rus, and S. Han, “Bevfusion: Multi-task multi-sensor fusion with unified bird’s- eye view representation,” in 2023 IEEE international conference on robotics and automation (ICRA) . IEEE, 2023, pp. 2774–2781

    2023

  26. [34]

    Pointpillars: Fast encoders for object detection from point clouds,

    A. H. Lang, S. V ora, H. Caesar, L. Zhou, J. Yang, and O. Beijbom, “Pointpillars: Fast encoders for object detection from point clouds,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2019, pp. 12 697–12 705

    2019

  27. [35]

    Deep residual learning for image recognition,

    K. He, X. Zhang, S. Ren, and J. Sun, “Deep residual learning for image recognition,” in Proceedings of the IEEE conference on computer vision and pattern recognition , 2016, pp. 770–778

    2016

  28. [36]

    Attention is all you need,

    A. Vaswani, N. Shazeer, N. Parmar, J. Uszkoreit, L. Jones, A. N. Gomez, Ł. Kaiser, and I. Polosukhin, “Attention is all you need,” Advances in neural information processing systems , vol. 30, 2017

    2017

  29. [37]

    Path tracking control for autonomous vehicles based on an improved mpc,

    H. Wang, B. Liu, X. Ping, and Q. An, “Path tracking control for autonomous vehicles based on an improved mpc,” IEEE access, vol. 7, pp. 161 064–161 073, 2019

    2019

  30. [38]

    CARLA: An open urban driving simulator,

    A. Dosovitskiy, G. Ros, F. Codevilla, A. Lopez, and V . Koltun, “CARLA: An open urban driving simulator,” in Proceedings of the 1st Annual Conference on Robot Learning , 2017, pp. 1–16

    2017

  31. [39]

    Carla autonomous driving leaderboard,

    C. team, “Carla autonomous driving leaderboard,” 2020, accessed: 2021-02-11. [Online]. Available: https://leaderboard.carla.org/

    2020

  32. [40]

    Gpt-4o mini: advancing cost-efficient intelligence,

    OpenAI, “Gpt-4o mini: advancing cost-efficient intelligence,” 2024, accessed: 2025-04-23. [Online]. Available: https://openai.com/index/ gpt-4o-mini-advancing-cost-efficient-intelligence/

    2024

  33. [41]

    Eureka: Human- level reward design via coding large language models,

    Y . J. Ma, W. Liang, G. Wang, D.-A. Huang, O. Bastani, D. Ja- yaraman, Y . Zhu, L. Fan, and A. Anandkumar, “Eureka: Human- level reward design via coding large language models,” arXiv preprint arXiv:2310.12931, 2023

    2023 arXiv

  34. [42]

    Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline,

    P. Wu, X. Jia, L. Chen, J. Yan, H. Li, and Y . Qiao, “Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline,” Advances in Neural Information Processing Systems , vol. 35, pp. 6119–6132, 2022

    2022

  35. [43]

    Hidden biases of end-to- end driving models,

    B. Jaeger, K. Chitta, and A. Geiger, “Hidden biases of end-to- end driving models,” in Proceedings of the IEEE/CVF International Conference on Computer Vision , 2023, pp. 8240–8249

    2023

  36. [44]

    Evaluation criteria for the carla autonomous driv- ing leaderboard v1.0,

    CARLA Team, “Evaluation criteria for the carla autonomous driv- ing leaderboard v1.0,” https://leaderboard.carla.org/evaluation v1 0/, 2020, accessed: 2025-04-23

    2020

Pith tools

Reviewed August 6, 2026 · model on record in the stance chip above.