Pith. sign in

REVIEW 4 major objections 6 minor 1 cited by

DriveMRP: Enhancing Vision-Language Models with Synthetic Motion Data for Motion Risk Prediction

T0 review · 4 major / 6 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 synthetic high-risk driving dataset plus visual trajectory prompting lifts VLM accident recognition from 27% to 88% and transfers zero-shot to real footage.

desk verdict A useful synthetic-data recipe and projection trick, but the real-world generalization claim is supported only by an unreleased private dataset. read the letter →

arxiv 2507.02948 v3 pith:6IMBDHYR submitted 2025-06-28 cs.CV cs.AIcs.RO

Zhiyi Hou , Enhui Ma , Fang Li , Zhiyi Lai , Kalok Ho , Zhanqian Wu , Lijun Zhou , Long Chen
show 6 more authors
Chitian Sun Haiyang Sun Bing Wang Guang Chen Hangjun Ye Kaicheng Yu
This is my paper
classification cs.CVcs.AIcs.RO
keywords motionriskpredictionvision-languagemodelssyntheticdataautonomousdrivingBEVsimulationvisualpromptingzero-shotgeneralizationchain-of-thought
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 sets out to show that rare, dangerous driving conditions, which are nearly impossible to collect in sufficient quantity from real roads, can be manufactured in simulation and used to teach vision-language models to judge whether an ego vehicle's planned trajectory will end in a collision, hard braking, abnormal acceleration, or lane violation. To do this it builds DriveMRP-10K, about 10,000 synthetic high-risk clips generated by a bird's-eye-view simulator on top of nuPlan scenes, with risk labels and GPT-4o-written explanations. It then introduces DriveMRP-Agent, a fine-tuning framework that projects trajectory waypoints onto the front-view camera image and prompts the model through a scene-understanding to motion-analysis to risk-prediction chain of thought. The paper reports that this lifts accident-recognition accuracy from 27.13% to 88.03% on its synthetic test set, and from 29.42% to 68.50% when the model is evaluated zero-shot on an in-house real-world high-risk driving dataset. A sympathetic reader would take the central claim to be that synthetic high-risk data plus visual prompting is a generalizable recipe for improving VLM motion risk prediction.

What carries the argument

The load-bearing mechanism is the motion projection-based visual prompting scheme: instead of handing the VLM trajectory waypoints as coordinate text, the future path is drawn onto the ego camera image and the BEV layout using camera intrinsics and extrinsics, turning a numerical sequence into a visually salient line whose spatial relation to other vehicles and lane markings the VLM can reason about directly. Supporting it is the BEV-based high-risk motion simulator, which generates risky polynomial trajectories from scene initial states under rule-based scenario definitions, and the chain-of-thought visual QA format that splits the task into scene understanding, motion analysis, and risk prediction. LoRA fine-tuning on DriveMRP-10K is the training mechanism that instills this behavior.

What would settle it

Re-run DriveMRP-Agent's zero-shot evaluation on a public high-risk driving dataset, or release the in-house set with its annotation guide, and compare per-class accuracy with an independent human-label audit; if accuracy drops to near chance on honestly labeled difficult cases, or if the current 68.50% reflects label leakage or easy negatives, the generalization claim fails.

Watch

Extended reading notes

Core claim

The central claim is that the learned ability to assess future-motion risk in autonomous driving, which real-world data cannot cover adequately because high-risk events are rare, can be acquired from synthetic data and transferred to the real world. DriveMRP-10K is synthesized by a BEV-based motion simulator that models risk from three sources: ego-vehicle dynamics (hard braking, abnormal acceleration), interactions with other vehicles (collisions), and environment constraints (lane violations, off-road departure). DriveMRP-Agent then fine-tunes a Qwen2.5-VL-7B base model with LoRA, feeding it a BEV layout for global context, the front-view scene image, and the future trajectory drawn onto that image by camera projection. The visual projection is argued to close the modality gap between numerical waypoints and images, and the chain-of-thought prompting forces the model to produce scene description, motion analysis, and final risk label. Experiments show consistent gains across multiple VLM baselines and a zero-shot real-world accuracy jump from 29.42% to 68.50%, which the paper presents as evidence of generalization.

Load-bearing premise

The load-bearing premise is that the in-house real-world dataset of 5,000 driving segments, whose collection protocol, annotation instructions, and class distribution are not described or released, is a correct and representative ground-truth benchmark for genuine high-risk events.

Editorial extensions

If this is right

  • If the central claim holds, autonomous driving stacks can cheaply expand long-tail risk coverage by simulating high-risk maneuvers atop existing driving logs, without waiting for real crashes.
  • The projection-based prompting recipe is VLM-agnostic, so future and stronger base VLMs should inherit the same risk-prediction improvement by fine-tuning on DriveMRP-10K.
  • Because the framework outputs explanations (scene description, motion analysis), risk predictions become auditable, which matters for safety-critical decision-making and for iterating planner behavior.
  • Zero-shot transfer at 68.50% accuracy implies that synthetic-data training can substitute for a large volume of hard-to-collect real risk data, at least for the risk categories modeled by the simulator.

Reading between the lines

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

  • The reported gains likely combine two effects: the synthetic data itself and the visual prompting format; a factorial experiment with the same data presented with and without projection would separate their contributions, and the paper's ablation suggests projection matters but does not isolate data-scale effects.
  • Because the real-world test set is in-house and its labeling protocol is undisclosed, the 68.50% figure should be re-verified on a public benchmark or with released data and annotation guidelines before generalizing beyond the four risk categories modeled in simulation.
  • The approach could extend to continuous risk scoring rather than discrete categories: the same simulator can generate trajectories with varying collision margins, letting a VLM output graded risk levels that could feed planner cost functions.
  • A testable extension is to generate perturbations of the same scene with multiple risk levels and check whether the VLM's confidence ordering matches the geometric margin, giving a fine-grained validation that accuracy alone misses.
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

4 major / 6 minor

Summary. The paper introduces DriveMRP-10K, a synthetic high-risk motion dataset generated by a BEV-based simulation on top of nuPlan, and DriveMRP-Agent, a VLM framework that projects planned ego trajectories into front-view images and uses a chain-of-thought visual QA schedule to predict risk categories (collision, abnormal acceleration, abnormal deceleration, illegal lane change, or no risk). The authors report that fine-tuning Qwen2.5-VL-7B and other VLMs on DriveMRP-10K raises risk classification accuracy from 27.13% to 88.03% on the synthetic test split, and that the fine-tuned agent reaches 68.50% accuracy on an in-house real-world high-risk dataset, compared with 29.42% for the base model.

Significance. If the findings hold, the paper would make two useful contributions: a scalable pipeline for synthesizing high-risk driving scenarios with automatic labels, and a simple projection-based visual prompting method that appears to close the modality gap between numerical waypoints and visual context. The paper includes experiments on multiple VLM backbones and a human-in-the-loop quality check for the synthetic data, which are strengths. However, the main evidence for real-world generalization rests entirely on a proprietary, undescribed dataset, and the synthetic benchmark is generated and labeled by the same rules. The central claim is therefore not yet supported to the standard required for publication.

major comments (4)
  1. [Section 3.2.1, 3.2.2, Table 1] The synthetic evaluation is confounded: trajectories are generated under rule-based definitions (collision distance threshold, acceleration magnitude and duration, lane-crossing geometry), and the same definitions are used to produce the labels in Section 3.2.2. Consequently, the 88.03% accuracy on the DriveMRP-10K test split may reflect rule mimicry rather than independent risk understanding. Please either provide an independent human-labeled subset to validate the synthetic labels, or explicitly present the synthetic benchmark as a controlled testbed for method development and place the generalization claim entirely on the real-world data.
  2. [Section 4.1, Table 2] The in-house real-world dataset is the sole external evidence for generalization, but it is described in only two sentences: 5,000 segments "exclusively feature genuine high-risk events," with no collection protocol, annotation instructions, class distribution, inter-annotator agreement, or release plan. Under these conditions the 68.50% accuracy cannot be interpreted: it could be inflated by an imbalanced class distribution or by label noise, and the absence of any "no risk" instances means the model is never tested on the full task definition. Please provide a detailed dataset description, class statistics, and a release plan, or replace this evaluation with a public benchmark.
  3. [Section 4.2, Table 2] The real-world comparison contrasts a fine-tuned model with zero-shot baselines; this conflates the contribution of synthetic fine-tuning with the contribution of the projection-based prompting framework. The paper does not report a zero-shot DriveMRP-Agent (projection prompting without fine-tuning) or a fine-tuned baseline with raw coordinate input on the real-world dataset, so the reader cannot tell what drives the 68.50% result. Adding these ablations on the real-world dataset would isolate the mechanism.
  4. [Section 3.3, Table 2] The evaluation metric section does not state how the free-text risk prediction is scored (exact match, keyword, or LLM-based) and does not report per-class metrics. With an all-high-risk dataset, accuracy alone is insufficient; balanced accuracy, precision and recall per category, and a confusion matrix are needed to assess whether the model actually discriminates among collision, acceleration, deceleration, and lane-change risks.
minor comments (6)
  1. [Section 3.2.2] The duplicated formula "(x,y)(x,y)" appears in the automated feature extraction paragraph; please correct it.
  2. [Appendix A.1] "We employed LoRA to supervise fine-tuning" should read "supervised fine-tuning."
  3. [Abstract] The abstract promises a GitHub repository, but no link or repository identifier appears in the manuscript; please include the URL or state that code will be released.
  4. [Figure 4] The prompt in Figure 4 says "past movement state," although the task is to evaluate the future planned trajectory; please align the wording with the method description.
  5. [Tables 1 and 3] Table 3 lists "Qwen2.5-VL-7B-Instruct" as the base model, while the method description says "Qwen2.5VL-7B"; please clarify whether the same checkpoint variant is used throughout.
  6. [Section 3.2.1] The polynomial trajectory generation is described only verbally; an explicit equation, degree, and sampling procedure would make the diversity and physical plausibility of the generated trajectories assessable.

Circularity Check

1 steps flagged · score 4.0 of 10

Synthetic benchmark is self-referential, but the real-world zero-shot test provides independent grounding; the central generalization claim is not forced by construction.

  1. self definitional [Section 3.2.1 (Rule Definition for High-Risk Scenarios), Section 3.2.2 (Automated Feature Extraction and Labeling), Section 4.1, Table 1]
    "we have formulated precise rule-based definitions for each selected high-risk scenario. For instance, collision events are determined by a predefined minimum safety distance threshold. ... Leveraging the simulator environment, we automatically extract exhaustive kinematic features from each synthesized trajectory. ... This dataset was partitioned into training and testing sets at an 8:2 ratio, yielding 8,000 samples for training and 2,000 samples for evaluation."

    The ground-truth labels for both the training and held-out test splits of DriveMRP-10K are produced by the same rule-based simulator that generates the high-risk trajectories. Accuracy on this test set (88.03% in Table 1) therefore measures how well the fine-tuned model reproduces the simulator's label function, not an independent estimate of real-world risk understanding. The 'prediction' is, by construction, a fit to the generator's own rule output. The paper partially escapes this loop with the in-house real-world zero-shot test, which is independent of the synthetic generator, though that dataset is proprietary, unbalanced (all high-risk), and not documented in enough detail to fully verify the transfer claim.

full rationale

The only concrete reduction-by-construction is the synthetic evaluation: DriveMRP-10K supplies both the training labels and the test labels through the same BEV simulator and rule definitions, so the reported 88.03% accuracy is a self-consistency measure rather than external validation. No load-bearing self-citation, uniqueness import, or ansatz-smuggling occurs; references to the authors' prior work (e.g., [38]) are not used to justify the central claim. The real-world zero-shot result on the in-house dataset is the key independent evidence, but it is unreleased, all-high-risk, and lacks collection/annotation details, which limits but does not constitute circularity. Overall partial circularity on the synthetic path, independent grounding on the real path: score 4.

Assumptions & free parameters 2 free parameters · 4 assumptions · 0 invented entities

The central claim rests on the validity of the private real-world labels, on nuPlan scenes being representative, and on GPT-4o captions being accurate. The synthetic benchmark is self-referential because labels come from the same rules that generate trajectories. No code or dataset is provided.

free parameters (2)
  • Risk-rule thresholds (collision distance, acceleration/deceleration magnitude and duration, lane-crossing geometry)
    Used to define high-risk scenarios and to auto-label trajectories in Section 3.2.1; no values are reported in the paper.
  • Polynomial trajectory coefficients and travel time per scenario
    Chosen per scene to reach estimated target points; the estimation procedure and polynomial order are not specified in detail.
assumptions (4)
  • domain assumption nuPlan provides realistic driving scenes that can support physically plausible high-risk simulation.
    Section 3.2 builds all synthetic data on nuPlan.
  • domain assumption GPT-4o generates accurate captions that correctly describe scene, trajectory, and risk.
    Section 3.2.3 uses GPT-4o for caption generation without quantitative validation of caption correctness.
  • domain assumption The in-house real-world dataset is correctly labeled and representative of long-tail risk.
    Section 4.1 introduces the dataset with no annotation protocol, class distribution, or release plan.
  • domain assumption Polynomial trajectories are valid approximations of vehicle dynamics for risk assessment.
    Section 3.2.1 uses polynomial generation to synthesize motions, relying on the assumption that these curves reflect real vehicle behavior.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DriveMRP: Enhancing Vision-Language Models with Synthetic Motion Data for Motion Risk Prediction." pith.science (2026). https://pith.science/paper/6IMBDHYR

@misc{pith2026250702948,
  author       = {Pith},
  title        = {Pith review of: DriveMRP: Enhancing Vision-Language Models with Synthetic Motion Data for Motion Risk Prediction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/6IMBDHYR}},
  note         = {Machine review of arXiv:2507.02948}
}
read the original abstract

Autonomous driving has seen significant progress, driven by extensive real-world data. However, in long-tail scenarios, accurately predicting the safety of the ego vehicle's future motion remains a major challenge due to uncertainties in dynamic environments and limitations in data coverage. In this work, we aim to explore whether it is possible to enhance the motion risk prediction capabilities of Vision-Language Models (VLM) by synthesizing high-risk motion data. Specifically, we introduce a Bird's-Eye View (BEV) based motion simulation method to model risks from three aspects: the ego-vehicle, other vehicles, and the environment. This allows us to synthesize plug-and-play, high-risk motion data suitable for VLM training, which we call DriveMRP-10K. Furthermore, we design a VLM-agnostic motion risk estimation framework, named DriveMRP-Agent. This framework incorporates a novel information injection strategy for global context, ego-vehicle perspective, and trajectory projection, enabling VLMs to effectively reason about the spatial relationships between motion waypoints and the environment. Extensive experiments demonstrate that by fine-tuning with DriveMRP-10K, our DriveMRP-Agent framework can significantly improve the motion risk prediction performance of multiple VLM baselines, with the accident recognition accuracy soaring from 27.13% to 88.03%. Moreover, when tested via zero-shot evaluation on an in-house real-world high-risk motion dataset, DriveMRP-Agent achieves a significant performance leap, boosting the accuracy from base_model's 29.42% to 68.50%, which showcases the strong generalization capabilities of our method in real-world scenarios.

Figures

Figures reproduced from arXiv: 2507.02948 by the authors.

Figure 1
Figure 1. (a) Previous rule-based pipelines rely on external models to predict other vehicles’ future [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. The proposed DriveMRP-10K simulation pipeline consists of the following stages: high [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. The proposed DriveMRP-Agent framework. Taking BEV layout, scene images, and [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (1 more)
Figure 4
Figure 4. Figure 4: This figure shows the performance comparison between our model and other VLM models [PITH_FULL_IMAGE:figures/full_fig_p008_4.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. NSF-HRPT: Neural Semantic Field meets Hierarchical Risk Perception Tree for Safety-Critical Scenario Assessment

    cs.AI 2026-08 conditional novelty 6.0 of 10

    A simulator-trained neural semantic field plus a hierarchical risk tree estimates per-agent collision risk and time-to-collision from monocular video, with foundation-model features used to close the sim-to-real gap w...

Reference graph

Works this paper leans on

39 extracted references · 14 canonical work pages · cited by 1 Pith paper

  1. [1]

    Gpt-4 technical report

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ahmad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report. arXiv preprint arXiv:2303.08774, 2023

    arXiv 2023

  2. [2]

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, et al. Qwen2. 5-vl technical report. arXiv preprint arXiv:2502.13923, 2025

    arXiv 2025

  3. [3]

    nuscenes: A multimodal dataset for autonomous driving

    Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Giancarlo Baldan, and Oscar Beijbom. nuscenes: A multimodal dataset for autonomous driving. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 11621–11631, 2020

    2020

  4. [4]

    nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles

    Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810, 2021

    arXiv 2021

  5. [5]

    Autonomous vehicles in mixed- autonomy traffic: game theoretic human-like decision making countermeasures

    Qitong Chen, Dong Zhao, Congzhi Liu, Meng Yang, and Yehui Shi. Autonomous vehicles in mixed- autonomy traffic: game theoretic human-like decision making countermeasures. Complex Engineering Systems, 4(4):N–A, 2024

    work page 2024

  6. [6]

    Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks

    Zhe Chen, Jiannan Wu, Wenhai Wang, Weijie Su, Guo Chen, Sen Xing, Muyan Zhong, Qinglong Zhang, Xizhou Zhu, Lewei Lu, et al. Internvl: Scaling up vision foundation models and aligning for generic visual-linguistic tasks. In Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 24185–24198, 2024

    2024

  7. [7]

    Rouge: A package for automatic evaluation of summaries

    Lin Chin-Yew. Rouge: A package for automatic evaluation of summaries. In Proceedings of the Workshop on Text Summarization Branches Out, 2004, 2004

    work page 2004

  8. [8]

    Flashattention: Fast and memory-efficient exact attention with io-awareness

    Tri Dao, Dan Fu, Stefano Ermon, Atri Rudra, and Christopher Ré. Flashattention: Fast and memory-efficient exact attention with io-awareness. Advances in neural information processing systems, 35:16344–16359, 2022

    2022

Show all 39 references
  1. [9]

    Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking

    Daniel Dauner, Marcel Hallgarten, Tianyu Li, Xinshuo Weng, Zhiyu Huang, Zetong Yang, Hongyang Li, Igor Gilitschenski, Boris Ivanovic, Marco Pavone, et al. Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking. Advances in Neural Information Processing...

    2024

  2. [10]

    Carla: An open urban driving simulator

    Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on robot learning, pages 1–16. PMLR, 2017

    2017

  3. [11]

    Vista: A generalizable driving world model with high fidelity and versatile controllability

    Shenyuan Gao, Jiazhi Yang, Li Chen, Kashyap Chitta, Yihang Qiu, Andreas Geiger, Jun Zhang, and Hongyang Li. Vista: A generalizable driving world model with high fidelity and versatile controllability. arXiv preprint arXiv:2405.17398, 2024

    2024 arXiv

  4. [12]

    Multi-frame, lightweight & efficient vision- language models for question answering in autonomous driving

    Akshay Gopalkrishnan, Ross Greer, and Mohan Trivedi. Multi-frame, lightweight & efficient vision- language models for question answering in autonomous driving. arXiv preprint arXiv:2403.19838, 2024

    2024 arXiv

  5. [13]

    A probabilistic interpretation of precision, recall and f-score, with implication for evaluation

    Cyril Goutte and Eric Gaussier. A probabilistic interpretation of precision, recall and f-score, with implication for evaluation. In European conference on information retrieval, pages 345–359. Springer, 2005

    2005

  6. [14]

    The llama 3 herd of models

    Aaron Grattafiori, Abhimanyu Dubey, Abhinav Jauhri, Abhinav Pandey, Abhishek Kadian, Ahmad Al- Dahle, Aiesha Letman, Akhil Mathur, Alan Schelten, Alex Vaughan, et al. The llama 3 herd of models. arXiv preprint arXiv:2407.21783, 2024

    2024 arXiv

  7. [15]

    Dme-driver: Integrating human decision logic and 3d scene perception in autonomous driving

    Wencheng Han, Dongqian Guo, Cheng-Zhong Xu, and Jianbing Shen. Dme-driver: Integrating human decision logic and 3d scene perception in autonomous driving. In Proceedings of the AAAI Conference on Artificial Intelligence, volume 39, pages 3347–3355, 2025

    2025

  8. [16]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Yelong Shen, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. ICLR, 1(2):3, 2022. 10

    2022

  9. [17]

    Planning-oriented autonomous driving

    Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853–17862, 2023

    2023

  10. [18]

    Gen-drive: Enhancing diffusion generative driving policies with reward modeling and reinforcement learning fine-tuning

    Zhiyu Huang, Xinshuo Weng, Maximilian Igl, Yuxiao Chen, Yulong Cao, Boris Ivanovic, Marco Pavone, and Chen Lv. Gen-drive: Enhancing diffusion generative driving policies with reward modeling and reinforcement learning fine-tuning. arXiv preprint arXiv:2410.05582, 2024

    arXiv

  11. [19]

    Vad: Vectorized scene representation for efficient autonomous driving

    Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representation for efficient autonomous driving. 2023 IEEE/CVF International Conference on Computer Vision (ICCV), pages 8306–8...

    2023

  12. [20]

    Comparison of waymo rider-only crash rates by crash type to human benchmarks at 56.7 million miles

    Kristofer D Kusano, John M Scanlon, Yin-Hsiu Chen, Timothy L McMurry, Tilia Gode, and Trent Victor. Comparison of waymo rider-only crash rates by crash type to human benchmarks at 56.7 million miles. arXiv preprint arXiv:2505.01515, 2025

    2025 arXiv

  13. [21]

    Metadrive: Composing diverse driving scenarios for generalizable reinforcement learning

    Quanyi Li, Zhenghao Peng, Lan Feng, Qihang Zhang, Zhenghai Xue, and Bolei Zhou. Metadrive: Composing diverse driving scenarios for generalizable reinforcement learning. IEEE transactions on pattern analysis and machine intelligence, 45(3):3461–3475, 2022

    2022

  14. [22]

    Bevformer: learning bird’s-eye-view representation from lidar-camera via spatiotemporal transformers

    Zhiqi Li, Wenhai Wang, Hongyang Li, Enze Xie, Chonghao Sima, Tong Lu, Qiao Yu, and Jifeng Dai. Bevformer: learning bird’s-eye-view representation from lidar-camera via spatiotemporal transformers. IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  15. [23]

    Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open-loop end-to-end autonomous driving? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14864–14873, 2024

    2024

  16. [24]

    Maptr: Structured modeling and learning for online vectorized hd map construction

    Bencheng Liao, Shaoyu Chen, Xinggang Wang, Tianheng Cheng, Qian Zhang, Wenyu Liu, and Chang Huang. Maptr: Structured modeling and learning for online vectorized hd map construction. arXiv preprint arXiv:2208.14437, 2022

    2022 arXiv

  17. [25]

    Improved baselines with visual instruction tuning

    Haotian Liu, Chunyuan Li, Yuheng Li, and Yong Jae Lee. Improved baselines with visual instruction tuning. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 26296–26306, 2024

    2024

  18. [26]

    Gpt-driver: Learning to drive with gpt

    Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023

    2023 arXiv

  19. [27]

    Autonomous driving in unstructured environments: How far have we come? arXiv preprint arXiv:2410.07701, 2024

    Chen Min, Shubin Si, Xu Wang, Hanzhang Xue, Weizhong Jiang, Yang Liu, Juan Wang, Qingtian Zhu, Qi Zhu, Lun Luo, et al. Autonomous driving in unstructured environments: How far have we come? arXiv preprint arXiv:2410.07701, 2024

  20. [28]

    Safe navigation in unstructured environments by minimizing uncertainty in control and perception

    Junwon Seo, Jungwi Mun, and Taekyung Kim. Safe navigation in unstructured environments by minimizing uncertainty in control and perception. arXiv preprint arXiv:2306.14601, 2023

    2023 arXiv

  21. [29]

    Drivelm: Driving with graph visual question answering

    Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beißwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. In European Conference on Computer Vision, pages 256–274. Springer, 2024

    2024

  22. [30]

    Drivevlm: The convergence of autonomous driving and large vision-language models

    Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Yang Wang, Zhiyong Zhao, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models. arXiv preprint arXiv:2402.12289, 2024

    2024 arXiv

  23. [31]

    He-drive: Human-like end-to-end driving with vision language models

    Junming Wang, Xingyu Zhang, Zebin Xing, Songen Gu, Xiaoyang Guo, Yang Hu, Ziying Song, Qian Zhang, Xiaoxiao Long, and Wei Yin. He-drive: Human-like end-to-end driving with vision language models. arXiv preprint arXiv:2410.05051, 2024

  24. [32]

    Omnidrive: A holistic llm-agent framework for autonomous driving with 3d perception, reasoning and planning

    Shihao Wang, Zhiding Yu, Xiaohui Jiang, Shiyi Lan, Min Shi, Nadine Chang, Jan Kautz, Ying Li, and Jose M Alvarez. Omnidrive: A holistic llm-agent framework for autonomous driving with 3d perception, reasoning and planning. arXiv preprint arXiv:2405.01533, 2024

    2024 arXiv

  25. [33]

    Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving

    Yuqi Wang, Jiawei He, Lue Fan, Hongxin Li, Yuntao Chen, and Zhaoxiang Zhang. Driving into the future: Multiview visual forecasting and planning with world model for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages ...

    2024

  26. [34]

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

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models. Advances in neural information processing systems, 35:24824–24837, 2022. 11

    2022

  27. [35]

    Bertscore: Evaluating text generation with bert

    Tianyi Zhang, Varsha Kishore, Felix Wu, Kilian Q Weinberger, and Yoav Artzi. Bertscore: Evaluating text generation with bert. arXiv preprint arXiv:1904.09675, 2019

    1904 arXiv

  28. [36]

    Perception and sensing for autonomous vehicles under adverse weather conditions: A survey

    Yuxiao Zhang, Alexander Carballo, Hanting Yang, and Kazuya Takeda. Perception and sensing for autonomous vehicles under adverse weather conditions: A survey. ISPRS Journal of Photogrammetry and Remote Sensing, 196:146–177, 2023

    2023

  29. [37]

    Llamafactory: Unified efficient fine-tuning of 100+ language models

    Yaowei Zheng, Richong Zhang, Junhao Zhang, Yanhan Ye, Zheyan Luo, Zhangchi Feng, and Yongqiang Ma. Llamafactory: Unified efficient fine-tuning of 100+ language models. In Proceedings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System ...

    2024

  30. [38]

    Ua-track: Uncertainty-aware end-to-end 3d multi-object tracking

    Lijun Zhou, Tao Tang, Pengkun Hao, Zihang He, Kalok Ho, Shuo Gu, Wenbo Hou, Zhihui Hao, Haiyang Sun, Kun Zhan, et al. Ua-track: Uncertainty-aware end-to-end 3d multi-object tracking. arXiv preprint arXiv:2406.02147, 2024

    2024 arXiv

  31. [39]

    Embodied understanding of driving scenarios

    Yunsong Zhou, Linyan Huang, Qingwen Bu, Jia Zeng, Tianyu Li, Hang Qiu, Hongzi Zhu, Minyi Guo, Yu Qiao, and Hongyang Li. Embodied understanding of driving scenarios. In European Conference on Computer Vision, pages 129–148. Springer, 2024. A Technical Details A.1 Training Detai...

    2024

Pith tools

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