pith. sign in

arxiv: 2605.19524 · v1 · pith:6X5Z4SSUnew · submitted 2026-05-19 · 💻 cs.RO · cs.CV

SafeAlign-VLA: A Negative-Enhanced Safe Alignment Framework for Risk-Aware Autonomous Driving

Kefei Tian , Yuansheng Lian , Kai Yang , Xiangdong Chen , Shen Li This is my paper

Pith reviewed 2026-05-20 05:19 UTC · model grok-4.3

classification 💻 cs.RO cs.CV
keywords autonomous drivingvision-language-actionsafe alignmentnegative datacounterfactual reasoningpolicy optimizationcollision avoidancerisk prediction
0
0 comments X

The pith

Negative data paired with counterfactual reasoning improves safety in vision-language-action autonomous driving models.

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

The paper tries to establish that end-to-end autonomous driving systems can learn safer behavior by training on both successful and failed driving examples rather than positive demonstrations alone. It shows how to turn risky scenarios into paired safety labels and corrected trajectories using counterfactual reasoning, then applies a two-stage process of supervised correction followed by contrastive policy optimization that penalizes high-risk actions. A sympathetic reader would care because current models perform well in ordinary conditions but fail in rare dangerous cases that cause accidents. The reported gains include higher overall driving scores and substantially lower collision rates on standard test sets.

Core claim

SafeAlign-VLA is a unified framework that incorporates negative data into VLA models for autonomous driving. It first applies a counterfactual safety pairing paradigm to produce structured safety labels and counterfactual positive trajectories from risky scenarios. This is followed by negative-enhanced supervised fine-tuning for trajectory correction and then anchor-based group relative policy optimization that treats positive and negative trajectories as contrastive anchors to calculate group-relative advantages and reduce high-risk sampling.

What carries the argument

The counterfactual safety pairing paradigm that generates structured safety labels and corrected trajectories from risky scenarios, together with anchor-based group relative policy optimization that contrasts positive and negative trajectories to steer sampling.

If this is right

  • Higher PDMS scores on NAVSIM v1 by adding negative data to the baseline.
  • Lower collision rates on DeepAccident while preserving language and risk prediction accuracy.
  • Explicit modeling of safety boundaries in long-tail scenarios through contrastive training.
  • Improved failure feedback during supervised fine-tuning for trajectory correction.

Where Pith is reading between the lines

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

  • The pairing technique could be reused to generate synthetic safety data for other sequential decision tasks outside driving.
  • Consistent safety gains might support regulatory testing requirements that demand evidence of handling rare events.
  • The contrastive anchor method could be combined with real-time risk sensors to adjust behavior on the fly.

Load-bearing premise

Counterfactual reasoning applied to risky driving scenarios can reliably produce accurate safety labels and safe positive trajectories.

What would settle it

A controlled experiment that replaces the generated counterfactual positive trajectories with deliberately flawed ones and checks whether the reported reductions in collision rate and gains in PDMS score disappear.

Figures

Figures reproduced from arXiv: 2605.19524 by Kai Yang, Kefei Tian, Shen Li, Xiangdong Chen, Yuansheng Lian.

Figure 1
Figure 1. Figure 1: The proposed SafeAlign-VLA framework for risk-aware autonomous driving. The framework [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overall framework of the proposed method. Stage 1 performs negative-enhanced curriculum [PITH_FULL_IMAGE:figures/full_fig_p008_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Overview of the negative-enhanced mixed SFT stage. Given input data and the paired [PITH_FULL_IMAGE:figures/full_fig_p009_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative results of our framework in urban intersection scenarios. (a) A safe right-turn [PITH_FULL_IMAGE:figures/full_fig_p016_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative comparison of predicted trajectories between Base SFT (ID-1) and Full GRPO [PITH_FULL_IMAGE:figures/full_fig_p017_5.png] view at source ↗
read the original abstract

End-to-end autonomous driving systems excel in common scenarios but struggle with safety-critical long-tail cases. Vision-Language-Action (VLA) models are promising due to their strong reasoning capabilities. However, most VLA-based approaches rely on positive expert demonstrations, rarely exploiting negative samples, leading to insufficient understanding of risky behaviors and safety boundaries. To address this limitation, we propose SafeAlign-VLA, a unified negative-enhanced safe alignment framework that incorporates negative data into supervised learning and reinforcement learning. First, we develop a counterfactual safety pairing paradigm to generate structured safety labels and counterfactual positive trajectories from risky scenarios via counterfactual reasoning. Then, a two-stage training strategy is adopted: negative-enhanced supervised fine-tuning for failure feedback and trajectory correction, followed by anchor-based group relative policy optimization that uses positive and negative trajectories as contrastive anchors to steer sampling and penalize high-risk behaviors via group-relative advantages. Experiments on NAVSIM and DeepAccident validate the proposed framework. SafeAlign-VLA achieves 89.1 PDMS on the NAVSIM v1 testset, improving over the baseline without negative data by 1.3%. On DeepAccident, it reduces the collision rate to 3.36%, while achieving 84.2% language accuracy and 85.8% risk prediction accuracy. These results demonstrate the effectiveness of the proposed negative-enhanced safe alignment framework for safe and robust autonomous driving.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

3 major / 2 minor

Summary. The manuscript proposes SafeAlign-VLA, a negative-enhanced safe alignment framework for Vision-Language-Action (VLA) models in autonomous driving. It introduces a counterfactual safety pairing paradigm to generate structured safety labels and counterfactual positive trajectories from risky scenarios via the model's reasoning. This feeds into a two-stage process: negative-enhanced supervised fine-tuning for failure feedback and trajectory correction, followed by anchor-based group-relative policy optimization that treats positive and negative trajectories as contrastive anchors to compute group-relative advantages and penalize high-risk behaviors. Experiments on the NAVSIM v1 test set and DeepAccident benchmark report a PDMS score of 89.1 (1.3% above the no-negative-data baseline), a collision rate of 3.36%, 84.2% language accuracy, and 85.8% risk prediction accuracy.

Significance. If the counterfactual pairing reliably produces accurate positive trajectories and labels, the framework could advance risk-aware end-to-end driving by systematically incorporating negative samples to clarify safety boundaries in long-tail cases. The use of established external benchmarks (NAVSIM, DeepAccident) keeps the evaluation grounded, and the contrastive anchor mechanism in the policy optimization stage offers a concrete way to steer sampling away from risky behaviors. These elements would represent a useful contribution to safe VLA alignment if the unverified data-generation step is substantiated.

major comments (3)
  1. [Section 3.2] Section 3.2 (Counterfactual Safety Pairing): The paradigm generates structured safety labels and positive trajectories from risky scenarios using the model's own counterfactual reasoning, yet the manuscript supplies no independent validation against ground-truth safe behaviors, simulator rollouts, or expert annotations. This step is load-bearing for the central claims, as both the negative-enhanced SFT and the subsequent anchor-based optimization optimize against these synthesized anchors; systematic errors here would propagate directly into the reported 89.1 PDMS and 3.36% collision-rate figures.
  2. [Section 4] Section 4 (Experiments): The performance numbers (89.1 PDMS on NAVSIM v1, 3.36% collision rate on DeepAccident) are presented without statistical significance tests, variance across multiple runs, or ablation studies that isolate the counterfactual pairing from other training components. Without these controls, it is unclear whether the 1.3% improvement and collision reduction are attributable to genuine safety-boundary learning or to artifacts in the generated data.
  3. [§3.3] §3.3 (Anchor-based Group Relative Policy Optimization): The group-relative advantages are computed from positive and negative trajectory pairs to penalize high-risk behaviors, but the absence of any external check on the accuracy of the counterfactual positives means the optimization may be reinforcing flawed or overly optimistic trajectories rather than true safety improvements.
minor comments (2)
  1. [Abstract] Abstract: The acronym PDMS is used without definition; expand it on first use and briefly state what it measures.
  2. [Throughout] Throughout: Several method diagrams and result tables would benefit from explicit captions that highlight the role of negative data versus the baseline, improving readability for readers unfamiliar with the two-stage pipeline.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive and detailed feedback on our manuscript. We address each major comment below and describe the revisions we intend to incorporate.

read point-by-point responses

  1. Referee: [Section 3.2] Section 3.2 (Counterfactual Safety Pairing): The paradigm generates structured safety labels and positive trajectories from risky scenarios using the model's own counterfactual reasoning, yet the manuscript supplies no independent validation against ground-truth safe behaviors, simulator rollouts, or expert annotations. This step is load-bearing for the central claims, as both the negative-enhanced SFT and the subsequent anchor-based optimization optimize against these synthesized anchors; systematic errors here would propagate directly into the reported 89.1 PDMS and 3.36% collision-rate figures.

    Authors: We agree that independent validation of the counterfactual safety pairing would strengthen the central claims. The current approach leverages the VLA model's reasoning to synthesize positive trajectories from risky scenarios, which is a core contribution. In the revised manuscript, we will expand Section 3.2 with a new validation subsection that includes (i) qualitative examples of the counterfactual reasoning process and (ii) quantitative consistency checks on a sampled subset of scenarios against simulator rollouts for safety metrics such as collision avoidance. These additions will help substantiate the quality of the generated anchors. revision: yes

  2. Referee: [Section 4] Section 4 (Experiments): The performance numbers (89.1 PDMS on NAVSIM v1, 3.36% collision rate on DeepAccident) are presented without statistical significance tests, variance across multiple runs, or ablation studies that isolate the counterfactual pairing from other training components. Without these controls, it is unclear whether the 1.3% improvement and collision reduction are attributable to genuine safety-boundary learning or to artifacts in the generated data.

    Authors: We acknowledge the value of greater statistical rigor and component isolation. In the revised Section 4, we will report all key metrics with standard deviations computed over multiple independent runs (different random seeds) and add ablation studies that systematically remove the counterfactual pairing step while keeping other components fixed. We will also include statistical significance tests (e.g., paired t-tests) comparing SafeAlign-VLA against the no-negative-data baseline to support the reported 1.3% PDMS gain and collision-rate reduction. revision: yes

  3. Referee: [§3.3] §3.3 (Anchor-based Group Relative Policy Optimization): The group-relative advantages are computed from positive and negative trajectory pairs to penalize high-risk behaviors, but the absence of any external check on the accuracy of the counterfactual positives means the optimization may be reinforcing flawed or overly optimistic trajectories rather than true safety improvements.

    Authors: This concern is directly linked to the validation issue raised for Section 3.2. Once the planned validation of counterfactual positives is added, we will augment §3.3 with a discussion and supporting analysis showing that the anchor-based group-relative advantages successfully steer the policy away from high-risk behaviors, as reflected in the observed collision-rate reductions on DeepAccident. We will clarify that the contrastive mechanism operates on the synthesized pairs to clarify safety boundaries even when the positives are model-generated rather than externally verified. revision: partial

Circularity Check

0 steps flagged

No significant circularity; framework grounded in external benchmarks

full rationale

The paper describes a counterfactual safety pairing paradigm and two-stage training (negative-enhanced SFT followed by anchor-based group-relative policy optimization) without presenting any equations, derivations, or first-principles results that reduce to fitted parameters or self-defined quantities by construction. Performance claims rest on evaluation against independent external benchmarks (NAVSIM v1 testset and DeepAccident), not on internal consistency checks or self-generated labels alone. No load-bearing self-citations, uniqueness theorems, or ansatzes imported from prior author work are invoked to justify core steps. The method introduces new components but does not rename known empirical patterns or force predictions via statistical construction from the same inputs. This is a standard non-circular empirical framework paper.

Axiom & Free-Parameter Ledger

0 free parameters · 1 axioms · 0 invented entities

Based on abstract only; the central claim rests on the ability of counterfactual reasoning to produce usable training pairs and on the effectiveness of the two-stage training without further specification of hyperparameters or loss terms.

axioms (1)
  • domain assumption Counterfactual reasoning applied to risky driving scenarios produces accurate positive trajectories and structured safety labels
    Invoked in the first component of the framework to generate training data from negative samples.

pith-pipeline@v0.9.0 · 5792 in / 1186 out tokens · 40334 ms · 2026-05-20T05:19:05.808193+00:00 · methodology

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

59 extracted references · 59 canonical work pages · 12 internal anchors

  1. [1]

    J. Zhao, Y. Wu, R. Deng, S. Xu, J. Gao, A. Burke, A survey of autonomous driving from a deep learning perspective, ACM Computing Surveys 57 (10) (2025) 1–60

    work page 2025

  2. [2]

    Y. Lian, K. Zhang, Y. Guo, S. Li, M. Li, Bap-srl: Bayesian adaptive priority safe reinforcement learning for vehicle motion planning at mixed traffic intersections, arXiv preprint arXiv:2601.21679 (2026)

    work page arXiv 2026

  3. [3]

    S. Li, K. Yang, Z. Wei, Y. Zheng, Z. Chen, X. Tang, A survey on interaction-aware decision-making for autonomous driving: Challenges, solutions, and perspectives, IEEE Transactions on Intelligent Transportation Systems (2026)

    work page 2026

  4. [4]

    Hussain, S

    R. Hussain, S. Zeadally, Autonomous cars: Research results, issues, and future chal- lenges, IEEE Communications Surveys & Tutorials 21 (2) (2018) 1275–1313

    work page 2018

  5. [5]

    Y. Lian, K. Zhang, M. Li, Cdkformer: Contextual deviation knowledge-based trans- former for long-tail trajectory prediction, Transportation Research Part C: Emerging Technologies 183 (2026) 105430

    work page 2026

  6. [6]

    S. Hu, L. Chen, P. Wu, H. Li, J. Yan, D. Tao, St-p3: End-to-end vision-based au- tonomous driving via spatial-temporal feature learning, in: European Conference on Computer Vision, Springer, 2022, pp. 533–549

    work page 2022

  7. [7]

    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. 17853–17862

    work page 2023

  8. [8]

    S. Chen, B. Jiang, H. Gao, B. Liao, Q. Xu, Q. Zhang, C. Huang, W. Liu, X. Wang, Vadv2: End-to-end vectorized autonomous driving via probabilistic planning, arXiv preprint arXiv:2402.13243 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  9. [9]

    L. Chen, P. Wu, K. Chitta, B. Jaeger, A. Geiger, H. Li, End-to-end autonomous driv- ing: Challenges and frontiers, IEEE Transactions on Pattern Analysis and Machine Intelligence 46 (12) (2024) 10164–10183

    work page 2024

  10. [10]

    Codevilla, E

    F. Codevilla, E. Santana, A. M. López, A. Gaidon, Exploring the limitations of behav- ior cloning for autonomous driving, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2019, pp. 9329–9338

    work page 2019

  11. [11]

    Omeiza, H

    D. Omeiza, H. Webb, M. Jirotka, L. Kunze, Explanations in autonomous driving: A survey, IEEE Transactions on Intelligent Transportation Systems 23 (8) (2022) 10142– 10162

    work page 2022

  12. [12]

    X. Hu, Y. Lian, M. Li, K. Zhang, Y. Li, Y. Su, Lift: Interpretable truck driving risk prediction with literature-informed fine-tuned llms, Transportation Research Part C: Emerging Technologies 185 (2026) 105570. 18

    work page 2026

  13. [13]

    EMMA: End-to-End Multimodal Model for Autonomous Driving

    J.-J. Hwang, R. Xu, H. Lin, W.-C. Hung, J. Ji, K. Choi, D. Huang, T. He, P. Covington, B. Sapp, et al., Emma: End-to-end multimodal model for autonomous driving, arXiv preprint arXiv:2410.23262 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  14. [14]

    C. Sima, K. Renz, K. Chitta, L. Chen, H. Zhang, C. Xie, J. Beißwenger, P. Luo, A. Geiger, H. Li, Drivelm: Driving with graph visual question answering, in: European Conference on Computer Vision, Springer, 2024, pp. 256–274

    work page 2024

  15. [15]

    Marcu, L

    A.-M. Marcu, L. Chen, J. Hünermann, A. Karnsund, B. Hanotte, P. Chidananda, S. Nair, V. Badrinarayanan, A. Kendall, J. Shotton, et al., Lingoqa: Visual ques- tion answering for autonomous driving, in: European Conference on Computer Vision, Springer, 2024, pp. 252–269

    work page 2024

  16. [16]

    Y. Wang, S. Wu, Y. Zhang, S. Yan, Z. Liu, J. Luo, H. Fei, Multimodal chain-of-thought reasoning: A comprehensive survey, arXiv preprint arXiv:2503.12605 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  17. [17]

    M. Nie, R. Peng, C. Wang, X. Cai, J. Han, H. Xu, L. Zhang, Reason2drive: Towards in- terpretable and chain-based reasoning for autonomous driving, in: European Conference on Computer Vision, Springer, 2024, pp. 292–308

    work page 2024

  18. [18]

    Z. Xu, Y. Zhang, E. Xie, Z. Zhao, Y. Guo, K.-Y. K. Wong, Z. Li, H. Zhao, Drivegpt4: Interpretable end-to-end autonomous driving via large language model, IEEE Robotics and Automation Letters 9 (10) (2024) 8186–8193

    work page 2024

  19. [19]

    H. Shao, Y. Hu, L. Wang, G. Song, S. L. Waslander, Y. Liu, 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. 15120–15130

    work page 2024

  20. [20]

    X. Zhou, X. Han, F. Yang, Y. Ma, V. Tresp, A. Knoll, Opendrivevla: Towards end- to-end autonomous driving with large vision language action model, in: Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 40, 2026, pp. 13782–13790

    work page 2026

  21. [21]

    X. Tian, J. Gu, B. Li, Y. Liu, Y. Wang, Z. Zhao, K. Zhan, P. Jia, X. Lang, H. Zhao, DriveVLM: The convergence of autonomous driving and large vision-language models, arXiv preprint arXiv:2402.12289 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  22. [22]

    S. Wang, Z. Yu, X. Jiang, S. Lan, M. Shi, N. Chang, J. Kautz, Y. Li, J. M. Alvarez, Om- nidrive: A holistic vision-language dataset for autonomous driving with counterfactual reasoning, in: Proceedings of the computer vision and pattern recognition conference, 2025, pp. 22442–22452

    work page 2025

  23. [23]

    Drivedpo: Policy learning via safety dpo for end-to-end autonomous driving.arXiv preprint arXiv:2509.17940, 2025

    S. Shang, Y. Chen, Y. Wang, Y. Li, Z. Zhang, Drivedpo: Policy learning via safety dpo for end-to-end autonomous driving, arXiv preprint arXiv:2509.17940 (2025)

    work page arXiv 2025

  24. [24]

    Jaeger, K

    B. Jaeger, K. Chitta, A. Geiger, Hidden biases of end-to-end driving models, in: Pro- ceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 8240–8249. 19

    work page 2023

  25. [25]

    Y. Lu, J. Fu, G. Tucker, X. Pan, E. Bronstein, R. Roelofs, B. Sapp, B. White, A. Faust, S. Whiteson, et al., Imitation is not enough: Robustifying imitation with reinforcement learning for challenging driving scenarios, in: 2023 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS), IEEE, 2023, pp. 7553–7560

    work page 2023

  26. [26]

    Z. Wang, S. Lan, X. Sun, N. Chang, Z. Li, Z. Yu, J. M. Alvarez, Enhancing autonomous driving safety with collision scenario integration, in: 2025 IEEE/RSJ International Con- ference on Intelligent Robots and Systems (IROS), IEEE, 2025, pp. 10116–10123

    work page 2025

  27. [27]

    Patrikar, A

    J. Patrikar, A. Sharma, S. Veer, B. Li, S. Scherer, M. Pavone, The case for nega- tive data: From crash reports to counterfactuals for reasonable driving, arXiv preprint arXiv:2509.18626 (2025)

    work page arXiv 2025

  28. [28]

    Vision-language- action models: Concepts, progress, applications and chal- lenges.arXiv preprint arXiv:2505.04769,

    R. Sapkota, Y. Cao, K. I. Roumeliotis, M. Karkee, Vision-language-action models: Con- cepts, progress, applications and challenges, arXiv preprint arXiv:2505.04769 (2025)

    work page arXiv 2025

  29. [29]

    T. Hu, X. Liu, S. Wang, Y. Zhu, A. Liang, L. Kong, G. Zhao, Z. Gong, J. Cen, Z. Huang, et al., Vision-language-action models for autonomous driving: Past, present, and future, arXiv preprint arXiv:2512.16760 (2025)

    work page arXiv 2025

  30. [30]

    Senna: Bridging Large Vision-Language Models and End-to-End Autonomous Driving

    B. Jiang, S. Chen, B. Liao, X. Zhang, W. Yin, Q. Zhang, C. Huang, W. Liu, X. Wang, Senna: Bridging large vision-language models and end-to-end autonomous driving, arXiv preprint arXiv:2410.22313 (2024)

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  31. [31]

    AlphaDrive: Unleashing the Power of VLMs in Autonomous Driving via Reinforcement Learning and Reasoning

    B. Jiang, S. Chen, Q. Zhang, W. Liu, X. Wang, Alphadrive: Unleashing the power of vlms in autonomous driving via reinforcement learning and reasoning, arXiv preprint arXiv:2503.07608 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  32. [32]

    H. Fu, D. Zhang, Z. Zhao, J. Cui, D. Liang, C. Zhang, D. Zhang, H. Xie, B. Wang, X. Bai, Orion: A holistic end-to-end autonomous driving framework by vision-language instructed action generation, in: Proceedings of the IEEE/CVF International Confer- ence on Computer Vision, 2025, pp. 24823–24834

    work page 2025

  33. [33]

    R. Zhao, Q. Yuan, J. Li, H. Hu, Y. Li, Z. Gao, F. Gao, Sce2drivex: A generalized mllm framework for scene-to-drive learning, IEEE Robotics and Automation Letters (2025)

    work page 2025

  34. [34]

    Y. Li, K. Xiong, X. Guo, F. Li, S. Yan, G. Xu, L. Zhou, L. Chen, H. Sun, B. Wang, et al., Recogdrive: A reinforced cognitive framework for end-to-end autonomous driving, arXiv preprint arXiv:2506.08052 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  35. [35]

    K. Renz, L. Chen, E. Arani, O. Sinavski, Simlingo: Vision-only closed-loop autonomous driving with language-action alignment, in: Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 11993–12003

    work page 2025

  36. [36]

    K. Tian, Y. Lian, K. Yang, X. Chen, S. Li, C-cot: Counterfactual chain-of-thought with vision-language models for safe autonomous driving, arXiv preprint arXiv:2605.10744 (2026). 20

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  37. [37]

    Pan, C.-A

    Y. Pan, C.-A. Cheng, K. Saigol, K. Lee, X. Yan, E. Theodorou, B. Boots, Ag- ile autonomous driving using end-to-end deep imitation learning, arXiv preprint arXiv:1709.07174 (2017)

    work page arXiv 2017

  38. [38]

    Pluto: Pushing the limit of imitation learning- based planning for autonomous driving.ArXiv, abs/2404.14327, 2024

    J. Cheng, Y. Chen, Q. Chen, Pluto: Pushing the limit of imitation learning-based planning for autonomous driving, arXiv preprint arXiv:2404.14327 (2024)

    work page arXiv 2024

  39. [39]

    W. Sun, X. Lin, Y. Shi, C. Zhang, H. Wu, S. Zheng, Sparsedrive: End-to-end au- tonomous driving via sparse scene representation, in: 2025 IEEE International Confer- ence on Robotics and Automation (ICRA), IEEE, 2025, pp. 8795–8801

    work page 2025

  40. [40]

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

    work page 2022

  41. [41]

    Q. Lu, X. Wang, R. Yuan, W. Lu, X. Gong, S. Feng, Controllable risk scenario generation from human crash data for autonomous vehicle testing, arXiv preprint arXiv:2512.07874 (2025)

    work page arXiv 2025

  42. [42]

    ChauffeurNet: Learning to Drive by Imitating the Best and Synthesizing the Worst

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

    work page internal anchor Pith review Pith/arXiv arXiv 2018

  43. [43]

    Z. Zhou, R. Yang, Y. Guo, S. X. Chen, T. Feng, K. Pistunova, Y. Shen, L. Su, J. Ma, et al., Spanvla: Efficient action bridging and learning from negative-recovery samples for vision-language-action model, arXiv preprint arXiv:2604.19710 (2026)

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  44. [44]

    Cheng, X

    M. Cheng, X. Xie, R. Wang, Y. Zhou, M. Hu, Adreft: Adaptive decision repair for safe autonomous driving via reinforcement fine-tuning, arXiv preprint arXiv:2506.23960 (2025)

    work page arXiv 2025

  45. [45]

    Dauner, M

    D. Dauner, M. Hallgarten, T. Li, X. Weng, Z. Huang, Z. Yang, H. Li, I. Gilitschenski, B. Ivanovic, M. Pavone, et al., Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking, Advances in Neural Information Processing Systems 37 (2024) 28706–28719

    work page 2024

  46. [46]

    Bengio, J

    Y. Bengio, J. Louradour, R. Collobert, J. Weston, Curriculum learning, in: Proceedings of the 26th annual international conference on machine learning, 2009, pp. 41–48

    work page 2009

  47. [47]

    S. Peng, K. Genova, C. Jiang, A. Tagliasacchi, M. Pollefeys, T. Funkhouser, et al., Openscene: 3d scene understanding with open vocabularies, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2023, pp. 815– 824

    work page 2023

  48. [48]

    T. Wang, S. Kim, J. Wenxuan, E. Xie, C. Ge, J. Chen, Z. Li, P. Luo, Deepaccident: A motion and accident prediction benchmark for v2x autonomous driving, in: Proceedings of the AAAI Conference on Artificial Intelligence, Vol. 38, 2024, pp. 5599–5606. 21

    work page 2024

  49. [49]

    National Transportation Safety Board, The use of forward collision avoidance systems to prevent and mitigate rear-end crashes, Special Investigation Report NTSB/SIR-15/01, National Transportation Safety Board, Washington, DC, pB2015-104098 (May 2015)

    work page 2015

  50. [50]

    Prakash, K

    A. Prakash, K. Chitta, A. Geiger, Multi-modal fusion transformer for end-to-end au- tonomous driving, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021, pp. 7077–7087

    work page 2021

  51. [51]

    X. Weng, B. Ivanovic, Y. Wang, Y. Wang, M. Pavone, Para-drive: Parallelized architec- ture for real-time autonomous driving, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 15449–15458

    work page 2024

  52. [52]

    R. Feng, N. Xi, D. Chu, R. Wang, Z. Deng, A. Wang, L. Lu, J. Wang, Y. Huang, Artemis: Autoregressive end-to-end trajectory planning with mixture of experts for autonomous driving, IEEE Robotics and Automation Letters 11 (1) (2025) 226–233

    work page 2025

  53. [53]

    B. Liao, S. Chen, H. Yin, B. Jiang, C. Wang, S. Yan, X. Zhang, X. Li, Y. Zhang, Q. Zhang, et al., Diffusiondrive: Truncated diffusion model for end-to-end autonomous driving, in: Proceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 12037–12047

    work page 2025

  54. [54]

    Y. Li, Y. Wang, Y. Liu, J. He, L. Fan, Z. Zhang, End-to-end driving with online trajec- tory evaluation via bev world model, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 27137–27146

    work page 2025

  55. [55]

    J. Zhu, W. Wang, Z. Chen, Z. Liu, S. Ye, L. Gu, Y. Duan, H. Tian, W. Su, J. Shao, et al., InternVL3: Exploring advanced training and test-time recipes for open-source multimodal models, arXiv preprint arXiv:2504.10479 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  56. [56]

    S. Bai, Y. Cai, R. Chen, K. Chen, X. Chen, Z. Cheng, L. Deng, W. Ding, C. Gao, C. Ge, et al., Qwen3-vl technical report, arXiv preprint arXiv:2511.21631 (2025)

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  57. [57]

    Y. Chen, Y. Wang, Z. Zhang, Drivinggpt: Unifying driving world modeling and plan- ning with multi-modal autoregressive transformers, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 26890–26900

    work page 2025

  58. [58]

    B. Xiao, C. Feng, Z. Huang, F. Yan, Y. Zhong, L. Ma, Robotron-sim: improving real- world driving via simulated hard-case, in: Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 27380–27389

    work page 2025

  59. [59]

    Z. Zhou, T. Cai, S. Zhao, Y. Zhang, Z. Huang, B. Zhou, J. Ma, Autovla: A vision- language-action model for end-to-end autonomous driving with adaptive reasoning and reinforcementfine-tuning, AdvancesinNeuralInformationProcessingSystems38(2026) 27920–27956. 22 Appendix A. Implementation Details Appendix A.1. Network Architecture Our model architecture cons...

    work page 2026