pith. sign in

arxiv: 2605.18895 · v1 · pith:LZSL35F4new · submitted 2026-05-17 · 💻 cs.RO · cs.AI

KG-ASG: Collision-Knowledge-Guided Closed-Loop Adversarial Scenario Generation With Primary-Support Attribution

Cheng Wang , Chen Xiong , Ziwen Wang , Yuchen Zhou , Qiang Liu This is my paper

Pith reviewed 2026-05-20 12:53 UTC · model grok-4.3

classification 💻 cs.RO cs.AI
keywords adversarial scenario generationautonomous drivingcollision knowledgeprimary-support attributionsafety validationclosed-loop testingmulti-vehicle interactions
0
0 comments X

The pith

KG-ASG generates adversarial driving scenarios by using collision knowledge to select one primary adversary and support vehicles for clearer, more executable tests of autonomous systems.

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

The paper introduces a framework that builds a structured knowledge base of collision types and trains a lightweight expert to identify a main colliding vehicle along with supporting vehicles that add risk without causing extra collisions. This semantic guidance turns multi-vehicle scenario creation into a primary-support process, with added hard constraints on physics, rules, and single-collider outcomes plus feedback from the ego vehicle's controller for refinement. The result targets scenarios that are more interpretable and controllable than those from low-level trajectory tweaks or single-adversary searches. A sympathetic reader would care because autonomous driving safety validation needs tests that expose specific weaknesses without producing ambiguous or unrealistic multi-vehicle pileups.

Core claim

KG-ASG constructs a structured collision knowledge base and trains a lightweight Collision Expert to infer the target collision mode, the unique primary adversary, support vehicles, and their interaction roles. Guided by this semantic prior, multi-vehicle adversarial generation is formulated as a primary-support process, where the primary adversary induces the main conflict and support vehicles shape the surrounding risk structure without becoming additional colliders. Rule, physical, interaction-safety, and single-collider constraints are imposed as hard gates to filter non-executable samples. To handle reactive ego behaviors, planner-controller feedback is further used for failure diagno

What carries the argument

The primary-support attribution process, in which a Collision Expert draws on a structured collision knowledge base to designate one primary adversary that causes the main conflict and support vehicles that shape risk without colliding.

If this is right

  • KG-ASG achieves strong adversarial effectiveness while improving Valid Primary Attack.
  • It reduces multi-collision rates in generated scenarios.
  • Closed-loop recovery gains appear under IDM, Cruise, and Expert controllers.
  • Collision-knowledge guidance and primary-support reasoning increase interpretability and executability for safety validation.

Where Pith is reading between the lines

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

  • The same attribution idea could help isolate failure modes during debugging of specific planner modules.
  • Growing the knowledge base with new collision patterns might extend coverage to additional urban driving edge cases.
  • Semantic priors of this kind may lower the search effort needed for effective adversarial generation compared with pure optimization.

Load-bearing premise

The Collision Expert trained on the structured collision knowledge base can accurately and uniquely infer the target collision mode, the primary adversary, support vehicles, and their interaction roles so that the process produces valid single-collider scenarios.

What would settle it

Run the method on WOMD scenarios in MetaDrive and check whether the output set shows markedly higher Valid Primary Attack rates and lower multi-collision counts than baselines; if the rates stay comparable or worse when the knowledge guidance is removed, the central claim does not hold.

Figures

Figures reproduced from arXiv: 2605.18895 by Cheng Wang, Chen Xiong, Qiang Liu, Yuchen Zhou, Ziwen Wang.

Figure 1
Figure 1. Figure 1: The first paradigm is data-driven safety-critical scenario [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 1
Figure 1. Figure 1: Comparison of scenario generation paradigms for autonomous driving [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: KG-ASG knowledge-guided closed-loop adversarial scenario generation framework. The framework uses high-level semantic priors to constrain low [PITH_FULL_IMAGE:figures/full_fig_p006_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Role-aware primary-support adversarial generation framework. The Collision Expert provides structured semantic guidance, including the target [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Statistics of the constructed collision knowledge base. (a) KG-ASG fit distribution for all knowledge entries and collision-related entries. (b) Distribution [PITH_FULL_IMAGE:figures/full_fig_p011_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Method-level qualitative comparison. KG-ASG preserves the original [PITH_FULL_IMAGE:figures/full_fig_p012_5.png] view at source ↗
Figure 7
Figure 7. Figure 7: Collision Expert versus base models. The trained Collision Expert [PITH_FULL_IMAGE:figures/full_fig_p013_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Qualitative progression from Stage 1 to KG-ASG Full in failure [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Multi-modal KG-ASG cases with primary-support roles. KG-ASG generates diverse high-risk interaction structures, where the primary adversary is [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Scene-type stratification of KG-ASG generated scenarios. The generated scenarios are grouped according to traffic structure, lane relation, collision [PITH_FULL_IMAGE:figures/full_fig_p015_10.png] view at source ↗
read the original abstract

Safety validation of autonomous driving systems requires high-risk scenario coverage, clear collision semantics, executable trajectories, and attributable multi-vehicle interactions. Existing safety-critical scenario generation methods often rely on low-level trajectory perturbations, collision-proxy optimization, or single-adversary search, which may produce adversarial samples with ambiguous collision causes or uncontrolled multi-vehicle collisions. This paper proposes KG-ASG, a collision-knowledge-guided closed-loop adversarial scenario generation framework with primary-support attribution. KG-ASG constructs a structured collision knowledge base and trains a lightweight Collision Expert to infer the target collision mode, the unique primary adversary, support vehicles, and their interaction roles. Guided by this semantic prior, multi-vehicle adversarial generation is formulated as a primary-support process, where the primary adversary induces the main conflict and support vehicles shape the surrounding risk structure without becoming additional colliders. Rule, physical, interaction-safety, and single-collider constraints are imposed as hard gates to filter non-executable samples. To handle reactive ego behaviors, planner-controller feedback is further used for failure diagnosis, candidate re-ranking, and terminal refinement. Experiments on WOMD scenarios reconstructed in MetaDrive show that KG-ASG achieves strong adversarial effectiveness while improving Valid Primary Attack, reducing multi-collision, and obtaining closed-loop recovery gains under IDM, Cruise, and Expert controllers. These results demonstrate that collision-knowledge guidance and primary-support single-collider reasoning improve adversarial effectiveness, interpretability, and executability for autonomous driving safety validation.

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

2 major / 1 minor

Summary. The paper introduces KG-ASG, a collision-knowledge-guided closed-loop adversarial scenario generation framework with primary-support attribution for autonomous driving safety validation. It constructs a structured collision knowledge base, trains a lightweight Collision Expert to infer target collision mode, unique primary adversary, support vehicles and interaction roles, then formulates multi-vehicle generation as a primary-support process subject to rule, physical, interaction-safety and single-collider hard constraints, with planner-controller feedback for diagnosis and refinement. Experiments on WOMD scenarios reconstructed in MetaDrive are reported to achieve strong adversarial effectiveness while improving Valid Primary Attack, reducing multi-collision, and obtaining closed-loop recovery gains under IDM, Cruise, and Expert controllers.

Significance. If substantiated, the approach could advance autonomous driving safety validation by supplying semantically attributable and executable adversarial scenarios that reduce ambiguous collision causes and uncontrolled multi-vehicle interactions, offering clearer interpretability than low-level perturbation or single-adversary search methods.

major comments (2)
  1. [Abstract] Abstract: the abstract reports positive outcomes on WOMD scenarios in MetaDrive but provides no quantitative metrics, error analysis, or detailed experimental controls, leaving the support for central claims difficult to verify.
  2. [Method] Collision Expert (method section): the claim that collision-knowledge guidance and primary-support reasoning improve Valid Primary Attack and reduce multi-collision rests on the unvalidated assumption that the Collision Expert accurately and uniquely infers collision mode, primary adversary, support vehicles, and roles; no precision, recall, or error-rate metrics on inference are supplied, which is load-bearing for attributing reported gains to the proposed guidance rather than filtering artifacts.
minor comments (1)
  1. [Method] Notation for primary-support attribution and hard gates could be accompanied by an explicit pseudocode listing or diagram to improve clarity of the filtering process.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which help strengthen the clarity and rigor of our work. We address each major comment point by point below and have revised the manuscript accordingly.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the abstract reports positive outcomes on WOMD scenarios in MetaDrive but provides no quantitative metrics, error analysis, or detailed experimental controls, leaving the support for central claims difficult to verify.

    Authors: We agree that the abstract would benefit from quantitative metrics and experimental details to better support the claims. In the revised manuscript, we have updated the abstract to include specific results such as the measured improvements in Valid Primary Attack rate, reductions in multi-collision occurrences, and closed-loop recovery gains, along with explicit references to the WOMD scenarios, MetaDrive simulator, and the three controller types (IDM, Cruise, Expert) used in evaluation. revision: yes

  2. Referee: [Method] Collision Expert (method section): the claim that collision-knowledge guidance and primary-support reasoning improve Valid Primary Attack and reduce multi-collision rests on the unvalidated assumption that the Collision Expert accurately and uniquely infers collision mode, primary adversary, support vehicles, and roles; no precision, recall, or error-rate metrics on inference are supplied, which is load-bearing for attributing reported gains to the proposed guidance rather than filtering artifacts.

    Authors: This is a valid observation. The original manuscript did not report direct accuracy metrics for the Collision Expert. To substantiate that the observed gains in Valid Primary Attack and multi-collision reduction stem from the semantic guidance rather than filtering effects, we have added a dedicated evaluation subsection in the experiments. This reports precision, recall, and overall accuracy of the Collision Expert on held-out collision annotations for mode inference, primary-adversary identification, support-vehicle selection, and role assignment. revision: yes

Circularity Check

0 steps flagged

No significant circularity detected in derivation chain

full rationale

The KG-ASG framework constructs an external structured collision knowledge base, trains a separate Collision Expert model to infer modes/roles, then applies hard rule/physical/interaction/single-collider gates plus planner-controller feedback loops to generate and refine scenarios. Experimental outcomes on WOMD reconstructions in MetaDrive are reported as direct empirical measurements under IDM/Cruise/Expert controllers. No step reduces a claimed prediction or uniqueness result to a fitted parameter or self-citation that is itself defined by the target claim; the derivation remains self-contained against external data and benchmarks.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

Review is based solely on the abstract; specific free parameters, axioms, and entities are inferred at high level from the described components. The central claim depends on the reliability of the constructed knowledge base and expert inference.

free parameters (1)
  • Collision Expert model parameters
    Lightweight Collision Expert is trained to infer collision modes and roles, implying learned parameters from the knowledge base.
axioms (1)
  • domain assumption Structured collision knowledge base accurately captures collision modes, primary adversaries, and interaction roles.
    Invoked when constructing the base and using it to guide multi-vehicle generation and filtering.
invented entities (1)
  • Collision Expert no independent evidence
    purpose: Infer target collision mode, unique primary adversary, support vehicles, and interaction roles.
    New trained model introduced to provide semantic prior for the primary-support generation process.

pith-pipeline@v0.9.0 · 5804 in / 1515 out tokens · 56104 ms · 2026-05-20T12:53:29.610580+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

44 extracted references · 44 canonical work pages

  1. [1]

    Curse of rarity for autonomous vehicles,

    H. X. Liu and S. Feng, “Curse of rarity for autonomous vehicles,”Nature Communications, vol. 15, p. 4808, 2024

    work page 2024

  2. [2]

    A matched case-control analysis of au- tonomous vs human-driven vehicle accidents,

    M. Abdel-Aty and S. Ding, “A matched case-control analysis of au- tonomous vs human-driven vehicle accidents,”Nature Communications, vol. 15, no. 1, p. 4931, 2024

    work page 2024

  3. [3]

    How to guarantee driving safety for autonomous vehicles in a real- world environment: A perspective on self-evolution mechanisms,

    S. Yang, Y . Huang, L. Li, S. Feng, X. Na, H. Chen, and A. Khajepour, “How to guarantee driving safety for autonomous vehicles in a real- world environment: A perspective on self-evolution mechanisms,”IEEE Intelligent Transportation Systems Magazine, vol. 16, no. 2, pp. 41–54, 2024

    work page 2024

  4. [4]

    How would autonomous vehicles behave in real-world crash scenarios?

    R. Zhou, G. Zhang, H. Huang, H. Huang, Z. Wei, H. Zhou, J. Jin, J. Jin, F. Chang, and J. Chen, “How would autonomous vehicles behave in real-world crash scenarios?”Accident Analysis & Prevention, vol. 202, p. 107572, 2024

    work page 2024

  5. [6]

    Interactive critical scenario generation for autonomous vehicles testing based on in-depth crash data using reinforcement learning,

    Z. Wei, H. Huang, G. Zhang, R. Zhouet al., “Interactive critical scenario generation for autonomous vehicles testing based on in-depth crash data using reinforcement learning,”IEEE Transactions on Intelligent Vehicles, vol. 10, no. 3, pp. 1471–1482, 2025

    work page 2025

  6. [7]

    Safety-critical scenario generation via reinforcement learning based editing,

    H. Liu, L. Zhang, S. K. S. Hari, and J. Zhao, “Safety-critical scenario generation via reinforcement learning based editing,” inProceedings of the 2024 IEEE International Conference on Robotics and Automation. IEEE, 2024, pp. 14 405–14 412

    work page 2024

  7. [8]

    Evaluating autonomous vehicle safety performance through analysis of pre-crash trajectories of powered two-wheelers,

    R. Zhou, Z. Lin, G. Zhang, H. Huang, H. Zhou, and J. Chen, “Evaluating autonomous vehicle safety performance through analysis of pre-crash trajectories of powered two-wheelers,”IEEE Transactions on Intelligent Transportation Systems, vol. 25, no. 10, pp. 13 560–13 572, 2024

    work page 2024

  8. [9]

    Adap- tive safety performance testing for autonomous vehicles with adaptive importance sampling,

    J. Yang, Z. Wang, D. Wang, Y . Zhang, Q. Lu, and S. Feng, “Adap- tive safety performance testing for autonomous vehicles with adaptive importance sampling,”Transportation Research Part C: Emerging Tech- nologies, vol. 179, p. 105256, 2025

    work page 2025

  9. [10]

    Critical test sce- nario generation for autonomous vehicles using reinforcement learning,

    S. Zhang, X. Sun, G. Li, Y . Pan, and T. Tak, “Critical test sce- nario generation for autonomous vehicles using reinforcement learning,” Transportmetrica A: Transport Science, 2025

    work page 2025

  10. [11]

    Hspg: An open-loop testing framework for autonomous driving based on proactive generation of hazardous scenario,

    C. Wang, Q. Liu, W. Fang, and C. Xiong, “Hspg: An open-loop testing framework for autonomous driving based on proactive generation of hazardous scenario,”Accident Analysis & Prevention, vol. 229, p. 108449, 2026

    work page 2026

  11. [12]

    Advdiffuser: Generating adversarial safety-critical driving scenarios via guided diffusion,

    Y . Xie, X. Guo, C. Wang, K. Liu, and L. Chen, “Advdiffuser: Generating adversarial safety-critical driving scenarios via guided diffusion,”arXiv preprint arXiv:2410.08453, 2024

    work page arXiv 2024

  12. [13]

    Diffscene: Diffusion- based safety-critical scenario generation for autonomous vehicles,

    C. Xu, A. Petiushko, D. Zhao, and B. Li, “Diffscene: Diffusion- based safety-critical scenario generation for autonomous vehicles,” in Proceedings of the AAAI Conference on Artificial Intelligence, vol. 39, no. 8, 2025, pp. 8797–8805

    work page 2025

  13. [14]

    World model-based end-to-end scene generation for accident anticipation in autonomous driving,

    Y . Guan, H. Liao, C. Wang, X. Liu, J. Zhang, and Z. Li, “World model-based end-to-end scene generation for accident anticipation in autonomous driving,”Communications Engineering, vol. 4, p. 144, 2025

    work page 2025

  14. [15]

    Chatscene: Knowledge-enabled safety- critical scenario generation for autonomous vehicles,

    J. Zhang, C. Xu, and B. Li, “Chatscene: Knowledge-enabled safety- critical scenario generation for autonomous vehicles,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 15 459–15 469

    work page 2024

  15. [16]

    Llm-attacker: Enhancing closed- loop adversarial scenario generation for autonomous driving with large language models,

    Y . Mei, T. Nie, J. Sun, and Y . Tian, “Llm-attacker: Enhancing closed- loop adversarial scenario generation for autonomous driving with large language models,”IEEE Transactions on Intelligent Transportation Systems, vol. 26, no. 10, pp. 15 068–15 076, 2025

    work page 2025

  16. [17]

    Traj-llm: A new exploration for empowering trajectory prediction with pre-trained large language models,

    Z. Lan, H. Li, L. Liu, B. Fan, Y . Lv, Y . Ren, and Z. Cui, “Traj-llm: A new exploration for empowering trajectory prediction with pre-trained large language models,”IEEE Transactions on Intelligent Vehicles, vol. 10, no. 2, pp. 794–807, 2025

    work page 2025

  17. [18]

    Asynchronous large language model enhanced planner for autonomous driving,

    Y . Chen, Z.-h. Ding, Z. Wang, Y . Wang, L. Zhang, and S. Liu, “Asynchronous large language model enhanced planner for autonomous driving,” inComputer Vision – ECCV 2024. Springer, 2024, pp. 22–38

    work page 2024

  18. [19]

    Traffic-it: Enhancing traffic scene understanding for multimodal large language models,

    S. Kuang, Y . Liu, X. Qu, and Y . Wei, “Traffic-it: Enhancing traffic scene understanding for multimodal large language models,”Transportation Research Part C: Emerging Technologies, vol. 180, p. 105325, 2025

    work page 2025

  19. [20]

    Uiram: An intention-uncertainty-based risk assessment framework for interactive traffic scenarios,

    C. Wang, C. Xiong, M. Hu, Y . Liu, C. Ma, J. Zhu, and Q. Liu, “Uiram: An intention-uncertainty-based risk assessment framework for interactive traffic scenarios,”Accident Analysis & Prevention, vol. 233, p. 108580, 2026

    work page 2026

  20. [21]

    Pre-crash scenario typology for crash avoidance research,

    W. G. Najm, J. D. Smith, and M. Yanagisawa, “Pre-crash scenario typology for crash avoidance research,” National Highway Traffic Safety Administration, U.S. Department of Transportation, Washington, DC, USA, Tech. Rep. DOT HS 810 767, Apr. 2007

    work page 2007

  21. [22]

    Generalization of cut-in pre-crash scenarios for autonomous vehicles based on accident data,

    P. Li, X. Zhu, Y . Renet al., “Generalization of cut-in pre-crash scenarios for autonomous vehicles based on accident data,”Scientific Reports, vol. 14, p. 17664, 2024

    work page 2024

  22. [23]

    Exploring master scenarios for autonomous driving tests from police-reported historical crash data using an adaptive search sampling framework,

    Y . Li, Z. Yang, J. Jin, and D. Wu, “Exploring master scenarios for autonomous driving tests from police-reported historical crash data using an adaptive search sampling framework,”Accident Analysis & Prevention, vol. 205, p. 107688, 2024

    work page 2024

  23. [24]

    Research on vehicle accident hazard scenario derivation based on improved ast,

    H. Zhou, L. Xu, Y . Renet al., “Research on vehicle accident hazard scenario derivation based on improved ast,”Scientific Reports, vol. 15, p. 26350, 2025

    work page 2025

  24. [25]

    High-risk test scenario generation for autonomous vehicles at roundabouts using naturalistic driving data,

    D. Ren, H. Huang, Y . Li, and J. Jin, “High-risk test scenario generation for autonomous vehicles at roundabouts using naturalistic driving data,” Applied Sciences, vol. 15, no. 8, p. 4505, 2025

    work page 2025

  25. [26]

    Cascaded safety analysis and test scenario generation techniques for autonomous driving: A case study with watonobus,

    C. Sun, R. Zhang, A. R. Alghoonehet al., “Cascaded safety analysis and test scenario generation techniques for autonomous driving: A case study with watonobus,”Automotive Innovation, vol. 8, pp. 252–263, 2025

    work page 2025

  26. [27]

    Specific scenario generation method for trustworthiness testing of autonomous vehicles based on interaction coding,

    Y . Chang, C. Xi, and Z. Luo, “Specific scenario generation method for trustworthiness testing of autonomous vehicles based on interaction coding,”Applied Sciences, vol. 15, no. 19, p. 10656, 2025

    work page 2025

  27. [28]

    Towards full-scenario safety evaluation of automated vehicles: A volume-based method,

    H. Zhou, C. Ma, S. Shen, Z. Liang, and X. Li, “Towards full-scenario safety evaluation of automated vehicles: A volume-based method,” Transportation Research Part C: Emerging Technologies, vol. 183, p. 105485, 2026

    work page 2026

  28. [29]

    Emergency lane-change simulation: A behavior-guided approach for safety-critical scenario generation,

    C. Xiong, C. Wang, Y . Liu, Z. Wu, and Y . Tian, “Emergency lane-change simulation: A behavior-guided approach for safety-critical scenario generation,” 2026

    work page 2026

  29. [30]

    Mjtg: A multi- vehicle joint trajectory generator for complex and rare scenarios,

    Y . Tian, W. Zheng, Y . Shao, H. Zhang, and J. Sun, “Mjtg: A multi- vehicle joint trajectory generator for complex and rare scenarios,”IEEE Transactions on Vehicular Technology, vol. 74, no. 10, pp. 15 026– 15 039, 2025

    work page 2025

  30. [31]

    Cat: Closed-loop adversarial training for safe end-to-end driving,

    L. Zhang, Z. Peng, Q. Li, and B. Zhou, “Cat: Closed-loop adversarial training for safe end-to-end driving,” inProceedings of the Conference on Robot Learning. PMLR, 2023, pp. 2357–2372

    work page 2023

  31. [32]

    King: Generating safety-critical driving scenarios for robust imitation via kinematics gradients,

    N. Hanselmann, K. Renz, K. Chitta, A. Bhattacharyya, and A. Geiger, “King: Generating safety-critical driving scenarios for robust imitation via kinematics gradients,” inComputer Vision – ECCV 2022, ser. Lecture Notes in Computer Science, vol. 13698. Springer, 2022, pp. 335–352

    work page 2022

  32. [33]

    On adversarial robustness of trajectory prediction for autonomous vehicles,

    Q. Zhang, S. Hu, J. Sun, Q. A. Chen, and Z. M. Mao, “On adversarial robustness of trajectory prediction for autonomous vehicles,” inPro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022, pp. 15 159–15 168

    work page 2022

  33. [34]

    Seal: Towards safe autonomous driving via skill-enabled adversary learning for closed-loop scenario generation,

    B. Stoler, I. Navarro, J. Francis, and J. Oh, “Seal: Towards safe autonomous driving via skill-enabled adversary learning for closed-loop scenario generation,”IEEE Robotics and Automation Letters, vol. 10, no. 9, pp. 9305–9312, 2025

    work page 2025

  34. [35]

    Goose: Goal-conditioned reinforcement learning for safety-critical scenario generation,

    J. Ransiek, J. Plaum, J. Langner, and E. Sax, “Goose: Goal-conditioned reinforcement learning for safety-critical scenario generation,” inPro- ceedings of the IEEE 27th International Conference on Intelligent Transportation Systems. IEEE, 2024, pp. 2651–2658

    work page 2024

  35. [36]

    Steerable adversarial scenario generation through test-time preference ARXIV PREPRINT 17 alignment,

    T. Nie, Y . Mei, Y . Tang, J. He, J. Sun, H. Shi, W. Ma, and J. Sun, “Steerable adversarial scenario generation through test-time preference ARXIV PREPRINT 17 alignment,” inInternational Conference on Learning Representations, 2026

    work page 2026

  36. [37]

    Large language models powered context-aware motion prediction in autonomous driving,

    X. Zheng, L. Wu, Z. Yanet al., “Large language models powered context-aware motion prediction in autonomous driving,”arXiv preprint arXiv:2403.11057, 2024

    work page arXiv 2024

  37. [38]

    Automating concrete simulation scenario generation for autonomous driving with large language models,

    J. Li and R. Wang, “Automating concrete simulation scenario generation for autonomous driving with large language models,” inSAE 2025 Intelligent and Connected Vehicles Symposium. SAE International, 2025

    work page 2025

  38. [39]

    Automated generation of test scenarios for autonomous driving using llms,

    A. A. Danso and U. B ¨uker, “Automated generation of test scenarios for autonomous driving using llms,”Electronics, vol. 14, no. 16, p. 3177, 2025

    work page 2025

  39. [40]

    Trajectory-llm: A language-based data generator for trajectory prediction in autonomous driving,

    K. Yang, Z. Guo, G. Linet al., “Trajectory-llm: A language-based data generator for trajectory prediction in autonomous driving,” in International Conference on Learning Representations, 2025

    work page 2025

  40. [41]

    A dynamic prompting and scenario generation method for autonomous driving perception via large-model optimization,

    S. Zhang, H. Lin, M. Wang, B. Wei, Y . Liu, and X. Qu, “A dynamic prompting and scenario generation method for autonomous driving perception via large-model optimization,”Transportation Research Part C: Emerging Technologies, vol. 188, p. 105672, 2026

    work page 2026

  41. [42]

    Curricuvlm: Towards safe autonomous driving via personalized safety- critical curriculum learning with vision-language models,

    Z. Sheng, Z. Huang, Y . Qu, Y . Leng, S. Bhavanam, and S. Chen, “Curricuvlm: Towards safe autonomous driving via personalized safety- critical curriculum learning with vision-language models,”Transporta- tion Research Part C: Emerging Technologies, vol. 185, p. 105549, 2026

    work page 2026

  42. [43]

    Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset,

    S. Ettinger, S. Cheng, B. Caine, C. Liu, H. Zhao, S. Pradhan, Y . Chai, B. Sapp, C. R. Qi, Y . Zhou, Z. Yang, A. Chouard, P. Sun, J. Ngiam, V . Vasudevan, A. McCauley, J. Shlens, and D. Anguelov, “Large scale interactive motion forecasting for autonomous driving: The waymo open motion dataset,” inProceedings of the IEEE/CVF International Conference on Com...

    work page 2021

  43. [44]

    Metadrive: Composing diverse driving scenarios for generalizable reinforcement learning,

    Q. Li, Z. Peng, L. Feng, Q. Zhang, Z. Xue, and B. Zhou, “Metadrive: Composing diverse driving scenarios for generalizable reinforcement learning,”IEEE Transactions on Pattern Analysis and Machine Intel- ligence, 2022

    work page 2022

  44. [45]

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

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

    work page 2021