Pith. sign in

REVIEW 3 major objections 4 minor 5 cited by

From Words to Collisions: LLM-Guided Evaluation and Adversarial Generation of Safety-Critical Driving Scenarios

T0 review · 3 major / 4 minor · reviewed 2026-08-09 · deepseek-v4-flash

Pith's one-line read The paper claims that an LLM prompted with ego-centric driving descriptions can evaluate collision and safety-critical scenarios and adversarially edit an agent's trajectory to turn a safe scene into a collision.

desk verdict The evaluation half is solid and the ego-prompt finding is the real contribution; the generation half leans on a self-scored LLM prompt and a single case study, so that claim is not yet supported. read the letter →

arxiv 2502.02145 v4 pith:5FZ5HJWT submitted 2025-02-04 cs.AI cs.CLcs.RO

classification cs.AIcs.CLcs.RO
keywords AutonomousDrivingLargeLanguageModelsScenario-basedTestingSafety-criticalScenarioEvaluationGenerationPromptEngineeringEgo-centricRepresentationAdversarial
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 is trying to establish that a general-purpose large language model, prompted with structured driving data described from the ego vehicle's viewpoint, can replace parts of the handcrafted metric pipeline in scenario-based testing. Its evaluation module turns 2D simulation scenarios into natural language and decides whether a collision occurs or whether the scene will become safety-critical; the ego-centric prompt reaches an F1 of 0.87 for collision evaluation and 0.83 for safety inference with one model, and similar scores with two other models. Its generation module identifies the most risk-inducing agent, rewrites that agent's trajectory, and reports a much lower risk score, turning a previously safe scenario into a collision in the shown case study. If this holds, test engineers could scale virtual safety testing by asking an LLM to flag and create critical cases instead of pre-defining every metric and reading the output by hand.

What carries the argument

The load-bearing mechanism is the ego-centric parser and prompt template, Promptego, contrasted with the Cartesian Promptcart. The parser re-expresses every agent's state in longitudinal and lateral coordinates relative to the ego vehicle, adding motion phrases such as 'approaching the EV longitudinally from the front', so the LLM receives a driver's-perspective description rather than absolute coordinates. The prompt stacks Contextual Prompting, which defines safety metrics such as time-to-collision and minimum distance to collision, with Chain-of-Thought reasoning and In-Context Learning examples. The generation module adds Self-Consistency, sampling several adversarial trajectories and selecting the one with the lowest LLM-assigned risk score.

What would settle it

Re-run every generated trajectory through the paper's 2D simulator with its motion planner and compare outcomes against the original safe runs; if the LLM-selected trajectories do not cause a significantly higher rate of collisions, near-misses, or metric values below critical thresholds, the generation claim fails. In parallel, run the collision evaluator on a fresh held-out set of simulated scenarios and check its verdicts against simulator ground truth.

Watch

Extended reading notes

Core claim

The authors claim that LLMs can act as both evaluator and generator of safety-critical driving scenarios. They compare two prompt families, Cartesian and ego-centric, and find the ego-centric representation is consistently superior: for Gemini-1.5Pro, F1 rises from 0.21 to 0.87 in collision evaluation and from 0.62 to 0.83 in safety inference, with DeepSeek-V3 reaching 0.90 and 0.87 on the same tasks. The generator uses a safety-metric prompt to score each obstacle, identifies an ego-attacker, creates multiple candidate trajectories through self-consistency, and keeps the candidate with the lowest risk score, producing a case where the ego-attacker's risk score falls over time and ends in collision. The paper concludes that this reduces dependence on handcrafted metrics while remaining controllable.

Load-bearing premise

The generation module's reported success is measured by the same LLM risk-score prompt that selected the trajectory, and the only validation is a single visualized case study, so the claim of realistic safety-critical generation assumes that self-scored risk drops correspond to genuinely critical driving situations.

Editorial extensions

If this is right

  • Scenario evaluators can flag collisions and predict imminent risk directly from structured log data, without a human re-reading metric traces.
  • The ego-centric framing, not the particular model, drives most of the gain, since all three tested LLMs perform far better with Promptego than with Promptcart.
  • Adding Contextual Prompting, Chain-of-Thought, and In-Context Learning together improves the ego-prompt collision F1 from 0.70 to 0.87.
  • The generation module offers a controllable way to produce targeted critical scenarios by editing one risk-inducing agent, complementing data-driven and knowledge-based generators.
  • The safety inference output can serve as an online supplementary risk signal for motion planners, although the paper notes the current response time is seconds rather than real time.

Reading between the lines

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

  • Beyond the paper, the same ego-centric relabeling could extend to vision-language models operating on camera views or rendered scenes, since the gain seems to come from spatial framing rather than from text generation itself.
  • Beyond the paper, the generation module should be validated by re-simulating the generated trajectories and checking collisions or metric thresholds; the current evidence is one case study whose outcome is scored by the same prompt that selected it.
  • Beyond the paper, a natural next test is to feed generated scenarios to a motion planner and measure how frequently it fails, converting scenario quality into a downstream safety statistic.
  • Beyond the paper, the method reduces rather than eliminates handcrafted metrics, because metric thresholds still appear as prompt context and as the ground truth used to label the evaluation data.
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

3 major / 4 minor

Summary. This paper proposes an LLM-based framework with two modules: an evaluation module that converts structured CommonRoad scenario data into natural-language descriptions via Cartesian or ego-centric parsers and prompts (with contextual prompting, chain-of-thought, and in-context learning), and a generation module that identifies an ego-attacker and adversarially modifies its trajectory to create safety-critical scenarios. The evaluation is tested on 200 randomly selected scenarios from a 6000-scenario simulation dataset; ego-centric prompts reach F1 0.87 for collision evaluation and 0.83 for safety inference with Gemini-1.5Pro, with ablations and comparisons across three LLMs. The generation module is validated by one case study and a plot of the selected trajectory's risk score over time. The paper argues that LLMs with domain-informed prompting can reduce dependence on handcrafted safety metrics.

Significance. If the evaluation results hold up, the ego-centric prompt design is a useful contribution: the ablations show clear gains from CP, CoT, and ICL, the comparison across three LLMs is informative, and the open-source release supports reproducibility. The collision-evaluation claim is not circular because the labels are external simulator collisions. The generation direction is novel—using an LLM to select an ego-attacker and synthesize adversarial trajectories—but the evidence for it is currently much weaker than for evaluation. The absence of baselines and uncertainty quantification for the evaluation, and the self-scored generation validation, mean the central two-part claim is only partially established.

major comments (3)
  1. [§III-C.2 and §IV-D] The central claim that the generation module 'synthesizes realistic, safety-critical scenarios' is supported only by the selected candidate's LLM-assigned risk score decreasing over time (Fig. 8) and by a single visualized case study (Fig. 9, BEL Antwerp-1_14_T-1). Because the same safety-critical metrics prompt is used both to score candidate trajectories under Self-Consistency and to produce the reported post-generation risk scores, the selection criterion and the evaluation metric coincide; a low risk score on a modified trajectory may reflect prompt bias rather than genuine safety-criticality. The authors should re-simulate the generated trajectories with the Frenetix planner for a distribution of generated scenarios, recompute TTC/MDC and collision outcomes from the modified geometry, and compare against a random trajectory-perturbation or rule-based attacker baseline.
  2. [§IV-B and §IV-C, Tables I–IV] The evaluation use cases report F1 scores without any comparison to the TTC/MDC threshold classifiers from which the ground-truth labels are derived (Section IV-A), and the prompts explicitly embed the same kinds of thresholds (Section III-B.2). A simple threshold-based baseline is therefore the natural lower bound for these tasks, and its omission leaves unclear how much the LLM adds over handcrafted metrics. Please add such baselines for both collision evaluation and safety inference, and report the threshold values used.
  3. [§IV-A and §III-C.2] The abstract's claim of 'reducing dependence on handcrafted metrics' is not reflected in the design details: TTC/MDC thresholds are handcrafted in the prompts, the same thresholds generate the ground truth for safety inference, and the generation module's ego-attacker selection depends on unspecified risk-score thresholds and a spatial search radius. Since these free parameters directly affect the reported F1 scores and the generation outcomes, they should be stated explicitly, and a sensitivity analysis should be provided for at least the main thresholds.
minor comments (4)
  1. [Table II] For GPT-4o and DeepSeek-V3, the confusion-matrix entries sum to 204 rather than 200 for both prompt types; please correct or explain the discrepancy.
  2. [Fig. 8] The mapping from TTC/MDC values to the 0–5 risk score is not defined; please state the thresholds and describe how the plotted risk scores are computed.
  3. [§IV-D] The text contains a likely typo: 'within 30 min front' should presumably read 'within 30 m in front'; other minor spacing errors (e.g., 'outcome,i.e.') should also be corrected.
  4. [§IV-C] The safety-inference results are point estimates from a single 100/100 scenario split without confidence intervals; given LLM API stochasticity, reporting intervals over repeated splits or samples would make the cross-model differences in Tables II and IV interpretable.

Circularity Check

2 steps flagged · score 6.0 of 10

Generation module is self-scored: the same LLM risk prompt both selects the adversarial trajectory and serves as the reported success metric; evaluation module remains independently grounded.

  1. self definitional [Section III-C.2, Prompt Formulations]
    "To enhance the generation performance, we apply Self-Consistency (SC) by prompting the model to produce multiple adversarial trajectory candidates for the same ego-attacker. Each candidate is scored using the safety-critical metrics prompt, and the one with a low-risk score for the ego-attacker is selected."

    The same safety-critical metrics prompt is used both to assess obstacle risk and to score every candidate trajectory. Selecting the candidate with the lowest LLM-assigned risk score means the generator's output is, by construction, the trajectory that this prompt deems most critical. Any later claim that the generated scenario is safety-critical because its risk score is low is therefore restating the selection criterion, not an independent evaluation.

  2. self definitional [Section IV-D, Generation of Safety-Critical Scenarios and Fig. 8]
    "To assess the effectiveness of our framework, we conduct a case study using a scenario from the previous 4,700 safe scenarios from Sec IV-A. As shown in Fig. 8 and 9, the original scenario is safe, and the ego-attacker's trajectory has a safe risk level. Conversely, in the generated scenario, the ego-attacker's modified trajectory has a decreasing risk score over time (Fig. 8), ultimately leading to a collision."

    The reported evidence for generation success is the decreasing risk score of the selected trajectory, but that trajectory was chosen precisely because the same safety-critical metrics prompt assigned it the lowest risk score. Thus Fig. 8's quantitative result is forced by the selection procedure. The only external anchor is a single visualized collision in Fig. 9; there is no re-simulation with the Frenetix planner, no recomputation of TTC/MDC on the modified geometry, and no distribution over generated scenarios. The generation module's central quantitative claim therefore reduces to the LLM grading its own output.

full rationale

Collision evaluation and safety inference (Sections IV-B and IV-C) are not circular: the ground-truth collision labels come from the CommonRoad/Frenetix simulator, the TTC/MDC risk labels are computed externally, and the LLM predictions are compared against those labels, e.g., F1=0.87 with Promptego. The ablation studies and cross-LLM comparisons are self-contained. The circularity is confined to the generation module. In Section III-C.2 the same safety-critical metrics prompt is used to score candidate adversarial trajectories, and the candidate with the lowest LLM-assigned risk score is selected. Section IV-D then uses that same risk score's decrease (Fig. 8) as the evidence that a safety-critical scenario was generated. The quantitative success measure is identical to the selection objective, so the reported risk-drop is forced by construction. The single visualized collision in Fig. 9 provides one external check, but it does not validate the quantitative claim across scenarios, and no re-simulation or independent metric recomputation is reported. The self-cited survey [17] and co-authored simulator reference [31] are used for context and tooling, not as load-bearing evidence. Overall score 6: the evaluation pillar is independent, while the generation pillar's central quantitative claim partially reduces to the model scoring its own output.

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

The paper's central claims rest on hand-selected risk thresholds, a 30 m search radius, and a 10-timestep history, none of which are derived or ablated. The most important domain assumption is that LLM risk scores track true collision risk, which is used both to select and to evaluate generated scenarios.

free parameters (3)
  • TTC/MDC risk-score thresholds = not reported
    The prompt assigns risk scores based on threshold values for TTC and MDC (Sec III-B.2); exact thresholds are not given in the paper and are hand-chosen domain parameters.
  • Ego-attacker search radius = 30 m
    The generation module considers obstacles within 30 m in front or to the side of the EV (Sec IV-D); this hand-set range affects which agents can be selected as ego-attackers.
  • Safety-inference history window = 10 timesteps (1 s)
    Safety inference uses the current state plus a 10 timestep history (Sec IV-C); this is a design choice that defines the prediction task.
assumptions (4)
  • domain assumption TTC and MDC are valid indicators of safety criticality
    Used both to label ground truth and to instruct the LLM in prompts (Sec III-B.2, IV-A); if these metrics are poor safety proxies, the ground truth and evaluation are questionable.
  • domain assumption LLM-assigned risk scores correlate with collision likelihood
    The generation module selects the ego-attacker and the final trajectory based on LLM risk scores from the safety-critical metrics prompt (Sec III-C.2); this correlation is assumed, not measured against re-simulated outcomes.
  • domain assumption CommonRoad 2D scenarios with Frenetix planner represent realistic driving situations
    The framework is validated only in the CommonRoad 2D simulator with the Frenetix planner (Sec IV-A); generalization to real-world or 3D traffic is assumed.
  • domain assumption The selected 200 scenarios are representative of the 6000-scenario dataset
    Performance is reported on 200 randomly selected scenarios (Sec IV-B); randomness is claimed but no seed or distribution analysis is given.

how reviews work

0 comments
Cite this review

Pith. "Pith review of From Words to Collisions: LLM-Guided Evaluation and Adversarial Generation of Safety-Critical Driving Scenarios." pith.science (2026). https://pith.science/paper/5FZ5HJWT

@misc{pith2026250202145,
  author       = {Pith},
  title        = {Pith review of: From Words to Collisions: LLM-Guided Evaluation and Adversarial Generation of Safety-Critical Driving Scenarios},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/5FZ5HJWT}},
  note         = {Machine review of arXiv:2502.02145}
}
read the original abstract

Ensuring the safety of autonomous vehicles requires virtual scenario-based testing, which depends on the robust evaluation and generation of safety-critical scenarios. So far, researchers have used scenario-based testing frameworks that rely heavily on handcrafted scenarios as safety metrics. To reduce the effort of human interpretation and overcome the limited scalability of these approaches, we combine Large Language Models (LLMs) with structured scenario parsing and prompt engineering to automatically evaluate and generate safety-critical driving scenarios. We introduce Cartesian and Ego-centric prompt strategies for scenario evaluation, and an adversarial generation module that modifies trajectories of risk-inducing vehicles (ego-attackers) to create critical scenarios. We validate our approach using a 2D simulation framework and multiple pre-trained LLMs. The results show that the evaluation module effectively detects collision scenarios and infers scenario safety. Meanwhile, the new generation module identifies high-risk agents and synthesizes realistic, safety-critical scenarios. We conclude that an LLM equipped with domain-informed prompting techniques can effectively evaluate and generate safety-critical driving scenarios, reducing dependence on handcrafted metrics. We release our open-source code and scenarios at: https://github.com/TUM-AVS/From-Words-to-Collisions.

Figures

Figures reproduced from arXiv: 2502.02145 by the authors.

Figure 1
Figure 1. Our paper presents a new LLM-based framework [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Overview of the proposed framework based on LLM to evaluate and generate safety-critical scenarios. [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Example of an overtaking scenario in the 2D [PITH_FULL_IMAGE:figures/full_fig_p003_3.png] view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Output from the Cartesian-coordinate parser (a) and [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]
Figure 6
Figure 6. Figure 6: Confusion matrices for collision evaluation using [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]
Figure 7
Figure 7. Figure 7: Confusion matrices for safety evaluation, using [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]
Figure 9
Figure 9. Figure 9: A case study of safety-critical scenario generation on [PITH_FULL_IMAGE:figures/full_fig_p007_9.png]
Figure 8
Figure 8. Figure 8: Risk score of the ego-attacker vehicle’s trajectory, in [PITH_FULL_IMAGE:figures/full_fig_p007_8.png]

Discussion (0). Continue with ORCID to comment.

Forward citations

Cited by 5 Pith papers

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

  1. CrashAgent: Crash Scenario Generation via Multi-modal Reasoning

    cs.RO 2025-05 conditional novelty 6.0 of 10

    A multi-agent vision-language framework converts NHTSA crash reports into simulation-ready road layouts and collision scenarios, with modest accuracy gains over direct VLM baselines.

  2. A Survey on Evaluating Quality and Trustworthiness in LLM-Generated Data

    cs.AI 2026-01 conditional novelty 5.0 of 10

    A metric-oriented survey that classifies intrinsic quality and trustworthiness metrics for LLM-generated data across six modalities and documents systematic evaluation gaps in the current literature.

  3. Generative AI for Testing of Autonomous Driving Systems: A Survey

    cs.SE 2025-08 conditional novelty 5.0 of 10

    A systematic survey that organizes 91 studies of generative AI for autonomous driving testing into six scenario-based tasks and catalogs 27 limitations.

  4. AutoODD: Agentic Audits via Bayesian Red Teaming in Black-Box Models

    cs.RO 2025-09 conditional novelty 4.0 of 10

    AutoODD combines an LLM agent with per-axis Gaussian Process uncertainty to automatically discover failure modes of black-box models, demonstrated on missing-digit MNIST and aircraft detect-and-avoid.

  5. LLM Harms: A Taxonomy and Discussion

    cs.CY 2025-12 unverdicted novelty 3.0 of 10

    This paper proposes a taxonomy of LLM harms in five categories and suggests mitigation strategies plus a dynamic auditing system for responsible development.

Reference graph

Works this paper leans on

36 extracted references · 19 canonical work pages · cited by 5 Pith papers

  1. [1]

    A new taxonomy for automated driving: Structuring applications based on their operational design domain, level of automation and automation readiness,

    J. Betz, M. Lutwitzi, and S. Peters, “A new taxonomy for automated driving: Structuring applications based on their operational design domain, level of automation and automation readiness,” in 2024 IEEE Intelligent V ehicles Symposium (IV) , 2024, pp. 1–7

  2. [2]

    Safety testing of automated driving systems: A literature review,

    F. Khan, M. Falco, H. Anwar, and D. Pfahl, “Safety testing of automated driving systems: A literature review,” IEEE Access , 2023

  3. [3]

    Survey on scenario-based safety assessment of automated vehicles,

    S. Riedmaier, T. Ponn, D. Ludwig, B. Schick, and F. Diermeyer, “Survey on scenario-based safety assessment of automated vehicles,” IEEE access , vol. 8, pp. 87 456–87 477, 2020

  4. [4]

    Simulation-based identification of critical scenarios for cooperative and automated vehicles,

    S. Hallerbach, Y . Xia, U. Eberle, and F. Koester, “Simulation-based identification of critical scenarios for cooperative and automated vehicles,” SAE International Journal of Connected and Automated V ehicles, vol. 1, no. 2018-01-1066, pp. 93–106, 2018

  5. [5]

    Language models are few-shot learners,

    T. Brown, B. Mann, N. Ryder, M. Subbiah, J. D. Kaplan, P. Dhariwal, A. Neelakantan, P. Shyam, G. Sastry, A. Askell et al. , “Language models are few-shot learners,” Advances in neural information processing systems, vol. 33, pp. 1877–1901, 2020

  6. [6]

    Attention is all you need,

    A. Vaswani, “Attention is all you need,” Advances in Neural Information Processing Systems , 2017

  7. [7]

    A systematic survey of prompt engineering in large language models: Techniques and applications,

    P. Sahoo, A. K. Singh, S. Saha, V . Jain, S. Mondal, and A. Chadha, “A systematic survey of prompt engineering in large language models: Techniques and applications,” arXiv preprint arXiv:2402.07927 , 2024

  8. [8]

    Language prompt for autonomous driving,

    D. Wu, W. Han, T. Wang, Y . Liu, X. Zhang, and J. Shen, “Language prompt for autonomous driving,” arXiv preprint arXiv:2309.04379 , 2023

Show all 36 references
  1. [9]

    Dualad: Dual-layer planning for reasoning in autonomous driving,

    D. Wang, M. Kaufeld, and J. Betz, “Dualad: Dual-layer planning for reasoning in autonomous driving,” arXiv preprint arXiv:2409.18053 , 2024

  2. [10]

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

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

  3. [11]

    Critical scenario identification for realistic testing of autonomous driving systems,

    Q. Song, K. Tan, P. Runeson, and S. Persson, “Critical scenario identification for realistic testing of autonomous driving systems,” Software Quality Journal , vol. 31, no. 2, pp. 441–469, 2023

  4. [12]

    Reality bites: Assessing the realism of driving scenarios with large language models,

    J. Wu, C. Lu, A. Arrieta, T. Yue, and S. Ali, “Reality bites: Assessing the realism of driving scenarios with large language models,” in Proceedings of the 2024 IEEE/ACM First International Conference on AI F oundation Models and Software Engineering , 2024, pp. 40–51

  5. [13]

    Deepscenario: An open driving scenario dataset for autonomous driving system testing,

    C. Lu, T. Yue, and S. Ali, “Deepscenario: An open driving scenario dataset for autonomous driving system testing,” in 2023 IEEE/ACM 20th International Conference on Mining Software Repositories (MSR). IEEE, 2023, pp. 52–56

  6. [14]

    A comprehensive llm-powered framework for driving intelligence evaluation,

    S. You, X. Luo, X. Liang, J. Yu, C. Zheng, and J. Gong, “A comprehensive llm-powered framework for driving intelligence evaluation,” arXiv preprint arXiv:2503.05164 , 2025

  7. [15]

    CARLA: An open urban driving simulator,

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

  8. [16]

    Multimodal large language model driven scenario testing for autonomous vehicles,

    Q. Lu, X. Wang, Y . Jiang, G. Zhao, M. Ma, and S. Feng, “Multimodal large language model driven scenario testing for autonomous vehicles,” arXiv preprint arXiv:2409.06450 , 2024

  9. [17]

    Foundation models in autonomous driving: A survey on scenario generation and scenario analysis,

    Y . Gao, M. Piccinini, Y . Zhang, D. Wang, K. Moller, R. Brusnicki, B. Zarrouki, A. Gambi, J. F. Totz, K. Stormset al., “Foundation models in autonomous driving: A survey on scenario generation and scenario analysis,” IEEE Open Journal of Intelligent Transportation Systems , 2...

  10. [18]

    A survey on safety-critical driving scenario generation—a methodological perspective,

    W. Ding, C. Xu, M. Arief, H. Lin, B. Li, and D. Zhao, “A survey on safety-critical driving scenario generation—a methodological perspective,” IEEE Transactions on Intelligent Transportation Systems, vol. 24, no. 7, pp. 6971–6988, 2023

  11. [19]

    Factor graph scene distributions for automotive safety analysis,

    T. A. Wheeler and M. J. Kochenderfer, “Factor graph scene distributions for automotive safety analysis,” in 2016 IEEE 19th International Conference on Intelligent Transportation Systems (ITSC). IEEE, 2016, pp. 1035–1040

  12. [20]

    A new multi-vehicle trajectory generator to simulate vehicle-to-vehicle encounters,

    W. Ding, W. Wang, and D. Zhao, “A new multi-vehicle trajectory generator to simulate vehicle-to-vehicle encounters,” arXiv preprint arXiv:1809.05680, 2018

  13. [21]

    Analyzing and improving neural networks by generating semantic counterexamples through differentiable rendering,

    L. Jain, V . Chandrasekaran, U. Jang, W. Wu, A. Lee, A. Yan, S. Chen, S. Jha, and S. A. Seshia, “Analyzing and improving neural networks by generating semantic counterexamples through differentiable rendering,” arXiv preprint arXiv:1910.00727 , 2019

  14. [22]

    Corner case generation and analysis for safety assessment of autonomous vehicles,

    H. Sun, S. Feng, X. Yan, and H. X. Liu, “Corner case generation and analysis for safety assessment of autonomous vehicles,” Transportation research record, vol. 2675, no. 11, pp. 587–600, 2021

  15. [23]

    Building safer autonomous agents by leveraging risky driving behavior knowledge,

    A. Rana and A. Malhi, “Building safer autonomous agents by leveraging risky driving behavior knowledge,” in 2021 International Conference on Communications, Computing, Cybersecurity, and Informatics (CCCI) . IEEE, 2021, pp. 1–6

  16. [24]

    Robust trajectory prediction against adversarial attacks,

    Y . Cao, D. Xu, X. Weng, Z. Mao, A. Anandkumar, C. Xiao, and M. Pavone, “Robust trajectory prediction against adversarial attacks,” in Conference on Robot Learning . PMLR, 2023, pp. 128–137

  17. [25]

    Microscopic traffic simulation using sumo,

    P. A. Lopez, M. Behrisch, L. Bieker-Walz, J. Erdmann, Y .-P. Flötteröd, R. Hilbrich, L. Lücken, J. Rummel, P. Wagner, and E. Wießner, “Microscopic traffic simulation using sumo,” in The 21st IEEE International Conference on Intelligent Transportation Systems . IEEE, 2018. [Onl...

  18. [26]

    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,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 15 459–15 469

  19. [27]

    Traffic scene generation from natural language description for autonomous vehicles with large language model,

    B.-K. Ruan, H.-T. Tsui, Y .-H. Li, and H.-H. Shuai, “Traffic scene generation from natural language description for autonomous vehicles with large language model,” arXiv preprint arXiv:2409.09575 , 2024

  20. [28]

    Chatsumo: Large language model for automating traffic scenario generation in simulation of urban mobility,

    S. Li, T. Azfar, and R. Ke, “Chatsumo: Large language model for automating traffic scenario generation in simulation of urban mobility,” IEEE Transactions on Intelligent V ehicles , 2024

  21. [29]

    Commonroad: Composable benchmarks for motion planning on roads,

    M. Althoff, M. Koschi, and S. Manzinger, “Commonroad: Composable benchmarks for motion planning on roads,” in 2017 IEEE Intelligent V ehicles Symposium (IV). IEEE, 2017, pp. 719–726

  22. [30]

    Lanelets: Efficient map representation for autonomous driving,

    P. Bender, J. Ziegler, and C. Stiller, “Lanelets: Efficient map representation for autonomous driving,” in 2014 IEEE Intelligent V ehicles Symposium Proceedings. IEEE, 2014, pp. 420–425

  23. [31]

    Frenetix: A high-performance and modular motion planning framework for autonomous driving,

    R. Trauth, K. Moller, G. Würsching, and J. Betz, “Frenetix: A high-performance and modular motion planning framework for autonomous driving,” IEEE Access , pp. 1–1, 2024

  24. [32]

    Gpt-4 technical report,

    J. Achiam, S. Adler, S. Agarwal, L. Ahmad, I. Akkaya, F. L. Aleman, D. Almeida, J. Altenschmidt, S. Altman, S. Anadkat et al. , “Gpt-4 technical report,” arXiv preprint arXiv:2303.08774 , 2023

  25. [33]

    Structgpt: A general framework for large language model to reason over structured data,

    J. Jiang, K. Zhou, Z. Dong, K. Ye, W. X. Zhao, and J.-R. Wen, “Structgpt: A general framework for large language model to reason over structured data,” arXiv preprint arXiv:2305.09645 , 2023

  26. [34]

    Gemini: a family of highly capable multimodal models,

    G. Team, R. Anil, S. Borgeaud, J.-B. Alayrac, J. Yu, R. Soricut, J. Schalkwyk, A. M. Dai, A. Hauth, K. Millican et al. , “Gemini: a family of highly capable multimodal models,” arXiv preprint arXiv:2312.11805, 2023

  27. [35]

    Deepseek llm: Scaling open-source language models with longtermism,

    X. Bi, D. Chen, G. Chen, S. Chen, D. Dai, C. Deng, H. Ding, K. Dong, Q. Du, Z. Fu et al. , “Deepseek llm: Scaling open-source language models with longtermism,” arXiv preprint arXiv:2401.02954 , 2024

  28. [36]

    Mptree: A sampling-based vehicle motion planner for real-time obstacle avoidance,

    M. Piazza, M. Piccinini, S. Taddei, and F. Biral, “Mptree: A sampling-based vehicle motion planner for real-time obstacle avoidance,” IF AC-PapersOnLine, vol. 58, no. 10, pp. 146–153, 2024, 17th IFAC Symposium on Control of Transportation Systems CTS 2024

Pith tools

Reviewed August 9, 2026 · model on record in the stance chip above.