Pith. sign in

REVIEW 4 major objections 5 minor 40 references

HySafe-AI: Hybrid Safety Architectural Analysis Framework for AI Systems: A Case Study

T0 review · 4 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

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

Pith's one-line read HySAFE-AI claims that classical FMEA/FTA can be adapted to foundation-model self-driving systems by adding architectural transparency and an AI-specific failure taxonomy mapped onto standard guidewords.

desk verdict A clear, standards-grounded framework for applying FMEA/FTA to foundation-model driving stacks, but the claimed safety improvement rests entirely on assumed detection-difficulty reductions rather than measured monitor performance. read the letter →

arxiv 2507.17118 v1 pith:JBLEBVJ5 submitted 2025-07-23 cs.AI

Mandar Pitale , Jelena Frtunikj , Abhinaw Priyadershi , Vasu Singh , Maria Spence This is my paper
classification cs.AI
keywords FMEAfaulttreeanalysisfoundationmodelsautonomousdrivingend-to-endarchitecturelatentspaceAIsafetyfunctional
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

This paper argues that the standard safety analyses used in automotive engineering, failure modes and effects analysis (FMEA) and fault tree analysis (FTA), can be made to work on end-to-end self-driving systems built from foundation models, despite those models being largely opaque. The route is HySAFE-AI: require enough architectural transparency to see components such as encoders, latent denoisers, and decoders, analyze them at multiple levels of abstraction, and translate generic AI failure modes such as hallucination, temporal misprediction, quantization effects, and dataset staleness into the familiar FMEA guidewords 'incorrect value' and 'missing value.' A case study on a reference generative driving architecture identifies high-risk failure modes and shows that adding policy monitors, physics-based safety evaluators, and plan arbitration lowers risk-priority numbers substantially. If the argument holds, safety engineers no longer need white-box access to a neural network to do systematic hazard analysis on it; they need a component-level map and a failure taxonomy.

What carries the argument

The load-bearing mechanism is the AI-specific failure taxonomy paired with the reference architecture. The taxonomy is a table that assigns each generic AI failure mode, such as hallucination, temporal reasoning failure, quantization-induced feature degradation, or dataset staleness, a classical FMEA guideword like 'incorrect value,' 'missing value,' or 'value too high/low,' and then a domain-specific manifestation such as 'phantom pedestrian' or 'undetected road boundary in unusual lighting.' The reference architecture decomposes a generative driving model into encoder, latent denoiser or U-Net, decoder, trajectory planner, and text conditioning blocks, so each failure can be localized to a system element and scored as severity, occurrence, and detection to yield a risk priority number (RPN = S × O × D). This mapping is what lets a closed-box model be analyzed with open-box techniques; without it, FMEA has no components to attach failure modes to.

What would settle it

Collect observed failure data from a deployed or simulated end-to-end driving model, classify every failure into the paper's taxonomy, and compare the empirically frequent and severe failures against the high-RPN entries (quantization-induced hallucination, temporal reasoning failure, dataset staleness). If the observed dominant failure modes do not appear in the taxonomy, or if the severity/occurrence/detection scores do not track real incident rates, the mapping loses its predictive value.

Watch

Extended reading notes

Core claim

On its own terms, the paper's central claim is that foundation-model-based end-to-end autonomous driving stacks can be subjected to classical FMEA and FTA if the analysis is hybridized: architectural transparency is treated as a precondition, the system is examined across entities from raw inputs to latent spaces, and an AI-specific failure taxonomy is mapped onto standard FMEA guidewords. Applying this to a reference architecture assembled from two recent generative driving models, the paper identifies the highest-risk failures as quantization-induced hallucination in the latent denoiser with an RPN of 252, temporal reasoning failure in causal temporal attention with an RPN of 252, and dataset staleness with an RPN of 216. It then shows that targeted mitigations, including a quantization-calibrated uncertainty monitor, trajectory consistency checks, an active learning pipeline, a physics-based safety evaluator, and an arbitration layer, cut those RPNs substantially by improving detectability and turning single-point failures into multi-point failures.

Load-bearing premise

The argument stands or falls on the claim that the reference architecture assembled from generative driving models is representative of real end-to-end systems, and that each internal piece (encoder, denoiser, U-Net, decoder) fails in the specific ways the taxonomy lists; the paper asserts this mapping without showing empirical evidence that those latent-space components actually fail that way.

Editorial extensions

If this is right

  • An end-to-end driving stack can be safety-analyzed without interpreting individual neurons; identifying the architectural entity that produces a latent representation is sufficient to attach failure modes.
  • Deployment optimizations such as quantization stop being a blind spot: precision loss is treated as an 'incorrect or missing value' failure and can be mitigated with quantization-aware training or mixed precision.
  • Runtime safety layers, namely a policy monitor, a safety evaluator, and an arbitrator, are presented as the architectural consequence of the analysis, turning single-point latent-space failures into dual- or multi-point failures.
  • The resulting fused architecture is claimed to preserve the predictive capability of foundation models while aligning with functional-safety standards such as the road-vehicle functional-safety standard and the automotive AI safety standard.

Reading between the lines

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

  • Because the failure taxonomy is domain-general, the same guideword mapping could be carried into other foundation-model safety-critical applications, such as surgical robotics or industrial control, with only the domain-specific manifestation column changed.
  • The S/O/D ratings in the case study are expert judgments with no empirical calibration; the most useful next test would be to compare the predicted high-RPN modes against incident logs or fault-injection experiments in a real deployment.
  • If such calibration were done, the RPN reductions could be read as quantitative safety claims; absent that, the reductions are best read as a relative ranking of where to put safety monitors.
Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 5 minor

Summary. The paper argues that traditional safety analyses such as FMEA and FTA are poorly matched to foundation-model-based end-to-end autonomous driving systems because of abstraction incompatibility, causal opacity, and temporal dynamism. It proposes HySAFE-AI, a framework that adds architectural transparency, multi-level abstraction, and an AI-specific failure taxonomy mapped to standard FMEA guidewords. The authors construct a reference E2E architecture from GenAD and GAIA-2 components, walk through an FMEA (Table III) and an FTA (Figure 2), propose architectural mitigations such as policy monitors, safety evaluators, and arbitration (Table IV), and present reduced RPN values after mitigation (Table V). The paper concludes that the resulting fused architecture reduces safety-critical risk while acknowledging computational overhead as future work.

Significance. If the framework's claims were supported by evidence, it would give safety engineers a structured way to trace AI-specific failures such as hallucination, temporal misprediction, and quantization effects through latent-space components and map them to established FMEA categories, aligning with emerging standards like ISO/PAS 8800. The paper is valuable as a worked example and a taxonomy proposal, and it is clearly organized around a concrete reference architecture. However, the quantitative safety-improvement claim is currently built from expert-assumed severity, occurrence, and detection scores, and the post-mitigation RPN reductions are obtained by assuming detection-score improvements without measured detector performance. The stress-test concern about Table V therefore lands: the central claim that the fused architecture is meaningfully safer is not yet evidenced by the manuscript.

major comments (4)
  1. [§III, Table V] The central quantitative conclusion that the fused architecture is significantly safer rests on assumed D reductions (4→1, 3→1, 2→1) with S and O held constant, while the paper itself admits in the walkthrough that the RPN values are based on expert judgment. This makes the RPN delta a re-statement of the input assumption rather than evidence of improvement. No detection experiment, false-positive budget, latency measurement, or failure-frequency data is provided for the Policy Monitor, Safety Evaluator, or arbitration logic. Please reframe Table V as a hypothesized post-mitigation assessment, or calibrate the D scores using measured monitor performance, or provide a sensitivity analysis showing which conclusions survive plausible variation in D.
  2. [§II, Reference E2E Architecture; Fig. 1] The reference architecture is an amalgam of GenAD and GAIA-2 components, both described in the paper as designed for simulation and motion planning, and its representativeness of a deployed end-to-end ADS control stack is asserted rather than established. The failure semantics assigned to latent-space components in Table III, such as quantization-induced hallucination in the Latent Denoiser, are plausible but are not supported by data showing that these components fail in the claimed manner. Please state explicitly that the case study is illustrative and either justify the representativeness claim or narrow its scope.
  3. [§III, HySafe-AI Outcome] The paper claims that HySAFE-AI improves traditional safety analyses, but it does not run a conventional FMEA/FTA baseline on the same reference architecture. Without such a comparison, it is not possible to determine which identified failure modes, traceability paths, or mitigations are attributable to the augmentation rather than to the analysts' prior knowledge. A side-by-side comparison, or at least an explicit enumeration of the additional failure modes uncovered by HySAFE-AI, would be needed to support the improvement claim.
  4. [§III, Fig. 2] The FTA diagrams are presented as qualitative illustrations; no minimal cut sets, failure probabilities, or common-cause analysis are provided. The statement that the mitigations convert single-point failures into dual- or multi-point failures is therefore not supported by the diagram alone. Please either label this as an illustrative fault tree or provide the corresponding cut-set analysis to substantiate the claim.
minor comments (5)
  1. [§III, walkthrough] The sentence 'The RPN values are based on expert judgment and and the reasoning for the top 3 RPN values' contains a duplicated 'and'.
  2. [§II, Table II] Cells in Table II cite the same reference number multiple times without distinguishing specific works, which makes the table difficult to interpret; consider listing the specific cited works in each cell.
  3. [§IV] The sentence 'the risk of safety critical failure risks is reduced' is grammatically unclear and should be rephrased.
  4. [References] Reference [28] is listed as 'arXiv preprint (2024)' without an arXiv identifier; please complete the citation.
  5. [Figures] Ensure all figures, especially Figure 1 and Figure 2, are legible in the final version and that their captions fully describe the content, since the printed text frequently refers to specific diagram details.

Circularity Check

1 steps flagged · score 6.0 of 10

The claimed safety improvement in Table V is an arithmetic restatement of the paper's own assumed D-score deltas, not an independently derived result.

  1. fitted input called prediction [Section III, 'HySafe-AI Outcome: Fused Stack Model' (Table V and preceding paragraph)]
    "The RPN values are based on expert judgment ... The FMEA in Table V illustrates how the AI/ML-specific failure modes of the reference architecture can be systematically mitigated through the architectural measures, i.e. safety-aware components like the Policy Monitor and Safety Evaluator, resulting in significantly reduced RPN values."

    Table V changes only the D column (e.g., 4→1, 3→1, 2→1) while S and O are unchanged; since RPN = S×O×D, the 'RPN Delta' column is the arithmetic product of the assumed D deltas, not an independent outcome. The paper supplies no measured detection rates, false-positive budgets, or validation results showing that the Policy Monitor, Safety Evaluator, or Quantization-Aware Training actually achieve D=1; instead it states that the RPN values are based on expert judgment. The claimed 'significantly reduced RPN values' therefore restates the authors' own assumptions: the same expert judgment that set baseline D also set the post-mitigation D, so the improvement is built into the input rather than derived or measured. This makes the demonstration self-referential and forced by construction.

full rationale

HySAFE-AI's core methodological contribution—augmenting FMEA/FTA with architectural transparency, multi-level abstraction, and an AI-specific failure taxonomy mapped to standard guidewords—is not circular in itself; it is a proposal illustrated on a synthetic reference architecture assembled from GenAD and GAIA-2 components. The paper's self-citations ([28], [34]) are not load-bearing: [28] supports only the general claim of cascading errors in modular architectures, and [34] appears only in future-work suggestions. The circularity is confined to the illustrative demonstration of value. Table III's baseline S/O/D values are expert judgments, which the paper acknowledges ('The RPN values are based on expert judgment'). Table V then changes only D (e.g., 4→1) and reports RPN reductions (e.g., 252→63) as a demonstration that the fused architecture is safer. Because RPN = S×O×D, every 'RPN Delta' is the deterministic product of the assumed D delta; no measurement of monitor detection performance, false-positive rate, latency, or real occurrence data is supplied. Thus the 'significantly reduced RPN' conclusion is an assumption encoded as an output, not evidence that the proposed monitors and arbitrators make the architecture safer. This is a partial circularity: it undermines the illustrative case study, but not the methodological mapping itself. Score 6 reflects that one or more 'predictions' (the reduced RPNs) reduce by construction to the input D deltas.

Assumptions & free parameters 2 free parameters · 4 assumptions · 3 invented entities

The paper's contributions are conceptual, and the quantitative elements (RPN scores and their reductions) are expert-assigned rather than empirically derived. This means the framework's apparent validation is largely self-constructed; no external data or benchmark constrains the numbers.

free parameters (2)
  • FMEA S/O/D scores for each failure mode = See Table III: e.g., 9,7,4 for quantized latent denoiser; 9,6,4 for dataset staleness
    Assigned by expert judgment; the ranking of failure modes and the top-RPN conclusions depend entirely on these numbers.
  • Mitigation D reductions = D reduced by 1 to 3 per row in Table V
    Assumed effectiveness of the proposed safety components; no operational or experimental evidence; these reductions drive the displayed RPN improvements.
assumptions (4)
  • domain assumption FMEA and FTA are suitable templates that can be extended via guideword mapping.
    Section III uses this premise to justify mapping AI failure modes to guidewords such as incorrect value and missing value.
  • domain assumption The GenAD/GAIA-2 reference architecture in Figure 1 represents state-of-the-art end-to-end ADS architectures.
    Section II constructs the unified reference model from [12] and [31]; if real production end-to-end stacks differ substantially, the FMEA results may not transfer.
  • domain assumption Latent-space components (encoder, denoiser, U-Net, decoder) can be analyzed as distinct architectural entities despite closed-box operation.
    Section III explicitly names architectural transparency as a precondition; no evidence is provided that such transparency is achievable.
  • ad hoc to paper Expert judgment is a valid source of severity, occurrence, and detection scores in the absence of field data.
    The entire RPN ranking rests on this; the paper gives no calibration or inter-rater validation.
invented entities (3)
  • Policy Monitor
    purpose: Runtime detection of out-of-distribution inputs and uncertainty quantification for end-to-end planner outputs.
    Proposed as mitigation; no implementation, data, or safety case validation is provided in the paper.
  • Safety Evaluator
    purpose: Applies rule-based and physics-derived checks to reject unsafe trajectories.
    Proposed component; effectiveness assumed in Table V D reductions.
  • Plan Arbitration
    purpose: Selects the highest-confidence trajectory that passes policy and safety checks.
    Proposed architecture element; no independent evidence of behavior.

how reviews work

0 comments
Cite this review

Pith. "Pith review of HySafe-AI: Hybrid Safety Architectural Analysis Framework for AI Systems: A Case Study." pith.science (2026). https://pith.science/paper/JBLEBVJ5

@misc{pith2026250717118,
  author       = {Pith},
  title        = {Pith review of: HySafe-AI: Hybrid Safety Architectural Analysis Framework for AI Systems: A Case Study},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/JBLEBVJ5}},
  note         = {Machine review of arXiv:2507.17118}
}
read the original abstract

AI has become integral to safety-critical areas like autonomous driving systems (ADS) and robotics. The architecture of recent autonomous systems are trending toward end-to-end (E2E) monolithic architectures such as large language models (LLMs) and vision language models (VLMs). In this paper, we review different architectural solutions and then evaluate the efficacy of common safety analyses such as failure modes and effect analysis (FMEA) and fault tree analysis (FTA). We show how these techniques can be improved for the intricate nature of the foundational models, particularly in how they form and utilize latent representations. We introduce HySAFE-AI, Hybrid Safety Architectural Analysis Framework for AI Systems, a hybrid framework that adapts traditional methods to evaluate the safety of AI systems. Lastly, we offer hints of future work and suggestions to guide the evolution of future AI safety standards.

Figures

Figures reproduced from arXiv: 2507.17118 by the authors.

Figure 1
Figure 1. Reference Architecture of Unified Framework [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. (a) HySafe-AI FTA, (b) HySafe-AI FTA with mitigations [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Fused Architecture [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

40 extracted references · 29 canonical work pages

  1. [1]

    In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR)

    Alahi, A., Goel, K., Ramanathan, V ., Robicquet, A., Fei-Fei, L., Savarese, S.: Social lstm: Human trajectory prediction in crowded spaces. In: Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition (CVPR). pp. 961–971 (2016)

    work page 2016

  2. [2]

    In: IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR) Workshops

    Bojarski, M., Choromanska, A., Choromanski, K., Firner, B., Jackel, L., Muller, U., Zieba, K.: Visualbackprop: Visualizing cnns for autonomous driving. In: IEEE Conference on Computer Vision and Pattern Recog- nition (CVPR) Workshops. pp. 470–477 (2017)

    work page 2017

  3. [3]

    Bojarski, M., Del Testa, D., Dworakowski, D., Firner, B., Flepp, B., Goyal, P., Jackel, L., Monfort, M., Muller, U., Zhang, J., Zhang, X., Zhao, J., Zieba, K.: End to end learning for self-driving cars (2016), available: https://arxiv.org/abs/1604.07316

    arXiv 2016

  4. [4]

    In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J

    Carion, N., Massa, F., Synnaeve, G., Usunier, N., Kirillov, A., Zagoruyko, S.: End-to-end object detection with transformers. In: Vedaldi, A., Bischof, H., Brox, T., Frahm, J. (eds.) ECCV 2020. LNCS, vol. 12346. pp. 213–229. Springer, Cham (2020)

    work page 2020

  5. [5]

    In: Proceedings of the IEEE International Conference on Robotics and Automation (ICRA)

    Codevilla, F., M ¨uller, M., L ´opez, A., Koltun, V ., Dosovitskiy, A.: End- to-end driving via conditional imitation learning. In: Proceedings of the IEEE International Conference on Robotics and Automation (ICRA). pp. 4693–4700 (2018). https://doi.org/10.1109/ICRA.2018.8460487

    work page arXiv 2018

  6. [6]

    Dal’Col, L., Oliveira, M., Santos, V .: Joint perception and prediction for autonomous driving: A survey (2024), available: https://arxiv.org/abs/2412.14088

    arXiv 2024

  7. [7]

    In: International Conference on Learning Representations (ICLR) (2021)

    Dosovitskiy, A., Beyer, L., Kolesnikov, A., Weissenborn, D., Zhai, X., Unterthiner, T., Dehghani, M., Minderer, M., Heigold, G., Gelly, S., Uszkoreit, J., Houlsby, N.: An image is worth 16x16 words: Trans- formers for image recognition at scale. In: International Conference on Learning Representations (ICLR) (2021)

    work page 2021

  8. [8]

    Erabati, G., Araujo, H.: Msf3ddetr: Multi-sensor fusion 3d detection transformer for autonomous driving (2022), available: https://arxiv.org/abs/2210.15316

    work page Pith review arXiv 2022

Show all 40 references
  1. [9]

    Gajewski, P., ˙Zurek, D., Pietro ´n, M., Faber, K.: Solving multi- goal robotic tasks with decision transformer (2024), available: https://arxiv.org/abs/2410.06347

    2024 arXiv

  2. [10]

    In: Annual Reliability and Maintainability Symposium

    Goddard, P.: Software fmea techniques. In: Annual Reliability and Maintainability Symposium. 2000 Proceedings. International Sympo- sium on Product Quality and Integrity (Cat. No.00CH37055). pp. 118– 123 (2000). https://doi.org/10.1109/RAMS.2000.816294

    2000

  3. [11]

    https://doi.org/10.1109/DASC62030.2024.10749313

    Hogen, A., Annighoefer, B., Daw, Z.: Failure modes or not failure modes? integrating machine learning in aerospace safety assessment processes (2024). https://doi.org/10.1109/DASC62030.2024.10749313

    2024

  4. [12]

    Hu, A., Russell, L., Yeo, H., Murez, Z., Fedoseev, G., Kendall, A., Shotton, J., Corrado, G.: Gaia-1: A generative world model for autonomous driving (2023), available: https://wayve.ai/research/GAIA-1/

    2023

  5. [13]

    arXiv.org (October 30 2024), available: https://arxiv.org/abs/2410.23262v1

    Hwang, J., Xu, R., Lin, H., Hung, W., Ji, J., Choi, K., Huang, D., He, T., Covington, P., Sapp, B., Zhou, Y ., Guo, J., Anguelov, D., Tan, M.: Emma: End-to-end multimodal model for autonomous driving. arXiv.org (October 30 2024), available: https://arxiv.org/abs/2410.23262v1

    2024 arXiv

  6. [14]

    International Electrotechnical Commission (IEC): IEC 61508:2010 – Functional Safety of Electrical/Electronic/Programmable Electronic Safety-related Systems (2010)

    2010

  7. [15]

    International Organization for Standardization: ISO 26262:2018 – Road Vehicles – Functional Safety (2018), available: https://www.iso.org/standard/68383.html

    2018

  8. [16]

    International Organization for Standardization: ISO 21448:2022 – Road Vehicles — Safety of the Intended Functionality (SOTIF) (2022), available: https://www.iso.org/standard/77490.html

    2022

  9. [17]

    International Organization for Standardization: ISO/IEC TR 5469:2024 – Artificial intelligence — Functional Safety and AI systems (2024), available: https://www.iso.org/standard/83303.html

    2024

  10. [18]

    International Organization for Standardization: ISO/PAS 8800:2024 – Road vehicles — Safety and Artificial Intelligence (2024), available: https://www.iso.org/standard/16012.html

    2024

  11. [19]

    International Organization for Standardization: ISO/TS 5083:2025 – Road vehicles — Safety for automated driving systems — Design, verification and validation (2025), available: https://www.iso.org/standard/81920.html

    2025

  12. [20]

    International Organization for Standardization: ISO/IEC AWI TS 22440-1 – Artificial intelligence — Functional safety and AI systems, Part 1: Requirements (Under Development), available: https://www.iso.org/standard/89535.html

  13. [21]

    International Organization for Standardization: ISO/IEC AWI TS 25223 – Artificial intelligence — Guidance and requirements for uncer- tainty quantification in AI systems (Under Development), available: https://www.iso.org/standard/89475.html

  14. [22]

    In: 2014 21st Asia-Pacific Software Engineering Conference

    Kim, H.H.: Sw fmea for iso-26262 software development. In: 2014 21st Asia-Pacific Software Engineering Conference. vol. 2 (2014). https://doi.org/10.1109/APSEC.2014.85

    2014 doi

  15. [23]

    In: AI Ethics, V olume 2, 2022

    Li, J., C.M.: Fmea-ai: Ai fairness impact assessment using failure mode and effects analysis. In: AI Ethics, V olume 2, 2022. Springer (2022)

    2022

  16. [24]

    In: Computer Vision – ECCV 2022: 17th European Conference on Computer Vision, Tel Aviv, Israel, October 23–27, 2022, Proceedings, Part IX

    Li, Z., Wang, W., Li, H., Xie, E., Sima, C., Lu, T., Qiao, Y ., Dai, J.: Bevformer: Learning bird’s-eye-view representation from multi-camera images via spatiotemporal transformers. In: Computer Vision – ECCV 2022: 17th European Conference on Computer Vision, Tel Aviv, Israel,...

    2022

  17. [25]

    In: 2023 7th Interna- tional Conference on System Reliability and Safety (ICSRS) (2023)

    Martinez, J., Eguia, A., Urretavizcaya, I., Amparan, E., Negro, P.L.: Fault tree analysis and failure modes and effects analysis for systems with artificial intelligence: A mapping study. In: 2023 7th Interna- tional Conference on System Reliability and Safety (ICSRS) (2023). ...

    2023

  18. [26]

    Mohanty, V ., Agrawal, S., Datta, S., Ghosh, A., Sharma, V ., Chakravarty, D.: Deepvo: A deep learning approach for monocular visual odometry (2016), available: https://arxiv.org/abs/1611.06069

    2016 arXiv

  19. [27]

    Mylius, S.: Systematic hazard analysis for frontier ai using stpa (2025), https://arxiv.org/abs/2506.01782

    2025 arXiv

  20. [28]

    arXiv preprint (2024)

    Pitale, M., Abbaspour, A., Upadhyay, D.: Inherent diverse redundant safety mechanisms for ai-based software elements in automotive appli- cations. arXiv preprint (2024)

    2024

  21. [29]

    Radford, A., Kim, J., Hallacy, C., Ramesh, A., Goh, G., Agarwal, S., Sastry, G., Askell, A., Mishkin, P., Clark, J., Krueger, G., Sutskever, I.: Learning transferable visual models from natural language supervision (2021), available: https://arxiv.org/abs/2103.00020

    2021 arXiv

  22. [30]

    Sensors (Basel) (2024)

    Rosero, L., Gomes, I., da Silva, J., Przewodowski, C., Wolf, D., Os ´orio, F.: Integrating modular pipelines with end-to-end learning: A hybrid approach for robust and reliable autonomous driving systems. Sensors (Basel) (2024). https://doi.org/10.3390/s24072097

    2024 doi

  23. [31]

    Russell, L., Hu, A., Bertoni, L., Fedoseev, G., Shotton, J., Arani, E., Corrado, G.: Gaia-2: A controllable multi-view gen- erative world model for autonomous driving (2025), available: https://wayve.ai/research/GAIA-2/

    2025

  24. [32]

    Shah, D., Osinski, B., Krueger, Ichter, B., Levine, S.: Lm-nav: Robotic navigation with large pre-trained models of language, vision, and action (2022), available: https://arxiv.org/abs/2207.04429

    2022 arXiv

  25. [33]

    IEEE Transactions on Pattern Analysis and Machine Intelligence 46(5), 3955–3971 (2024)

    Shi, S., Jiang, L., Dai, D., Schiele, B.: Mtr++: Multi-agent motion pre- diction with symmetric scene modeling and guided intention querying. IEEE Transactions on Pattern Analysis and Machine Intelligence 46(5), 3955–3971 (2024). https://doi.org/10.1109/TPAMI.2024.3352811

    2024

  26. [34]

    In: 4th IEEE International Conference on Industrial Cyber-Physical Systems (2021)

    Singh, V ., Pitale, M.: Impact of automotive system safety design on machine learning based perception systems. In: 4th IEEE International Conference on Industrial Cyber-Physical Systems (2021)

    2021

  27. [35]

    In: IEEE International Conference on Robotics and Automation (2017)

    Wang, S., Clark, R., Wen, H., Trigoni, N.: Deepvo: Towards end-to-end visual odometry with deep recurrent convolutional neural networks. In: IEEE International Conference on Robotics and Automation (2017)

    2017

  28. [36]

    Wang, Y ., Jiao, R., Zhan, S., Lang, C., Huang, C., Wang, Z., Yang, Z., Zhu, Q.: Empowering autonomous driving with large language models: A safety perspective (2024), available: https://arxiv.org/abs/2403.12345

    2024 arXiv

  29. [37]

    Wang, Y ., Liu, D., Jeon, H., Chu, Z., Matson, E.: End-to-end learning approach for autonomous driving: A convolutional neural network model. Tech. rep., Department of Computer and Information Technol- ogy, Purdue University, U.S.A

  30. [38]

    Information Fusion 114, 102671 (2025)

    Xu, H., Chen, J., Meng, S., Wang, Y ., Chau, L.: A survey on occupancy perception for autonomous driving: The informa- tion fusion perspective. Information Fusion 114, 102671 (2025). https://doi.org/10.1016/j.inffus.2024.102671

    2025

  31. [39]

    Yang, J., Gao, S., Qiu, Y ., Chen, L., Li, T., Dai, B., Chitta, K., Wu, P., Zeng, J., Luo, P., Zhang, J., Geiger, A., Qiao, Y ., Li, H.: Genad: Generalized predictive model for autonomous driving (2023), available: https://arxiv.org/abs/2403.09630

    2023 arXiv

  32. [40]

    Yang, Z., Jia, X., Li, H., Ya, J.: Llm4ad: A survey of large language models for autonomous driving (2023), available: https://arxiv.org/abs/2311.01043v1

    2023 arXiv

Pith tools

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