Pith. sign in

REVIEW 3 major objections 5 minor 1 cited by

SEAL: Vision-Language Model-Based Safe End-to-End Cooperative Autonomous Driving with Adaptive Long-Tail Modeling

T0 review · 3 major / 5 minor · reviewed 2026-08-06 · deepseek-v4-flash

Pith's one-line read SEAL claims a VLM-based cooperative driving pipeline can stay accurate and collision-free in snow and fog by training on GPT-4o-generated scenes, with reported average L2 error of 0.6779 m and collision rate of 0.0662 on augmented DAIR-V2X.

desk verdict SEAL's headline robustness result is undermined by a circular test set and an oracle scenario label, though the mechanisms themselves are concrete. read the letter →

arxiv 2506.21041 v2 pith:XF73NSMC submitted 2025-06-26 cs.RO cs.AIcs.CV

classification cs.ROcs.AIcs.CV
keywords end-to-endautonomousdrivingV2Xcooperationvision-languagemodellong-tailscenariostrajectoryplanningadverseweatherscenario-awareattentioncontrastivelearning
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

SEAL is a vision-language-model-based end-to-end cooperative driving framework aimed at a specific safety gap: car-and-infrastructure planners that work in clear weather often fail under rare snow and fog. The paper tries to show that this gap can be closed without collecting real adverse-weather data, by (1) using GPT-4o prompt-based image transformation to synthesize photorealistic snow and fog on both vehicle and infrastructure views, (2) a gated multi-scenario adaptive attention module that uses scenario labels to recalibrate visual features, and (3) a multi-task contrastive objective that aligns image-text embeddings while separating scenarios. On the DAIR-V2X benchmark augmented through this same pipeline, SEAL reports the lowest average trajectory error ($L_2$ 0.6779 m) and lowest average collision rate (6.62%) among the compared cooperative driving baselines. If the results hold, the paper would establish a scalable route to long-tail robustness for cooperative autonomy at real-time speed (~15.4 FPS).

What carries the argument

The two load-bearing mechanisms are GMSAA and MSCL, supported by the GPT-4o data-generation pipeline. GMSAA (gated multi-scenario adaptive attention) pools visual tokens into a global descriptor, computes attention logits over scenario types, adds a scenario-label bias and a snow/fog mutual-exclusion penalty, blends the result with a similarity-guided prior from a learnable scenario similarity matrix, then extracts scenario-specific features and gates their residual addition to the token stream; it is the component designed to recalibrate ambiguous or corrupted features using the scenario prior. MSCL (multi-task scenario-aware contrastive learning) combines a modality contrastive loss between visual and textual embeddings with an intra-modal scenario contrastive term that clusters same-scenario images and separates different ones, weighted by inverse scenario frequency to emphasize rare conditions. The prompt-driven pipeline supplies the scenario labels and the augmented training and evaluation data: GPT-4o transforms both vehicle- and infrastructure-side images under shared weather prompts, and a composite score over LPIPS, BRISQUE, FID, FADE, and Semantic IoU is used to filter for realism and semantic preservation.

What would settle it

A decisive test would run the same SEAL training procedure but evaluate on real collected snow and fog driving frames rather than transformed images; if the reported $L_2$ error and collision-rate advantage over V2X-VLM shrinks or disappears, the claimed long-tail robustness is an artifact of the synthetic test distribution. A second decisive test would replace the ground-truth scenario label $d$ with a classifier's prediction during inference; a large performance drop would show the mechanism depends on oracle information that deployment would not provide.

Watch

Extended reading notes

Core claim

SEAL's central claim is that a vision-language cooperative planner can be made safe in long-tail weather by conditioning the visual stream on scenario identity and by training on prompt-generated adverse-weather data. The architecture builds on a V2X-VLM-style design that fuses vehicle and infrastructure images with scene descriptions, and adds three components: a GPT-4o-based transformation pipeline that turns normal frames into heavy snow and dense fog views, a GMSAA module that injects the scenario label into attention and gates scenario-specific feature refinement, and an MSCL loss that pulls image-text pairs together while clustering same-scenario images and separating different scenarios. The claimed outcome is an end-to-end model that plans trajectories with the lowest average $L_2$ error (0.6779 m over 2.5/3.5/4.5 s horizons) and the lowest average collision rate (0.0662) versus V2X-VLM, UniV2X, and CooperNaut on the augmented DAIR-V2X test set, while keeping communication cost and latency comparable to V2X-VLM.

Load-bearing premise

SEAL's central claim rests on the premise that synthetic snow and fog images generated by the image-generation model preserve the geometry and semantics needed for trajectory planning, and that the true scenario label is available as an input to the gating module at test time.

Editorial extensions

If this is right

  • If SEAL's reported numbers hold, cooperative end-to-end planners can be made resilient to snow and fog without collecting real adverse-weather data: prompt-based synthesis on existing V2X datasets is a sufficient training signal.
  • The GMSAA module adds only modest overhead, so scenario-aware feature recalibration is compatible with real-time operation; the paper reports 15.36 FPS and 255 ms total latency.
  • Ablations attribute large portions of the gain to infrastructure views, scene text, and the GMSAA and MSCL modules, implying each of those inputs contributes to long-tail planning accuracy and safety.
  • Because the scenario-split results stay bounded in snow and fog where UniV2X and CooperNaut errors inflate strongly, the mechanism appears to improve worst-case conditions rather than only average performance.

Reading between the lines

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

  • An extension the paper does not develop: at deployment the scenario label $d$ must be predicted rather than taken from ground truth; coupling SEAL with a lightweight weather or domain classifier is the natural next step, and its success would determine whether the gating benefit survives outside the benchmark's oracle setting.
  • A testable extension would apply the same prompt-driven pipeline to rain, night, and partial occlusion; since GMSAA and MSCL treat scenario identity as a categorical prior, the framework's benefit should transfer to any condition that can be labeled and synthesized.
  • The composite generation-quality score uses manually assigned per-weather weights; a data-driven variant that tunes those weights by downstream planning error could make the augmentation selection criterion more directly safety-relevant.
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 / 5 minor

Summary. The paper proposes SEAL, a vision-language-model-based cooperative end-to-end autonomous driving framework that extends the authors' prior V2X-VLM work. SEAL has three main components: a prompt-driven GPT-4o pipeline that synthesizes snow and fog versions of existing DAIR-V2X frames; a gated multi-scenario adaptive attention (GMSAA) module that recalibrates visual features using a scenario label; and a multi-task scenario-aware contrastive learning (MSCL) objective that aligns vision-language embeddings while separating scenario domains. Experiments on the augmented DAIR-V2X dataset report that SEAL achieves the lowest average L2 trajectory error and collision rate compared with V2X-VLM, UniV2X, and CooperNaut, with additional ablations and runtime analysis.

Significance. If the reported results were valid, SEAL would offer a practical and inexpensive way to improve cooperative trajectory planning under adverse weather: the prompt-driven augmentation avoids costly real long-tail data collection, GMSAA is a lightweight modular add-on, and MSCL is a clean extension of contrastive learning. The paper also provides useful ablations and runtime measurements. However, two load-bearing concerns identified in the reader's report are confirmed by the manuscript: the long-tail test images are generated by the same GPT-4o pipeline used for training augmentation, and GMSAA consumes the ground-truth scenario label at inference while the baselines do not. These issues mean that the central claim of robust generalization to real long-tail conditions is not established by the current experiments.

major comments (3)
  1. [Section 4.1 and Section 3.2] The long-tail evaluation set is synthesized with the same GPT-4o prompt-driven transformation pipeline used to create training augmentations. Section 4.1 states that the augmented dataset is used for both training and evaluation, so the snow/fog test cases are drawn from the same distribution the model was trained to fit. This does not measure generalization to real snow or fog. The five image-quality metrics and composite score in Eqs. (2)-(3) and Table 3 assess image fidelity, not downstream planning transfer, and no real adverse-weather benchmark is used. Because the abstract and conclusion make robustness claims under challenging driving conditions, this circular evaluation is load-bearing.
  2. [Section 3.3, Eqs. (6), (8), (11), with Section 4.1] GMSAA is conditioned on the ground-truth scenario label d at inference time. Eq. (6) injects the one-hot vector e_d and the cross-domain penalty gamma_neg(d) into attention logits, Eq. (8) computes the guided weights from e_d, and Eq. (11) selects the scenario-specific feature extractor based on d. Section 4.1 explains that labels are available because every synthetic frame is annotated by construction. The baselines in Tables 5 and 6 do not receive such privileged inputs, so the comparison is between SEAL-plus-oracle and the baselines. No scenario-recognition module, label-noise analysis, or deployment-time estimator is presented. If d is wrong, gamma_neg(d) suppresses the correct scenario logit and the wrong feature extractor is selected, so this is an internal correctness risk, not merely an external generalization issue.
  3. [Section 4.3 and Section 4.5] No error bars, random seeds, or significance tests are reported for any table, and the baseline training protocol is not described. In particular, it is unclear whether V2X-VLM, UniV2X, and CooperNaut were retrained on the augmented snow/fog data with the same splits and hyperparameters, or whether they were evaluated only on the synthetic test frames. The abstract's claim that SEAL 'significantly outperforms' baselines therefore lacks statistical support, and some differences are small or equal in Table 6 (e.g., fog collision rate is 0.2000 for both V2X-VLM and SEAL).
minor comments (5)
  1. [Section 4.1] The paragraph beginning 'To ensure the quality and realism of these generated scenes...' is duplicated verbatim; one copy should be removed.
  2. [Table 5 and Section 2.1] The table labels the baseline as 'UniV2X (2024 AAAI)' while the text and reference list cite Yu et al. (2025); this inconsistency should be corrected.
  3. [Section 4.5.1] The phrase 'multimodel contrastive loss' should read 'multimodal contrastive loss'.
  4. [Section 2.1] There is a typo in 'whichb limits its adaptability'; it should be 'which limits its adaptability'.
  5. [Section 3.3, Eq. (6)] The broadcast semantics of the Hadamard product between the vector beta_self and the one-hot e_d is not fully specified; a short notational clarification would improve reproducibility.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: SEAL's trajectory outputs are learned functions of visual, textual, and scenario inputs, not algebraic rewrites of those inputs.

full rationale

The paper's central claim is an empirical performance comparison on an augmented benchmark, not a derivation from first principles. The GMSAA module does condition on the ground-truth scenario label d at inference (Eqs. 6, 8, 11), and Section 4.1 supplies d at test time; this is a real evaluation-design limitation because it gives SEAL privileged information unavailable at deployment, but it is not circular: the predicted trajectory is still produced by a learned transformer decoder and could be wrong even with perfect d. Similarly, the long-tail training and test sets are both generated by the same GPT-4o prompt pipeline (Sections 3.2 and 4.1), so the benchmark does not independently measure generalization to real snow or fog; however, the reported L2 and collision numbers are empirically measured outputs, not quantities that are equal to the inputs by construction. The self-citations to V2X-VLM are used for the base architecture, contrastive objective, and distillation recipe, but SEAL's advantage is tested against V2X-VLM as a baseline rather than asserted from the citation. No equation in the paper makes the predicted trajectory algebraically identical to the scenario label or to the synthetic transformation, so the claimed improvements do not reduce to the inputs by construction. The concerns raised about oracle labels and same-source synthetic data are validity threats for external deployment, but they are not circularity in the derivation chain.

Assumptions & free parameters 10 free parameters · 4 assumptions · 0 invented entities

The central claim depends on the synthetic data generation assumption, the availability of true scenario labels at inference, and hand-tuned hyperparameters. No new physical entities or forces are introduced. The many hand-set weights and temperatures are free parameters that affect the reported performance and are not justified by an external principle.

free parameters (10)
  • GMSAA self-attention bias vector beta_self = [2.0, 2.5, 2.5]
    Hand-chosen per-scenario bias added to attention logits in Eq. (6); Table 4.
  • Scenario similarity matrix S initialization = 3x3 matrix in Table 4
    Hand-set prior over scenario relations; used in Eq. (7)-(8).
  • Attention temperature tau (GMSAA) = 0.5
    Controls sharpness of scenario attention in Eq. (6); Table 4.
  • Guidance blending coefficient lambda = 0.85
    Blends learned and prior-guided attention in Eq. (9); Table 4.
  • Feature fusion weights = 0.7 scenario context, 0.3 raw context
    Residual fusion weighting in GMSAA; Table 4.
  • MSCL modality temperature tau_mod = 0.07
    Temperature in modality contrastive similarity, Eq. (14); Table 4.
  • MSCL scenario temperature tau_d = 0.1
    Temperature in scenario contrastive similarity, Eq. (17); Table 4.
  • MSCL scenario loss weight lambda_d = 0.3
    Balances modality alignment and scenario discrimination, Eq. (19); Table 4.
  • Overall loss weights alpha and beta = alpha=0.2, beta=0.5
    Weights for contrastive and distillation losses, Eq. (21), Section 4.2.
  • Weather-specific composite score weights = Table 3 (snow: LPIPS 0.30, BRISQUE 0.25, FID 0.20, FADE 0.05, IoU 0.20; fog: 0.20, 0.10, 0.25, 0.30, 0.15)
    Hand-selected weights for the generated-image quality score in Section 3.2.
assumptions (4)
  • domain assumption GPT-4o image generation produces photorealistic snow and fog transforms that preserve road geometry, object positions, and semantics needed for trajectory planning.
    Section 3.2 and Figure 5; no independent validation that the transformed images retain trajectory-relevant geometric fidelity.
  • domain assumption Synthetic long-tail data generated by the prompt-driven pipeline is representative of real-world long-tail scenarios.
    Section 4.1; the test set is generated by the same pipeline, so distribution shift to real weather is untested.
  • ad hoc to paper The ground-truth scenario label d is known and available as input to GMSAA at inference.
    Eq. (6)-(11), Section 3.1; the paper does not state how d is obtained at deployment, yet scenario-split evaluation in Table 6 uses true labels.
  • domain assumption Florence-2 pretrained representations and DAIR-V2X ground-truth trajectories are reliable for the planning task.
    Section 4.2; standard backbone and dataset assumptions adopted without re-validation.

how reviews work

0 comments
Cite this review

Pith. "Pith review of SEAL: Vision-Language Model-Based Safe End-to-End Cooperative Autonomous Driving with Adaptive Long-Tail Modeling." pith.science (2026). https://pith.science/paper/XF73NSMC

@misc{pith2026250621041,
  author       = {Pith},
  title        = {Pith review of: SEAL: Vision-Language Model-Based Safe End-to-End Cooperative Autonomous Driving with Adaptive Long-Tail Modeling},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/XF73NSMC}},
  note         = {Machine review of arXiv:2506.21041}
}
read the original abstract

Autonomous driving technologies face significant safety challenges while operating under rare, diverse, and visually degraded weather scenarios. These challenges become more critical in cooperative settings, where vehicles and infrastructure jointly perceive and reason across complex environments. To address these issues, we propose SEAL, a vision-language model-based framework with adaptive multimodal learning for robust cooperative autonomous driving under long-tail scenarios. SEAL introduces three core innovations: (i) a prompt-driven long-tail scenario generation and evaluation pipeline that leverages foundation models to synthesize realistic long-tail conditions such as snow and fog across vehicle- and infrastructure-side views, enriching training diversity efficiently; (ii) a gated multi-scenario adaptive attention module that modulates the visual stream using scenario priors to recalibrate ambiguous or corrupted features; and (iii) a multi-task scenario-aware contrastive learning objective that improves multimodal alignment and promotes cross-scenario feature separability. Extensive experiments demonstrate that SEAL significantly outperforms existing baselines in reasoning, safety, and planning accuracy under complex, challenging driving conditions, advancing the safety, robustness, and scalability of autonomous driving.

Figures

Figures reproduced from arXiv: 2506.21041 by the authors.

Figure 1
Figure 1. Overview of representative end-to-end autonomous driving pipelines. (a) Traditional pipeline separates perception, prediction, and planning. (b) VLM-based framework integrates semantic reasoning but lacks cross-agent collaboration. (c) UniV2X (Yu et al. (2025)) introduces cooperative planning without semantic alignment. (d) V2X-VLM (You et al. (2024)) introduces multimodal fusion but lacks long-tail robustness. (e) … view at source ↗
Figure 2
Figure 2. Overview of the SEAL framework. The vehicle-side and infrastructure-side views 𝐼𝑣 , 𝐼𝑖 are concatenated and processed by a frozen image encoder 𝐼 to produce visual token embeddings 𝑧. The GMSAA module  injects scenario￾aware priors into 𝑧 based on scenario label 𝑑. A trainable text encoder 𝑇 encodes the scene description 𝐸 into textual embeddings ℎ. The two modalities are fused via a transformer encoder , and de… view at source ↗
Figure 3
Figure 3. Architecture of the Gated Multi-Scenario Adaptive Attention (GMSAA) module. Attention weights are computed from global image features and blended with similarity-guided supervision. The resulting context is refined by scenario￾specific extractors and adaptively gated before fusion with the original token stream. metrics such as FADE and FID receive higher importance. This adjustment is necessary, since adverse condi… view at source ↗
Figures from the paper (5 more)
Figure 4
Figure 4. Figure 4: Multi-Task Scenario-Aware Contrastive Learning (MSCL). Image-text pairs are encoded and optimized with cross-modal similarity. Simultaneously, same-scenario image embeddings are encouraged to cluster, guided by scenario frequency-aware weighting. 3.5. Training Objectiv…
Figure 5
Figure 5. Figure 5: Examples of long-tail data generation and quality assessment. For each pair of original and transformed scenes (snow or fog), we visualize the generated images, scene semantics, edge overlays, and corresponding composite metric scores. These quantitative evaluations va…
Figure 6
Figure 6. Figure 6: Scenario-aware attention learned by GMSAA. Left: attention weight heatmap across domains. Right: adaptation distribution for each input scenario. 4.4. Analysis of Scenario-Aware Adaptation To further understand how SEAL achieves robustness across diverse driving condit…
Figure 7
Figure 7. Figure 7: Latent feature space before and after GMSAA adaptation. Left: original features. Right: adapted features with clearer domain separation. 4.5.1. General Modality and Module Contributions We begin by disabling each key input or architectural module in isolation to examin…
Figure 8
Figure 8. Figure 8: Qualitative visualization of planned trajectories under fog (a), normal (b), and snow (c) scenarios. The top row shows vehicle-side and infrastructure-side views, where bounding boxes in infrastructure-side images highlight the ego vehicle’s position when visible. Bott…

Discussion (0). Sign in to comment.

Forward citations

Cited by 1 Pith paper

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

  1. A Review of Learning-Based Motion Planning: Toward a Data-Driven Optimal Control Approach

    cs.RO 2025-12 conditional novelty 3.0 of 10

    A position/review paper argues data-driven model predictive control is the best route to safe, adaptive, human-like autonomous-driving motion planning, but provides no new derivation or experiment.

Reference graph

Works this paper leans on

20 extracted references · 13 canonical work pages · cited by 1 Pith paper

  1. [1]

    Spatialvlm:Endowingvision-languagemodelswithspatialreasoning capabilities, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp

    Chen,B.,Xu,Z.,Kirmani,S.,Ichter,B.,Sadigh,D.,Guibas,L.,Xia,F.,2024. Spatialvlm:Endowingvision-languagemodelswithspatialreasoning capabilities, in: Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pp. 14455–14465. Chen, Q., Tang, S., Yang, Q., Fu, S.,

  2. [6]

    arXiv preprint arXiv:2410.23262

    Emma: End-to-end multimodal model for autonomous driving. arXiv preprint arXiv:2410.23262 . Li,X.,Liu,J.,Li,J.,Yu,W.,Cao,Z.,Qiu,S.,Hu,J.,Wang,H.,Jiao,X.,2023. Graphstructure-basedimplicitriskreasoningforlong-tailscenarios of automated driving, in: 2023 4th International Conference on Big Data, Artificial Intelligence and Internet of Things Engineering (IC...

  3. [7]

    3373–3382

    Macp: Efficient model adaptation for cooperative perception, in: Proceedings of the IEEE/CVF Winter Conference on Applications of Computer Vision, pp. 3373–3382. Mehra,A.,Mandal,M.,Narang,P.,Chamola,V.,2020. Reviewnet:Afastandresourceoptimizednetworkforenablingsafeautonomousdrivingin hazy weather conditions. IEEE Transactions on Intelligent Transportation...

  4. [8]

    Lot-nuscenes: A virtual long-tail scenario dataset for parallel vision and parallel vehicles, in: 2024 IEEE 4th International Conference on Digital Twins and Parallel Intelligence (DTPI), IEEE. pp. 194–199. Mittal, A., Moorthy, A.K., Bovik, A.C.,

  5. [11]

    Optimizing intersection-over-union in deep neural networks for image segmentation, in: International symposium on visual computing, Springer. pp. 234–244. Snell,J.,Ridgeway,K.,Liao,R.,Roads,B.D.,Mozer,M.C.,Zemel,R.S.,2017. Learningtogenerateimageswithperceptualsimilaritymetrics,in: 2017 IEEE international conference on image processing (ICIP), IEEE. pp. 4...

  6. [15]

    LTDA-Drive: LLMs-guided Generative Models based Long-tail Data Augmentation for Autonomous Driving

    Ltda-drive: Llms-guided generative models based long-tail data augmentation for autonomous driving. arXiv preprint arXiv:2505.18198 . Zang,S.,Ding,M.,Smith,D.,Tyler,P.,Rakotoarivelo,T.,Kaafar,M.A.,2019. Theimpactofadverseweatherconditionsonautonomousvehicles: How rain, snow, fog, and hail affect the performance of a self-driving car. IEEE vehicular techno...

  7. [16]

    Co-http: Cooperative trajectory prediction with heterogeneous graph transformer for autonomous driving, in: 2024 IEEE 27th International Conference on Intelligent Transportation Systems (ITSC), IEEE. pp. 3033–3039. Zhang, X., Zhou, Z., Wang, Z., Ji, Y., Huang, Y., Chen, H.,

  8. [17]

    arXiv preprint arXiv:2502.16589

    Co-mtp: A cooperative trajectory prediction framework with multi-temporal fusion for autonomous driving. arXiv preprint arXiv:2502.16589 . Zhao, S.Z., Xiang, H., Xu, C., Xia, X., Zhou, B., Ma, J.,

Show all 20 references
  1. [18]

    arXiv preprint arXiv:2408.11241

    Coopre: Cooperative pretraining for v2x cooperative perception. arXiv preprint arXiv:2408.11241 . Zhou, R., Huang, H., Lee, J., Huang, X., Chen, J., Zhou, H.,

  2. [19]

    Accident Analysis & Prevention 191, 107218

    Identifying typical pre-crash scenarios based on in-depth crash data with deep embedded clustering for autonomous vehicle safety testing. Accident Analysis & Prevention 191, 107218. Zhou,W.,Cao,Z.,Deng,N.,Liu,X.,Jiang,K.,Yang,D.,2022. Dynamicallyconservativeself-drivingplanner...

  3. [20]

    arXiv preprint arXiv:2503.23463

    Opendrivevla: Towards end-to-end autonomous driving with large vision language action model. arXiv preprint arXiv:2503.23463 . Junwei You et al.:Preprint submitted to Elsevier Page 21 of 21

  4. [936]

    Obukhov,A.,Krasnyanskiy,M.,2020. Qualityassessmentmethodforganbasedonmodifiedmetricsinceptionscoreandfréchetinceptiondistance, in: Software Engineering Perspectives in Intelligent Systems: Proceedings of 4th Computational Methods in Systems and Software 2020, Vol. 1 4, Springe...

  5. [2011]

    Blind/referenceless image spatial quality evaluator, in: 2011 conference record of the forty fifth asilomar conference on signals, systems and computers (ASILOMAR), IEEE. pp. 723–727. Mo, Y., Vijay, R., Rufus, R., Boer, N.d., Kim, J., Yu, M.,

  6. [2016]

    IEEE Intelligent Transportation Systems Magazine 8, 8–22

    Cooperative maneuver planning for cooperative driving. IEEE Intelligent Transportation Systems Magazine 8, 8–22. Feng,B.,Mei,Z.,Li,B.,Ost,J.,Girgis,R.,Majumdar,A.,Heide,F.,2025. Verdi:Vlm-embeddedreasoningforautonomousdriving. arXivpreprint arXiv:2505.15925 . Feng, C., Xu, Z.,...

  7. [2019]

    Cooper: Cooperative perception for connected autonomous vehicles based on 3d point clouds, in: 2019 IEEE 39th International Conference on Distributed Computing Systems (ICDCS), IEEE. pp. 514–524. Chen, Y., Veer, S., Karkus, P., Pavone, M.,

  8. [2020]

    Digital Communications and Networks 6, 444–451

    Generating routes for autonomous driving in vehicle-to-infrastructure communications. Digital Communications and Networks 6, 444–451. Yao,H.,Li,P.,Jin,B.,Zheng,Y.,Liu,A.,Mu,L.,Su,Q.,Zhang,Q.,Chen,Y.,Li,P.,2025. Lilodriver:Alifelonglearningframeworkforclosed-loop motion plannin...

  9. [2022]

    IEEE Transactions on Intelligent Vehicles 7, 286–299

    Parallel vision for long-tail regularization: Initial results from ivfc autonomous driving testing. IEEE Transactions on Intelligent Vehicles 7, 286–299. Wang,T.H.,Manivasagam,S.,Liang,M.,Yang,B.,Zeng,W.,Urtasun,R.,2020. V2vnet:Vehicle-to-vehiclecommunicationforjointperception...

  10. [2023]

    Vehicle-to-infrastructure communication for real-time object detection in autonomous driving, in: 2023 18th Wireless On-Demand Network Systems and Services Conference (WONS), IEEE. pp. 40–46. Huang, Z., Sheng, Z., Qu, Y., You, J., Chen, S.,

  11. [2024]

    arXiv preprint arXiv:2412.15544

    Vlm-rl: A unified vision language models and reinforcement learning framework for safe autonomous driving. arXiv preprint arXiv:2412.15544 . Hwang, J.J., Xu, R., Lin, H., Hung, W.C., Ji, J., Choi, K., Huang, D., He, T., Covington, P., Sapp, B., et al.,

  12. [2025]

    1001–1009

    Openemma: Open-source multimodal model for end-to-end autonomous driving, in: Proceedings of the Winter Conference on Applications of Computer Vision, pp. 1001–1009. Xu,H.,Bian,Y.,Li,Y.,Qin,H.,Zhou,H.,Chang,F.,Wang,S.,Ye,Q.,2025. Learnableoperationaldesignconditionmonitorforfa...

Pith tools

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