Pith. sign in

REVIEW 4 major objections 5 minor 21 references

Multimodal Framework for Explainable Autonomous Driving: Integrating Video, Sensor, and Textual Data for Enhanced Decision-Making and Transparency

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 A multimodal framework that fuses video, sensor, and text features predicts driving actions at 92.5% accuracy and generates natural-language explanations with BLEU-4 0.75 on nuScenes.

desk verdict The evaluation measures the wrong thing: the model receives the human explanation as an input, and the diagnostic figures are admitted simulations. read the letter →

arxiv 2507.07938 v1 pith:R7SNVUHT submitted 2025-07-10 cs.MM

Abolfazl Zarghani , Amirhossein Ebrahimi , Amir Malekesfandiari This is my paper
classification cs.MM
keywords multimodalfusionautonomousdrivingexplainableAIactionpredictionnaturallanguageexplanationVideoMAEBERTBART
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 claims that a single multimodal network can both predict a driver's next action and explain it in natural language, and that doing both does not sacrifice accuracy. The framework combines spatiotemporal video features from VideoMAE, a small sensor-fusion network reading speed and GPS, and BERT text embeddings; these are concatenated and fed to a fused vector that drives an action classifier and a BART-based explanation generator. On the nuScenes and BDD-X urban driving datasets, the paper reports 92.5% and 91.3% action accuracy and BLEU-4 scores of 0.75 and 0.72, beating unimodal baselines and prior multimodal models. The significance, if the numbers hold, is that late fusion of three modalities plus a pretrained language decoder gives both transparent reasoning and state-of-the-art performance in one architecture.

What carries the argument

The load-bearing mechanism is the multimodal fusion layer and the two-headed decoder. A 1664-dimensional vector (video 768 + sensor 128 + text 768) goes through one fully connected layer with ReLU to 768 dimensions; that fused vector is the input to a classification head and is linearly projected and added to BART's decoder input embeddings so that generated explanations are conditioned on the combined context. Inference uses beam search with five beams. This design lets a single representation serve both discriminative action prediction and generative explanation, which is why the paper attributes the simultaneous accuracy and BLEU gains to the fusion layer rather than to any single encoder.

What would settle it

Run the released code and pretrained weights on the actual BDD-X and nuScenes test splits and check whether the action accuracy and BLEU-4 reproduce the reported 92.5%/91.3% and 0.75/0.72; if the confusion matrix or attention maps turn out to be simulated rather than computed from model outputs, or the numbers differ materially, the claim is falsified.

Watch

Extended reading notes

Core claim

The paper's central claim is that a fused representation built by concatenating a 768-dimensional VideoMAE video feature, a 128-dimensional sensor feature, and a 768-dimensional BERT text feature, then projecting the 1664-dimensional concatenation to 768, supports two heads at once: a linear action classifier and a decoder conditioning step that adds the projected fusion to BART's input embeddings. Trained for five epochs with an unweighted sum of cross-entropy losses, the model is reported to reach 92.5% action prediction accuracy and 0.75 BLEU-4 on nuScenes, and 91.3% and 0.72 on BDD-X, outperforming DeepDriving, SensorFusionTransformer, AttentionMultimodal, and RuleBasedXAI baselines. The paper's framing is that explainability and performance are jointly achievable when each modality is encoded by a strong pretrained transformer and fused late, rather than after early feature extraction.

Load-bearing premise

The entire evaluation rests on the assumption that the reported accuracy, BLEU, confusion-matrix, and human-evaluation numbers come from real model outputs on held-out data; the paper explicitly says the attention maps and action distribution are simulated, so if the headline numbers are similarly simulated rather than measured, the central claim is unverified.

Editorial extensions

If this is right

  • If the reported results hold, late fusion of video, sensor, and text suffices for both accurate action classification and fluent explanation generation on urban driving datasets.
  • The 120 ms per-sample inference time places the architecture near real-time, though the paper notes that 50 ms response budgets on highways would still require optimization.
  • Ablation results indicate that video contributes the most to performance, with removal dropping accuracy to 75.6%, while sensor and text each add smaller but measurable gains.
  • The 231-million-parameter model would need pruning or distillation before running on embedded vehicle hardware, a direction the paper explicitly flags for future work.

Reading between the lines

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

  • Because the paper's diagnostic visualizations (attention maps and the action distribution) are admitted to be simulated, the ablation numbers should be treated with caution until the model is rerun with real attention and real action statistics.
  • A natural next test would replace the single projection injected into BART's decoder with cross-attention between the fused features and the token history, letting the explanation attend to specific video regions or sensor readings.
  • BLEU-4 scores reward surface n-gram overlap, so the reported 0.75 may underestimate semantic quality; the paper's own human ratings (4.4-4.7 out of 5) are the more meaningful evidence for explanation usefulness.
  • If the fusion-plus-decoder pattern generalizes, the same recipe could be applied to other safety-critical domains that need both a class decision and a justification, such as clinical alerts or industrial control.
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 proposes a multimodal driving framework that integrates VideoMAE video features, a small sensor-fusion MLP, and BERT text features, then fuses them to predict driving actions and to condition BART for explanation generation. The authors report 92.5% action accuracy and a 0.75 BLEU-4 on nuScenes, with 91.3% and 0.72 on BDD-X, plus ablation, human-evaluation, and qualitative results. The manuscript's own limitation text (Sections 5.3 and 5.5) states that the attention maps, the action distribution, and the attention visualizations in Figures 3, 5, and 6 are simulated rather than produced by the trained model. The evaluation as described also feeds the human-written explanation text into the model as an input modality, which is the central flaw examined in this report.

Significance. If the claimed numbers were obtained from a clean, text-free evaluation protocol, the paper would provide a useful data point on late fusion of video, sensor, and language features for explainable driving. The high-level architecture is plausible, and the inclusion of ablations and a clearly stated limitations section is a positive feature. However, the central empirical claims are not supported as described: the text modality at inference is the ground-truth explanation, so the BLEU-4 score measures reconstruction of an input rather than explanation generation, and the action accuracy can be inflated by reading the answer from the text branch. In addition, the explainability evidence is explicitly simulated. These are load-bearing problems rather than presentation issues, so the paper's main conclusions cannot be accepted on the evidence provided.

major comments (4)
  1. [§3.1, §3.2.3, §3.2.4, §4.4] The model receives the ground-truth explanation as an input at inference. Section 3.1 defines the textual modality with examples such as "Slow down due to pedestrian crossing"; Section 3.2.3 passes these descriptions through BERT; Section 3.2.4 sends the resulting vector into both the action classifier and the BART decoder; and Section 4.4 uses the same kind of text as the ground truth for explanation quality. The reported BLEU-4 of 0.75 is therefore a reconstruction score of an input, not a measure of explanations generated from video and sensor context, and the claimed 92.5% action accuracy can be inflated by the action classifier reading the answer from the text branch. Table 3 corroborates this dependency: removing text makes BLEU-4 N/A and reduces accuracy by only about four points, which indicates that the text branch is the source of the explanation and partly the source of the action prediction.
  2. [§5.3, §5.5, Figs. 3, 5, 6] The paper states explicitly in Section 5.3 that "the simulated nature of the attention maps and action distribution (Fig. 5, Fig. 6, Fig. 3) highlights the need for real data validation in future work," and Section 5.5 calls these figures "simulated" and "illustrative." Since these figures are the main qualitative evidence for the framework's explainability, and since Section 4.6 relies on the simulated action distribution to explain the confusion matrix, the interpretability contribution is not supported by actual model outputs. This is a load-bearing gap, not a minor caveat.
  3. [§3.1, §4.1, §4.6] The dataset description is internally inconsistent. Section 3.1 states that BDD-X samples include a 40-second video clip, sensor readings (speed, GPS latitude, and longitude), and textual descriptions of the driving context; however, BDD-X is a video-and-annotation dataset and does not provide the sensor readings described. Section 3.1 also states nuScenes has 1,000 scenes, while Section 4.6 refers to "16,890 training samples in nuScenes" and gives class counts such as 4,223 samples for Stop and Decelerate, which appear to come from the BDD-X split. These inconsistencies make it impossible to determine which data were actually used and how the 80:20 split was applied.
  4. [§3.3, §4.2, Tables 1–3] No standard deviations, confidence intervals, random-seed information, or multiple-run statistics are reported for any accuracy or BLEU-4 number, and the training description is limited to five epochs with no learning curves or per-task losses. More importantly, the baseline comparisons in Tables 1 and 2 are not apples-to-apples: the proposed model receives the ground-truth text as an input, while the unimodal and partial baselines do not. Even if all reported numbers are honestly computed, the comparison does not support the claim of state-of-the-art performance.
minor comments (5)
  1. [§3.1] The paper says each BDD-X sample is a "40-second video clip (16 frames at 224×224 resolution)" but does not specify how the 16 frames were sampled from the 40-second clip; please clarify the temporal sampling rate and whether the same protocol was used for nuScenes.
  2. [§3.4] The code availability statement says code and pretrained models "will be made publicly available," but no repository link or release plan is provided; the experimental setup is therefore not reproducible from the manuscript alone.
  3. [§4.2.2, §4.2.3] Figures 2 and 4 are referenced as showing the confusion matrix and human-evaluation results, but the actual figure content is not included in the provided text, so the reader cannot verify the stated numbers such as 93.5% for Stop or the human-evaluation scores.
  4. [§4.5] The error-analysis percentages (3%, 2%, 1% of test cases) are presented without explaining how the error categories were identified or annotated; please describe the error-analysis procedure.
  5. [§5.3] The parameter count is given as 231 million without a breakdown across VideoMAE, BERT, BART, and the small fusion MLP; a per-module count would help the reader assess the stated computational-complexity limitation.

Circularity Check

1 steps flagged · score 6.0 of 10

Explanation scores and action accuracy are partially forced because the ground-truth explanation text is also an input modality.

  1. self definitional [Sections 3.1, 3.2.3, 3.2.4, 4.1]
    "Each sample includes a 40-second video clip (16 frames at 224 ×224 resolution), sensor readings (speed, GPS latitude, and longitude), and textual descriptions of the driving context (e.g., “Slow down due to pedestrian crossing”). ... BLEU-4 Score: ... measuring n-gram overlap (up to 4-grams) between generated explanations and ground-truth textual descriptions"

    The same textual descriptions are (a) encoded by BERT as an input modality in Sec. 3.2.3, (b) injected into the fusion layer and BART decoder in Sec. 3.2.4, and (c) used as the ground-truth reference for BLEU-4 in Sec. 4.1 and for the qualitative examples in Table 4. The “generated” explanation is therefore a reconstruction or paraphrase of its own input text. BLEU-4 = 0.75 measures text-to-text copying from the BERT branch, not generation from video and sensor context. The ablation in Table 3 confirms the dependency: removing text makes explanation generation undefined (BLEU-4 = N/A) while accuracy remains 88.3%.

full rationale

The paper is an empirical systems paper, not a formal derivation, so the classic derivation-chain circularity modes are mostly absent: there are no self-citations carrying a uniqueness theorem, no fitted parameter renamed as a prediction, and no known result re-packaged as new. The substantive circularity is in the evaluation design. Section 3.1 defines the textual modality as “textual descriptions of the driving context,” with examples such as “Slow down due to pedestrian crossing”; Section 3.2.3 routes those descriptions through BERT; Section 3.2.4 sends the resulting vector into both the action head and the BART decoder; Section 4.1 then scores explanation quality with BLEU-4 against “ground-truth textual descriptions.” Thus the target explanation is supplied as an input, making the BLEU-4 score a reconstruction metric. The ablation study (Table 3) makes the leakage visible: without text, explanation generation is impossible (N/A) and accuracy drops only about four points, confirming that the text branch carries the explanatory answer. The paper also concedes in Sections 5.3 and 5.5 that the attention maps and action distribution are simulated and that future work will replace them with real data; that is a serious evidence-quality weakness, but it is not itself circular. Because the central quantitative claims are partially forced by the ground-truth text being an input, the score is 6 rather than 0–2, but it is not 8–10 because the video and sensor modalities still contribute measurably to action accuracy in the ablations.

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

The ledger shows that the paper's central evaluation rests on unverified premises about the BDD-X data format and on simulated diagnostic figures. No code or data is released, so the claimed metrics cannot be checked. The free parameters are standard architectural choices, but without released artifacts they are not independently testable.

free parameters (4)
  • Fusion layer width (1664->768 with ReLU) = 1664 input, 768 hidden
    Chosen by grid search on the validation set (Section 3.4). The central fusion design depends on these dimensions.
  • Sensor MLP layer sizes (3->64->128) = 64 then 128
    Chosen by the authors for the sensor branch; no derivation or comparison is provided.
  • Text truncation length (50 tokens) = 50
    Chosen to balance efficiency and richness (Section 3.1), but this is a hand-set preprocessing choice.
  • Action class prevalence weights = Stop 25%, Decelerate 25%, Accelerate 20%, Turns 15% each
    The simulated action distribution in Fig. 3 is used in Section 4.6 to explain the accuracy numbers; if the distribution is simulated, the class-prior explanation is itself an invented quantity.
assumptions (3)
  • domain assumption The BDD-X dataset contains 21,113 samples with paired video, speed, GPS, and textual descriptions.
    Section 3.1 describes BDD-X with 6,970 videos and 21,113 samples including sensor and text data. This does not match the public BDD-X/BDD100K distribution, and the cited reference [14] is the BDD100K paper. The evaluation numbers depend on this premise.
  • domain assumption VideoMAE, BERT, and BART pretrained checkpoints provide suitable initializations for the driving tasks.
    The model is built on pretrained VideoMAE (MCG-NJU/videomae-base), BERT (bert-base-uncased), and BART (bart-base). This is a standard transfer-learning assumption, but no fine-tuning details are released.
  • domain assumption The reported loss, accuracy, and BLEU numbers are computed from real model outputs on the described test splits.
    Section 5.3 and 5.5 state that attention maps and action distribution figures are simulated; no statement confirms that the scalar metrics are real. The correctness of the headline numbers rests on this unverified premise.
invented entities (2)
  • Simulated action distribution (Stop 25%, Decelerate 25%, Accelerate 20%, Turn Left/Right 15% each)
    purpose: Used in Fig. 3 and Section 4.6 to explain why Stop and Decelerate have high accuracy.
    The paper explicitly admits the figure is simulated (Section 5.3, 5.5). It is presented as the dataset context for interpreting performance, but it is not measured from nuScenes.
  • Simulated attention maps for VideoMAE and BERT (Figs. 5 and 6)
    purpose: Presented as evidence of where the model focuses when generating explanations.
    The authors state these are simulated and plan to replace them with real data in future revisions. As presented, they do not provide evidence about the actual model.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Multimodal Framework for Explainable Autonomous Driving: Integrating Video, Sensor, and Textual Data for Enhanced Decision-Making and Transparency." pith.science (2026). https://pith.science/paper/R7SNVUHT

@misc{pith2026250707938,
  author       = {Pith},
  title        = {Pith review of: Multimodal Framework for Explainable Autonomous Driving: Integrating Video, Sensor, and Textual Data for Enhanced Decision-Making and Transparency},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/R7SNVUHT}},
  note         = {Machine review of arXiv:2507.07938}
}
read the original abstract

Autonomous vehicles (AVs) are poised to redefine transportation by enhancing road safety, minimizing human error, and optimizing traffic efficiency. The success of AVs depends on their ability to interpret complex, dynamic environments through diverse data sources, including video streams, sensor measurements, and contextual textual information. However, seamlessly integrating these multimodal inputs and ensuring transparency in AI-driven decisions remain formidable challenges. This study introduces a novel multimodal framework that synergistically combines video, sensor, and textual data to predict driving actions while generating human-readable explanations, fostering trust and regulatory compliance. By leveraging VideoMAE for spatiotemporal video analysis, a custom sensor fusion module for real-time data processing, and BERT for textual comprehension, our approach achieves robust decision-making and interpretable outputs. Evaluated on the BDD-X (21113 samples) and nuScenes (1000 scenes) datasets, our model reduces training loss from 5.7231 to 0.0187 over five epochs, attaining an action prediction accuracy of 92.5% and a BLEU-4 score of 0.75 for explanation quality, outperforming state-of-the-art methods. Ablation studies confirm the critical role of each modality, while qualitative analyses and human evaluations highlight the model's ability to produce contextually rich, user-friendly explanations. These advancements underscore the transformative potential of multimodal integration and explainability in building safe, transparent, and trustworthy AV systems, paving the way for broader societal adoption of autonomous driving technologies.

Figures

Figures reproduced from arXiv: 2507.07938 by the authors.

Figure 1
Figure 1. Architecture of the proposed multimodal framework. [PITH_FULL_IMAGE:figures/full_fig_p005_1.png] view at source ↗
Figure 2
Figure 2. Normalized confusion matrix for ac￾tion prediction (%), showing high accuracy (e.g., 93.5% for Stop) and minor misclassifications, con￾sistent with the 92.5% accuracy on nuScenes. at 4.7 for Relevance, 4.5 for Clarity, and 4.4 for Helpfulness (on a 0–5 scale), closely approaching ground truth scores (4.8, 4.6, 4.7). These scores were obtained from 50 human evaluators assess￾ing 100 randomly selected explanations, en… view at source ↗
Figure 4
Figure 4. Human evaluation scores for explanation [PITH_FULL_IMAGE:figures/full_fig_p008_4.png] view at source ↗
Figures from the paper (1 more)
Figure 6
Figure 6. Figure 6: Simulated attention map for BERT on a 50-token explanation, highlighting early tokens (0–10) for key phrases, aligning with high BLEU￾4 scores. 4.6 Dataset Context The simulated distribution of driving actions, as shown in [PITH_FULL_IMAGE:figures/full_fig_p009_6.png]

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

21 extracted references · 7 canonical work pages

  1. [1]

    A survey of autonomous driving: Common practices and emerging technolo- gies,

    E. Yurtsever, J. Lambert, A. Carballo, and K. Takeda, “A survey of autonomous driving: Common practices and emerging technolo- gies,” IEEE Access , vol. 8, pp. 58443–58469, 2020, doi: 10.1109/ACCESS.2020.2983149

    arXiv 2020

  2. [2]

    Safety-enhanced autonomous driving using interpretable sensor fusion transformer,

    H. Shao, L. Wang, R. Chen, H. Li, and Y. Liu, “Safety-enhanced autonomous driving using interpretable sensor fusion transformer,” in Proc. Conf. Robot Learn. (CoRL) , pp. 726–737, 2023, doi: 10.48550/arXiv.2310.12345

    arXiv doi:10.48550/arxiv.2310.12345 2023

  3. [3]

    Planning-oriented autonomous driving,

    Y. Hu, J. Yang, L. Chen, K. Li, C. Sima, X. Zhu, S. Chai, S. Du, T. Lin, W. Wang, L. Lu, X. Jia, Q. Liu, J. Dai, Y. Qiao, and H. Li, “Planning-oriented autonomous driving,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pp. 17853–17862, 2023, doi: 10.1109/CVPR52729.2023.01711

    arXiv 2023

  4. [4]

    Deep multi-modal object detection and semantic segmentation for au- tonomous driving,

    D. Feng, C. Haase-Sch ¨utz, L. Rosenbaum, H. Hertlein, C. Glaeser, F. Timm, W. Wiesbeck, and K. Dietmayer, “Deep multi-modal object detection and semantic segmentation for au- tonomous driving,” IEEE T rans. Intell. T ransp. Syst., vol. 22, no. 3, pp. 1341–1360, 2021, doi: 10.1109/TITS.2020.2970884

    work page arXiv 2021

  5. [5]

    Ex- plainable artificial intelligence (XAI),

    S. Ali, M. D. Abeyratne, M. A. Ediris- inghe, A. M. Ali, and M. Zahidh, “Ex- plainable artificial intelligence (XAI),” Inf. Fusion, vol. 99, p. 101805, 2023, doi: 10.1016/j.inffus.2023.101805

    arXiv 2023

  6. [6]

    Why did the AI make that decision?,

    J. Dong, S. Huang, and C. Lv, “Why did the AI make that decision?,” T ransp. Res. Part C: Emerg. T echnol., vol. 158, p. 104465, 2024, doi: 10.1016/j.trc.2023.104465

    work page arXiv 2024

  7. [7]

    Ex- plainable artificial intelligence (XAI),

    A. B. Arrieta, N. D ´ıaz-Rodr´ıguez, J. Del Ser, A. Bennetot, S. Tabik, A. Barbado, S. Garc´ıa, S. Gil-L ´opez, D. Molina, R. Ben- jamins, R. Chatila, and F. Herrera, “Ex- plainable artificial intelligence (XAI),” Inf. Fusion, vol. 58, pp. 82–115, 2020, doi: 10.1016/j.inffus.2019.12.012

    doi:10.1016/j.inffus.2019.12.012 2020

  8. [8]

    Interpretable autonomous driving: A sur- vey of recent advances,

    S. Li, Y. Yang, J. Zhang, and X. Wang, “Interpretable autonomous driving: A sur- vey of recent advances,” IEEE T rans. Intell. Veh., vol. 8, no. 1, pp. 123–135, 2023, doi: 10.1109/TIV .2022.3214567

    arXiv 2023

Show all 21 references
  1. [9]

    Attention- based multimodal framework for au- tonomous driving,

    J. Kim, J. Park, and Y. Lee, “Attention- based multimodal framework for au- tonomous driving,” IEEE T rans. Robot. , vol. 38, no. 4, pp. 2100–2115, 2022, doi: 10.1109/TRO.2022.3156789

    2022

  2. [10]

    DeepDriving: Learning affordance for direct perception in autonomous driving,

    Y. Chen, H. Hu, and G. Gong, “DeepDriving: Learning affordance for direct perception in autonomous driving,” in Proc. IEEE Int. Conf. Comput. Vis. (ICCV) , pp. 2722–2730, 2015, doi: 10.1109/ICCV .2015.312

    2015 doi

  3. [11]

    CARLA: An open urban driving simulator,

    A. Dosovitskiy, G. Ros, F. Codevilla, A. Lopez, and V . Koltun, “CARLA: An open urban driving simulator,” in Proc. Conf. Robot Learn. (CoRL) , pp. 1–16, 2017, doi: 10.48550/arXiv.1711.03938

    2017 doi arXiv

  4. [12]

    VideoMAE: Masked autoencoders are data- efficient learners for self-supervised video pre-training,

    Z. Tong, Y. Song, J. Wang, and L. Wang, “VideoMAE: Masked autoencoders are data- efficient learners for self-supervised video pre-training,” in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS) , pp. 10078–10093, 2022, doi: 10.48550/arXiv.2203.12602

    2022 doi arXiv

  5. [13]

    BERT: Pre-training of deep bidi- rectional transformers for language under- standing,

    J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, “BERT: Pre-training of deep bidi- rectional transformers for language under- standing,” in Proc. Conf. North Amer. Chapter Assoc. Comput. Linguistics: Human Lang. T ech- nol. (NAACL-HLT), pp. 4171–4186, 2019, doi: 10.18653/v1/N19-1423

    2019 doi

  6. [15]

    nuScenes: A mul- timodal dataset for autonomous driving,

    H. Caesar, V . Bankiti, A. H. Lang, S. Vora, V . E. Liong, Q. Xu, A. Krishnan, Y. Pan, G. Baldan, and O. Beijbom, “nuScenes: A mul- timodal dataset for autonomous driving,” in Proc. IEEE/CVF Conf. Comput. Vis. Pattern Recognit. (CVPR), pp. 11621–11631, 2020, doi: 10.1109/CVPR4...

    2020

  7. [16]

    BART: Denoising sequence- to-sequence pre-training for natural lan- guage generation, translation, and com- prehension,

    M. Lewis, Y. Liu, N. Goyal, M. Ghazvinine- jad, A. Mohamed, O. Levy, V . Stoyanov, and L. Zettlemoyer, “BART: Denoising sequence- to-sequence pre-training for natural lan- guage generation, translation, and com- prehension,” in Proc. Assoc. Comput. Lin- guistics (ACL) , pp. 78...

    2020 doi

  8. [17]

    Language-augmented Bird’s-eye View Maps for autonomous driving,

    D. K. Dewangan, S. Sahu, R. R. Sahay, and A. K. Sahoo, “Language-augmented Bird’s-eye View Maps for autonomous driving,” IEEE Robot. Autom. Lett. , vol. 8, no. 5, pp. 2900– 2907, 2023, doi: 10.1109/LRA.2023.3264589

    2023

  9. [18]

    Vi- sual question answering and natural lan- guage explanations for autonomous driv- ing,

    H. Xu, Y. Zhang, and Q. Wu, “Vi- sual question answering and natural lan- guage explanations for autonomous driv- ing,” IEEE T rans. Intell. T ransp. Syst. , vol. 24, no. 8, pp. 8900–8912, 2023, doi: 10.1109/TITS.2023.3256789

    2023

  10. [19]

    LINGO-1: A vision- language-action model for real-time driving explanations,

    Wayve Team, “LINGO-1: A vision- language-action model for real-time driving explanations,” arXiv preprint , 2023, doi: 10.48550/arXiv.2309.12345

    2023 doi arXiv

  11. [20]

    S. Bird, E. Klein, and E. Loper, Natural Language Processing with Python , 1st ed. Se- bastopol, CA, USA: O’Reilly Media, 2009

    2009

  12. [21]

    ROUGE: A package for automatic evaluation of summaries,

    C.-Y. Lin, “ROUGE: A package for automatic evaluation of summaries,” in Proc. Workshop T ext Summarization Branches Out , pp. 74–81, 2004

    2004

  13. [22]

    METEOR: An automatic metric for MT evaluation with im- proved correlation with human judgments,

    S. Banerjee and A. Lavie, “METEOR: An automatic metric for MT evaluation with im- proved correlation with human judgments,” in Proc. ACL Workshop Intrinsic Extrinsic Eval. Measures Mach. T ransl. Summarization , pp. 65–72, 2005

    2005

Pith tools

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