Pith. sign in

REVIEW 3 major objections 2 minor 41 references

Planning-aligned conditional VQ-VAE compresses long driving context tokens while retaining decision-critical information, improving success rates by more than 6 percent under fixed budgets.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · grok-4.3

2026-06-27 21:33 UTC pith:QTHJUBP4

load-bearing objection COMPACT-VA adds a planning-intent latent to conditional VQ-VAE compression for driving tokens and reports closed-loop gains, but the abstract gives no numbers on how well the prior actually predicts that latent. the 3 major comments →

arxiv 2606.07464 v1 pith:QTHJUBP4 submitted 2026-06-05 cs.RO cs.AIcs.CV

Planning-aligned Token Compression for Long-Context Autonomous Driving

Zhixuan Liang , Yuxiao Chen , Yurong You , Peter Karkus , Wenhao Ding , Boyi Li , Alexander Popov , Yan Wang
show 7 more authors
Maximilian Igl Yiming Li Danfei Xu Nikolai Smolyanskiy Boris Ivanovic Ping Luo Marco Pavone
This is my paper
classification cs.RO cs.AIcs.CV
keywords token compressionautonomous drivingvision-action modelsconditional VQ-VAEplanning intentlong contextclosed-loop evaluationworking memory
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

The paper develops a token compression approach for vision-action models in autonomous driving that keeps extended temporal context within real-time limits. It builds a conditional VQ-VAE where compression is guided by a planning intent latent that the posterior encoder extracts from future trajectories during training and the prior encoder learns to predict from the compressed observations. This coupling feeds the policy with both the compressed memory and the predicted latent so that end-to-end optimization preserves the information needed for correct behaviors such as stop, yield, or proceed. Evaluation on high-signal dynamic scenarios reports over 6 percent higher success rates to 68.3 percent together with consistent metric gains, and closed-loop tests show 3.3 times speedup and 2.7 times memory reduction while maintaining general performance.

Core claim

COMPACT-VA is a planning-aligned working memory framework built on conditional VQ-VAE that compresses extended context into bounded representations conditioned on both historical trajectory and a learned planning intent. The posterior encoder distills the intent from future trajectories at training time while the prior encoder predicts it from compressed observations; the compressed memory concatenated with this predicted latent then drives the policy under end-to-end optimization, enabling planning that retains decision-critical information.

What carries the argument

Conditional VQ-VAE whose compression is conditioned on planning intent distilled from future trajectories, with the prior encoder trained to predict that latent from the compressed observations.

Load-bearing premise

The training procedure that distills planning intent from future trajectories produces a latent the prior encoder can reliably predict from compressed observations without introducing systematic biases that degrade closed-loop behavior.

What would settle it

A closed-loop driving test in which the prior encoder cannot accurately reconstruct the planning-intent latent from the compressed tokens and the resulting policy exhibits lower success rates than the uncompressed baseline on scenarios that require historical context.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • Under comparable token budgets the method reaches 68.3 percent success rate with gains across behavioral metrics.
  • Closed-loop evaluation confirms maintained general driving performance.
  • Processing achieves 3.3 times speedup and 2.7 times memory reduction relative to uncompressed input.
  • Ablation studies confirm that the planning-aligned coupling is responsible for the observed gains.

Where Pith is reading between the lines

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

  • The same conditioning mechanism could support still longer context windows before token budgets are exceeded.
  • Analogous planning-intent alignment might improve compression in other long-horizon sequential decision domains.
  • Explicit measurement of intent-prediction accuracy on held-out scenarios would quantify how much decision information survives compression.

Editorial analysis

A structured set of objections, weighed in public.

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

Referee Report

3 major / 2 minor

Summary. The manuscript proposes COMPACT-VA, a planning-aligned token compression framework for long-context vision-action models in autonomous driving. It employs a conditional VQ-VAE where the posterior distills a planning-intent latent from future trajectories during training, the prior predicts this latent from compressed observation tokens at inference, and the policy receives the compressed memory concatenated with the predicted latent. The central empirical claim is a success-rate improvement to 68.3% (>6% gain) under comparable token budgets on high-signal dynamic scenarios, with ablations supporting the coupling and closed-loop results showing 3.3× speedup and 2.7× memory reduction while preserving general driving performance.

Significance. If the reported gains are shown to be robust with explicit baselines, token budgets, and controls for prior-posterior fidelity, the work would offer a practical architectural contribution to efficient long-horizon context handling in end-to-end driving policies without requiring backbone changes. The emphasis on closed-loop evaluation and behavioral metrics tailored to context-critical maneuvers (stop/yield) is a constructive element.

major comments (3)
  1. [Abstract / Evaluation] Abstract and evaluation section: the central claim of a >6% success-rate improvement to 68.3% under comparable token budgets is presented without naming the baseline methods, reporting exact token counts per comparison, dataset statistics, or statistical significance tests; this leaves the empirical support for the planning-aligned compression benefit unverifiable from the provided text.
  2. [Method / Ablations] Method and ablation sections: the architecture relies on the prior encoder accurately reconstructing the planning-intent latent from compressed observations alone, yet no quantitative metric (e.g., reconstruction error, KL divergence, or distribution-shift measure between posterior and prior) is reported to validate that decision-critical cues are retained; the ablations are described only at a high level as confirming “planning-aligned coupling effectiveness.”
  3. [Closed-loop evaluation] Closed-loop evaluation: the claim that COMPACT-VA “maintained general driving performance” with the stated speedups is load-bearing for the practical utility argument, but the text provides no details on the closed-loop simulator, scenario distribution, or comparison against uncompressed and alternative compression baselines at matched latency/memory.
minor comments (2)
  1. [Method] Notation for the planning-intent latent dimension and VQ-VAE codebook size should be introduced explicitly with symbols rather than prose only.
  2. [Figures] Figure captions and axis labels for any token-budget or success-rate plots should include the precise token counts and baseline names used.

Simulated Author's Rebuttal

3 responses · 0 unresolved

We thank the referee for the constructive feedback, which identifies key areas for improving the clarity and verifiability of our empirical claims. We address each major comment below and will revise the manuscript to incorporate additional details where feasible.

read point-by-point responses

  1. Referee: [Abstract / Evaluation] Abstract and evaluation section: the central claim of a >6% success-rate improvement to 68.3% under comparable token budgets is presented without naming the baseline methods, reporting exact token counts per comparison, dataset statistics, or statistical significance tests; this leaves the empirical support for the planning-aligned compression benefit unverifiable from the provided text.

    Authors: We agree that the abstract and evaluation sections would be strengthened by explicit details. The comparisons are against standard token compression baselines and uncompressed models on the high-signal dynamic scenarios, but we will revise to name the baselines explicitly, report exact token counts (e.g., under 512-token and 1024-token budgets), include dataset statistics (scenario counts and characteristics), and add statistical significance tests from multiple runs with p-values. revision: yes

  2. Referee: [Method / Ablations] Method and ablation sections: the architecture relies on the prior encoder accurately reconstructing the planning-intent latent from compressed observations alone, yet no quantitative metric (e.g., reconstruction error, KL divergence, or distribution-shift measure between posterior and prior) is reported to validate that decision-critical cues are retained; the ablations are described only at a high level as confirming “planning-aligned coupling effectiveness.”

    Authors: We acknowledge the value of quantitative metrics to directly validate prior-posterior fidelity. While the existing ablations show performance gains from the coupling, we will add metrics such as KL divergence between the prior and posterior distributions and latent reconstruction error in the revised ablation section to provide explicit evidence that decision-critical cues are retained. revision: yes

  3. Referee: [Closed-loop evaluation] Closed-loop evaluation: the claim that COMPACT-VA “maintained general driving performance” with the stated speedups is load-bearing for the practical utility argument, but the text provides no details on the closed-loop simulator, scenario distribution, or comparison against uncompressed and alternative compression baselines at matched latency/memory.

    Authors: We will expand the closed-loop evaluation section to specify the simulator environment and its configuration, detail the scenario distribution (including proportions of stop/yield/proceed cases), and include direct comparisons against uncompressed models and alternative compression methods at matched latency and memory budgets to substantiate the maintained performance alongside the 3.3× speedup and 2.7× memory reduction. revision: yes

Circularity Check

0 steps flagged

No circularity: empirical gains from architectural change, not by-construction reduction

full rationale

The paper presents COMPACT-VA as an end-to-end trained conditional VQ-VAE architecture that compresses tokens while distilling a planning-intent latent. Reported gains (68.3% success rate, 3.3× speedup) are measured on external closed-loop driving metrics and ablations; no equations, derivations, or self-citations reduce these quantities to fitted parameters or prior outputs by construction. The prior-posterior training is standard VAE practice and the evaluation regime (observation-only at test time) is independent of the training objective.

Axiom & Free-Parameter Ledger

1 free parameters · 1 axioms · 1 invented entities

The central claim rests on the effectiveness of distilling and predicting a planning intent latent inside a conditional VQ-VAE; the abstract does not enumerate explicit free parameters or background axioms beyond standard VQ-VAE training assumptions.

free parameters (1)
  • planning intent latent dimension
    The dimensionality of the latent variable that carries planning intent is a design choice required for the conditional VQ-VAE.
axioms (1)
  • domain assumption Future trajectories contain extractable planning intent that can be used to supervise compression during training
    Invoked when the posterior encoder distills intent from future trajectories.
invented entities (1)
  • planning intent latent no independent evidence
    purpose: To condition token compression on decision-critical future information
    Introduced as the quantity distilled by the posterior and predicted by the prior encoder.

pith-pipeline@v0.9.1-grok · 5833 in / 1427 out tokens · 23234 ms · 2026-06-27T21:33:55.311109+00:00 · methodology

0 comments
read the original abstract

Monolithic vision-action models represent an emerging paradigm in autonomous driving. However, this architecture produces token sequences that quickly exceed real-time computational budgets when encoding extended temporal context for complex interactions. While approaches like linear transformers and external memory try to make the context lightweight, token compression is most compatible with the architecture as it requires no backbone modifications. Yet existing compression adopts rule-based heuristics like temporal decay, decoupled from planning, risking loss of decision-critical information. We propose COMPACT-VA, a planning-aligned working memory framework built on conditional VQ-VAE, compressing extended context into bounded representations. Compression is conditioned on both historical trajectory and a learned planning intent that the posterior encoder distills from future trajectories during training, while the prior encoder learns to predict it from compressed observations. The compressed memory, concatenated with the predicted latent, feeds the policy for end-to-end optimization, planning with retained decision-critical information. We evaluate on high-signal dynamic scenarios where historical context is most critical for behavior correctness (e.g., stop, yield, or proceed), and accordingly design behavioral metrics. Under comparable token budgets, we achieve $>$6% improvement (68.3%) on success rates with consistent gains across metrics. Ablations validate planning-aligned coupling effectiveness. Closed-loop evaluation confirms that COMPACT-VA maintained general driving performance with 3.3* speedup and 2.7* memory reduction over uncompressed processing.

Figures

Figures reproduced from arXiv: 2606.07464 by Alexander Popov, Boris Ivanovic, Boyi Li, Danfei Xu, Marco Pavone, Maximilian Igl, Nikolai Smolyanskiy, Peter Karkus, Ping Luo, Wenhao Ding, Yan Wang, Yiming Li, Yurong You, Yuxiao Chen, Zhixuan Liang.

Figure 1
Figure 1. Figure 1: From limited context to planning-aligned memory. At an all-way sign intersection, determining right-of-way requires observing traffic over extended periods. Rule-based temporal decay with a 2s window (t=-2s to t=0s) discards distant observations, losing information about which vehicle arrived first (observed at t=-3s) and causing incorrect go decisions (62% success). Our planning-aligned compression retain… view at source ↗
Figure 2
Figure 2. Figure 2: Overall architecture of COMPACT-VA. Multi-view raw observation tokens from current and past are compressed via Q-former conditioning on camera embeddings, historical trajectory and a learned driving intent latent. The posterior encoder (Q-net) distills intent from future trajectories during training, while the prior encoder (P-net) predicts this latent from compressed observations alone. The hierarchical F… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of high-signal dynamic scenarios. appends it as a special token, then autoregressively generates trajectory tokens. This preserves full compatibility with the unified VA architecture while enabling effective long-horizon planning under strict token budgets. IV. MEMORY-DEPENDENT DRIVING SCENARIOS AND EVALUATION FRAMEWORK We focus on scenarios where extended historical context plays a particular… view at source ↗
Figure 4
Figure 4. Figure 4: Closed-loop example at an all-way sign controlled right turn. Our method correctly identifies the all-way sign (top right) and yields to the oncoming through vehicle that arrived first. The bird’s-eye view (left) shows the reference trajectory (yellow) and predicted trajectory (blue) which are both short, indicating the decision to stop and wait. Camera views (right) capture the ego vehicle decelerating as… view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

41 extracted references · 5 canonical work pages · 3 internal anchors

  1. [1]

    Alpamayo-r1: Bridging reasoning and action prediction for generalizable autonomous driving in the long tail,

    Y . Wang, W. Luo, J. Bai, Y . Cao, T. Che, K. Chen, Y . Chen, J. Diamond, et al., “Alpamayo-r1: Bridging reasoning and action prediction for generalizable autonomous driving in the long tail,”arXiv, 2025

    2025

  2. [2]

    Planning-oriented autonomous driving,

    Y . Hu, J. Yang, L. Chen,et al., “Planning-oriented autonomous driving,” inProceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 17 853–17 862

    2023

  3. [3]

    Para-drive: Parallelized architecture for real-time autonomous driving,

    X. Weng, B. Ivanovic,et al., “Para-drive: Parallelized architecture for real-time autonomous driving,” inIEEE/CVF CVPR, 2024

    2024

  4. [4]

    Mamba: Linear-time sequence modeling with selective state spaces,

    A. Gu and T. Dao, “Mamba: Linear-time sequence modeling with selective state spaces,” inConference on language modeling, 2024

    2024

  5. [5]

    Memoryvla: Perceptual-cognitive memory in vision-language-action models for robotic manipulation,

    H. Shi, B. Xie, Y . Liu, L. Sun, F. Liu, T. Wang, E. Zhou, H. Fan, X. Zhang, and G. Huang, “Memoryvla: Perceptual-cognitive memory in vision-language-action models for robotic manipulation,”arXiv, 2025

    2025

  6. [6]

    Autonomous driving at unsignalized in- tersections: A review of decision-making challenges and reinforcement learning-based solutions,

    M. Al-Sharman, L. Edes,et al., “Autonomous driving at unsignalized in- tersections: A review of decision-making challenges and reinforcement learning-based solutions,”arXiv, 2024

    2024

  7. [7]

    Blip-2: Bootstrapping language-image pre- training with frozen image encoders and large language models,

    J. Li, D. Li,et al., “Blip-2: Bootstrapping language-image pre- training with frozen image encoders and large language models,” in International conference on machine learning. PMLR, 2023

    2023

  8. [8]

    Diffusion policy: Visuomotor policy learning via action diffusion,

    C. Chi, Z. Xu, S. Feng, E. Cousineau, Y . Du, B. Burchfiel, R. Tedrake, and S. Song, “Diffusion policy: Visuomotor policy learning via action diffusion,”The International Journal of Robotics Research, vol. 44, no. 10-11, pp. 1684–1704, 2025

    2025

  9. [9]

    π0.5: a vision- language-action model with open-world generalization,

    K. Black, N. Brown, J. Darpinian, K. Dhabalia, D. Driess, A. Esmail, M. R. Equi, C. Finn, N. Fusai, M. Y . Galliker,et al., “ π0.5: a vision- language-action model with open-world generalization,” in9th Annual Conference on Robot Learning, 2025

    2025

  10. [10]

    Adaptdiffuser: diffusion models as adaptive self-evolving planners,

    Z. Liang, Y . Mu, M. Ding, F. Ni, M. Tomizuka, and P. Luo, “Adaptdiffuser: diffusion models as adaptive self-evolving planners,” in Proceedings of the 40th International Conference on Machine Learning, 2023, pp. 20 725–20 745

    2023

  11. [11]

    Discrete diffusion vla: Bringing discrete diffusion to action decoding in vision-language-action policies,

    Z. Liang, Y . Li, T. Yang, C. Wu, S. Mao, T. Nian, L. Pei, S. Zhou, X. Yang, J. Pang,et al., “Discrete diffusion vla: Bringing discrete diffusion to action decoding in vision-language-action policies,” in Proceedings of the 43rd International Conference on Machine Learning, 2026

    2026

  12. [12]

    GR00T N1: An Open Foundation Model for Generalist Humanoid Robots

    J. Bjorck, F. Casta˜neda, N. Cherniadev, X. Da, R. Ding, L. Fan, Y . Fang, D. Fox, F. Hu, S. Huang,et al., “Gr00t n1: An open foundation model for generalist humanoid robots,”arXiv preprint arXiv:2503.14734, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  13. [13]

    Skilldiffuser: Interpretable hierarchical planning via skill abstractions in diffusion-based task execution,

    Z. Liang, Y . Mu, H. Ma, M. Tomizuka, M. Ding, and P. Luo, “Skilldiffuser: Interpretable hierarchical planning via skill abstractions in diffusion-based task execution,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 16 467–16 476

    2024

  14. [14]

    Embodied navigation foundation model.arXiv preprint arXiv:2509.12129,

    J. Zhang, A. Li, Y . Qi, M. Li, J. Liu, S. Wang, H. Liu, G. Zhou, Y . Wu, X. Li,et al., “Embodied navigation foundation model,”arXiv preprint arXiv:2509.12129, 2025

    work page arXiv 2025

  15. [15]

    Uni-navid: A video-based vision-language-action model for unifying embodied navigation tasks,

    J. Zhang, K. Wang, S. Wang, M. Li, H. Liu, S. Wei, Z. Wang, Z. Zhang, and H. Wang, “Uni-navid: A video-based vision-language-action model for unifying embodied navigation tasks,”Robotics: Science and Systems, 2025

    2025

  16. [16]

    Impromptu vla: Open weights and open data for driving vision-language-action models,

    H. Chi, H.-a. Gao, Z. Liu, J. Liu, C. Liu, J. Li, K. Yang, Y . Yu, Z. Wang, W. Li,et al., “Impromptu vla: Open weights and open data for driving vision-language-action models,”Advances in Neural Information Processing Systems, vol. 38, 2026

    2026

  17. [17]

    dvlm-ad: Enhance diffusion vision-language- model for driving via controllable reasoning,

    Y . Ma, Y . Cao, W. Ding, S. Zhang, Y . Wang, B. Ivanovic, M. Jiang, M. Pavone, and C. Xiao, “dvlm-ad: Enhance diffusion vision-language- model for driving via controllable reasoning,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 1050–1061

    2026

  18. [18]

    Probabilistic robotics,

    S. Thrun, “Probabilistic robotics,”Communications of the ACM, vol. 45, no. 3, pp. 52–57, 2002

    2002

  19. [19]

    Dream to control: Learning behaviors by latent imagination,

    D. Hafner, T. Lillicrap, J. Ba, and M. Norouzi, “Dream to control: Learning behaviors by latent imagination,” inICLR, 2019

    2019

  20. [20]

    Planning and acting in partially observable stochastic domains,

    L. P. Kaelbling, M. L. Littman, and A. R. Cassandra, “Planning and acting in partially observable stochastic domains,”Artificial intelligence, vol. 101, no. 1-2, pp. 99–134, 1998

    1998

  21. [21]

    Palm-e: an embodied multimodal language model,

    D. Driess, F. Xia, M. S. Sajjadi,et al., “Palm-e: an embodied multimodal language model,” inICML, 2023

    2023

  22. [22]

    Compressive transformers for long-range sequence modelling,

    J. W. Rae, A. Potapenko, S. M. Jayakumar, C. Hillier, and T. P. Lillicrap, “Compressive transformers for long-range sequence modelling,” in International Conference on Learning Representations, 2019

    2019

  23. [23]

    Transformer-xl: Attentive language models beyond a fixed-length context,

    Z. Dai, Z. Yang, Y . Yang, J. G. Carbonell, Q. Le, and R. Salakhutdinov, “Transformer-xl: Attentive language models beyond a fixed-length context,” inProceedings of the 57th ACL, 2019, pp. 2978–2988

    2019

  24. [24]

    Learning to (learn at test time): Rnns with expressive hidden states,

    Y . Sun, X. Li, K. Dalal,et al., “Learning to (learn at test time): Rnns with expressive hidden states,” inForty-second International Conference on Machine Learning, 2024

    2024

  25. [25]

    Titans: Learning to Memorize at Test Time

    A. Behrouz, P. Zhong, and V . Mirrokni, “Titans: Learning to memorize at test time,”arXiv preprint arXiv:2501.00663, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  26. [26]

    Token merging: Your vit but faster,

    D. Bolya, C.-Y . Fu, X. Dai, P. Zhang, C. Feichtenhofer, and J. Hoffman, “Token merging: Your vit but faster,” inThe Eleventh International Conference on Learning Representations, 2023

    2023

  27. [27]

    Efficient streaming language models with attention sinks,

    G. Xiao, Y . Tian, B. Chen, S. Han, and M. Lewis, “Efficient streaming language models with attention sinks,” inThe Twelfth International Conference on Learning Representations, 2024

    2024

  28. [28]

    Longformer: The Long-Document Transformer

    I. Beltagy, M. E. Peters, and A. Cohan, “Longformer: The long- document transformer,”arXiv preprint arXiv:2004.05150, 2020

    work page internal anchor Pith review Pith/arXiv arXiv 2004

  29. [29]

    Big bird: Transformers for longer sequences,

    M. Zaheer, G. Guruganesh,et al., “Big bird: Transformers for longer sequences,”Advances in neural information processing systems, vol. 33, pp. 17 283–17 297, 2020

    2020

  30. [30]

    Hymba: A hybrid-head architecture for small language models,

    X. Dong, Y . Fu, S. Diao,et al., “Hymba: A hybrid-head architecture for small language models,” inThe Thirteenth International Conference on Learning Representations, 2025

    2025

  31. [31]

    Hint-ad: Holistically aligned interpretability in end-to-end autonomous driving,

    K. Ding, B. Chen, Y . Su, H.-a. Gao, B. Jin, C. Sima, X. Li, W. Zhang, P. Barsch, H. Li,et al., “Hint-ad: Holistically aligned interpretability in end-to-end autonomous driving,” inConference on Robot Learning. PMLR, 2025, pp. 3742–3765

    2025

  32. [32]

    Dinov2: Learning robust visual features without supervision,

    M. Oquab, T. Darcet, T. Moutakanni, H. V . V o, M. Szafraniec, V . Khalidov, P. Fernandez, D. HAZIZA, F. Massa, A. El-Nouby, et al., “Dinov2: Learning robust visual features without supervision,” Transactions on Machine Learning Research, 2024

    2024

  33. [33]

    Finite scalar quantization: Vq-vae made simple,

    F. Mentzer, D. Minnen, E. Agustsson, and M. Tschannen, “Finite scalar quantization: Vq-vae made simple,” inThe Twelfth International Conference on Learning Representations, 2024

    2024

  34. [34]

    Scaling rectified flow transformers for high-resolution image synthesis,

    P. Esser, S. Kulal, A. Blattmann,et al., “Scaling rectified flow transformers for high-resolution image synthesis,” inICML, 2024

    2024

  35. [35]

    Roformer: En- hanced transformer with rotary position embedding,

    J. Su, M. Ahmed, Y . Lu, S. Pan, W. Bo, and Y . Liu, “Roformer: En- hanced transformer with rotary position embedding,”Neurocomputing, vol. 568, p. 127063, 2024

    2024

  36. [36]

    Tikhonov regularization and total least squares,

    G. H. Golub, P. C. Hansen, and D. P. O’Leary, “Tikhonov regularization and total least squares,”SIAM journal on matrix analysis and applications, vol. 21, no. 1, pp. 185–194, 1999

    1999

  37. [37]

    Neural discrete representation learning,

    A. Van Den Oordet al., “Neural discrete representation learning,” Advances in neural information processing systems, vol. 30, 2017

    2017

  38. [38]

    You’ll never walk alone: Modeling social behavior for multi-target tracking,

    S. Pellegrini, A. Ess, K. Schindler, and L. Van Gool, “You’ll never walk alone: Modeling social behavior for multi-target tracking,” in International conference on computer vision, 2009, pp. 261–268

    2009

  39. [39]

    What matters to enhance traffic rule compliance of imitation learning for end-to-end autonomous driving,

    H. Zhou, W. Cao, A. Sui, and Z. Bing, “What matters to enhance traffic rule compliance of imitation learning for end-to-end autonomous driving,”arXiv preprint arXiv:2309.07808, 2023

    work page arXiv 2023

  40. [40]

    Physicalai autonomous vehicles dataset,

    NVIDIA, “Physicalai autonomous vehicles dataset,” https://huggingface. co/datasets/nvidia/PhysicalAI-Autonomous-Vehicles, 2025, one of the largest, geographically diverse datasets with 1,727+ hours of driving data (camera, LiDAR, radar) across 25 countries and 2500+ cities

    2025

  41. [41]

    AlpaSim: A modular, lightweight, and data-driven research simulator for autonomous driving,

    NVIDIAet al., “AlpaSim: A modular, lightweight, and data-driven research simulator for autonomous driving,” 2025. [Online]. Available: https://github.com/NVlabs/alpasim

    2025