Pith. sign in

REVIEW 3 major objections 4 minor 29 references

A fixed-depth planner that generates 48 proposals in a single pass and aggregates them without trajectory-quality labels beats prior one-step planners on NAVSIM.

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 · deepseek-v4-flash

2026-08-02 01:52 UTC pith:FM2RURDM

load-bearing objection DRIFT is a solid, clearly-written planner paper whose main caveat is an underspecified test-time prior; worth reviewing, not worth rejecting. the 3 major comments →

arxiv 2607.14507 v1 pith:FM2RURDM submitted 2026-07-16 cs.RO

DRIFT: Drift and Aggregation for Motion Planning

Yining Xing , Zhiyuan Liu , Zehong Ke , Wenhao Yu , Jianqiang Wang This is my paper
classification cs.RO
keywords motion planningtrajectory generationlatent space driftproposal aggregationend-to-end autonomous drivingNAVSIMone-step generationimitation learning
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 claims that an end-to-end planner can handle multiple driving hypotheses efficiently by generating 48 proposal features in a single batched pass inside a compact trajectory latent space, then learning to aggregate those hidden features into one executable trajectory without any trajectory-quality labels. On the NAVSIM benchmark, this design reports 89.6 PDMS and 90.4 EPDMS, slightly ahead of the strongest one-step baselines, with proposal generation and aggregation running in 10.82 ms on an RTX 4090. The central ablation shows the learned aggregation head beats the arithmetic mean of the same 48 proposals by 4.5 PDMS, indicating that the way proposals are combined, not just how they are generated, carries much of the benefit. A map-derived boundary regularizer and a drift intensity factor that interpolates between exploratory and expert-attracted proposal distributions round out the method.

Core claim

DRIFT's central claim is that one-step drifting in a PCA trajectory latent space, followed by scene-aware pooling of the resulting hidden features, is an effective and efficient alternative to iterative generative planning and to evaluator-supervised trajectory scoring. The decoder produces 48 hidden proposal features in one pass (32 at a low drift intensity, 16 at a high intensity), and the aggregation head cross-attends to these features along with scene, navigation, and ego-state tokens to directly regress the final trajectory. Training uses imitation of the expert trajectory and a piecewise road-boundary regularizer; no trajectory-level quality labels are used for aggregation. The author

What carries the argument

The load-bearing mechanism is the combination of latent-space drift and feature-level aggregation. The drift field, computed with an RBF affinity kernel and Sinkhorn symmetrization, attracts each generated latent proposal toward expert trajectories while repelling it from sibling proposals, and a scalar alpha interpolates between an unconditional map-compatible prior and the conditional expert pull. The Aggregation Head then pools the 48 hidden proposal features through learnable query attention, refines them with gated cross-attention to scene features, and decodes the result into waypoints and heading. The PCA latent space parameterizes proposal generation only; the final output is decoded

Load-bearing premise

The method assumes that a library of training-set expert trajectories, filtered by road-boundary compatibility, is a sufficient geometric prior for any test scene; if a test road shape is under-represented in that library, the exploratory proposals and therefore the final plan may be biased.

What would settle it

Evaluate DRIFT on a test set dominated by road geometries that do not appear in the training trajectory library (for example, roundabouts, hairpins, or parking lots) and compare PDMS against the mean-of-proposals baseline; if the gap collapses or the method falls below the strongest one-step baseline, the gains are attributable to the prior library rather than to label-free aggregation.

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

If this is right

  • Multi-hypothesis planning can run at fixed depth with one batched decoder pass, removing iterative denoising or autoregressive steps from the critical path.
  • Proposal-to-plan conversion can be trained imitatively with geometric regularizers, avoiding the need for evaluator-derived trajectory quality labels.
  • Pooling hidden features before physical decoding insulates the final trajectory from PCA reconstruction error.
  • The 10.82 ms proposal-and-aggregation latency is small relative to the full-model 66.43 ms, so the planning step is not the bottleneck.
  • The learned aggregation head contributes substantially more (4.5 PDMS) than the proposal generator or boundary regularizer alone, highlighting aggregation as the key design choice.

Where Pith is reading between the lines

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

  • The unconditional prior is built from ground-truth trajectories of other scenes filtered by compatibility with the current map; at test time on maps under-represented in the training set, the exploratory proposals may inherit a geometric bias, and the paper does not describe how the prior is constructed for novel maps.
  • If the gap between learned aggregation and mean-averaging persists across diverse map geometries, it would suggest the learned head is genuinely scene-aware rather than merely compensating for a poor proposal distribution.
  • The paper's own future-work call suggests the strongest test is pseudo-closed-loop and interactive evaluation, where non-reactive simulation may not capture the safety value of the boundary regularizer.
  • A direct measurement of proposal diversity and aggregation weights across scenes would clarify how much exploration the latent drift actually provides.

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 / 4 minor

Summary. DRIFT is an end-to-end trajectory planner for NAVSIM that generates 48 latent-space trajectory proposals in one batched forward pass (32 at drift intensity α=0.5 and 16 at α=0.9) and aggregates them into a single executable trajectory with a learned, label-free aggregation head. The aggregation head is trained with expert-trajectory imitation and a map-derived boundary regularizer, without trajectory-quality labels. On NAVSIM navtest, DRIFT reports 89.6 PDMS (v1) and 90.4 EPDMS (v2), outperforming the compared non-scorer-based planners (MeanFuser, Drive-JEPA) by 0.6 and 0.9 points, with proposal generation and aggregation taking 10.82 ms on an RTX 4090.

Significance. If the reported results hold, DRIFT provides a compelling demonstration that one-step latent-space proposal generation plus scene-aware, label-free aggregation can match or exceed more complex iterative or scorer-based designs on a standard open-loop benchmark. The paper includes several commendable elements: a clear architectural description, a fixed-depth inference profile, an ablation showing that learned aggregation is substantially better than averaging the same proposals (89.6 vs 85.1 PDMS), and a multi-seed robustness analysis demonstrating that the gains are not due to favorable random sampling. The method avoids evaluator-derived trajectory labels for aggregation, which is an interesting and timely design point.

major comments (3)
  1. [Section IV-A, Eq. (4)] The unconditional prior V_unc is defined only for training scenes: 'we extract structured road graph data for each training scene... ground-truth trajectories from other scenes... retained, and 32 diverse trajectories are selected.' The inference-time construction of V_unc on navtest is never specified. Since Eq. (4) sets V_total = V_unc + α(V_cond − V_unc), the α=0.5 proposals used in the main results are 50% determined by this prior. If V_unc at test time is built from a navtrain trajectory library filtered by map compatibility, the paper must report coverage statistics (e.g., how many navtest scenes have fewer than 32 compatible trajectories) and the diversity selection procedure; if V_unc is instead built from navtest ground-truth trajectories, the main numbers include test-set leakage. Either way, the central result as reported is not reproducible without this specification. Please
  2. [Section IV-C, Tables I and II] The comparison is restricted to non-scorer-based planners, which is stated, but the claim 'state-of-the-art' is not qualified in the abstract. Scorer-based methods (e.g., GRADE or others trained with trajectory-quality labels) are excluded, so the paper's stated 'new state-of-the-art results in this setting' is accurate only within the chosen subset. Please either include a representative scorer-based baseline in the main comparison or explicitly qualify the claim as 'among non-scorer-based planners' in the abstract and conclusion to avoid overclaiming.
  3. [Section III-E, Eq. (6)] The road boundary loss is described as 'differentiably computed' but the manuscript does not state how the unsigned minimum distance to the nearest road boundary is computed in a differentiable manner (e.g., via a signed distance field, polygonal distance, or a neural approximation). Since this loss contributes to training and the paper's DAC improvements are partly attributed to it, the implementation must be specified for reproducibility. Also, the hard constants in Eq. (6) — C_out, m_safe, m_soft, ε, γ — are not given numerically; please list them or state that they follow a prior setting.
minor comments (4)
  1. [Section III-E] Typo: 'differentiably computed, the unsigned minimum distance' should be 'differentiably computed the unsigned minimum distance' or similar. Also, Eq. (5) defines w=1/(γ+ℓ) but γ is never specified; please state its value or that it is a small smoothing constant.
  2. [Section IV-B] The PCA latent dimension is stated as 8, but the paper does not report the fraction of variance explained or sensitivity to this choice. A brief sentence on the PCA reconstruction error or a small ablation on latent dimension would strengthen confidence in the latent-space design.
  3. [Section III-C] The drift loss is written as minimizing ||x_k − sg(x_k + V_total_k)||^2, but the weight λ_2 and its scheduling are given only as '1 in the first stage, 0.1 in the second.' The Sinkhorn temperature set {τ_s} and the number of positive/negative samples are not reported. Please include these hyperparameters, either in the text or in a table.
  4. [Section IV-F, Table IV] The ablation 'w/o Decoder' removes the DRIFT Decoder, but it is unclear whether the 48 proposals are then replaced by something else (e.g., random PCA samples) or whether the aggregation head operates on a different proposal set. Please specify how the 'w/o Decoder' variant generates its proposals; otherwise the ablation is difficult to interpret.

Circularity Check

0 steps flagged

No significant circularity: DRIFT's proposals and final trajectory are trained on expert trajectories and map geometry, not on the reported PDMS/EPDMS metrics; self-citations are contextual, not load-bearing.

full rationale

The derivation chain is not circular. The drift loss (Eqs. 3-4) uses the expert trajectory as positive samples and a training-time unconditional prior built from other scenes' ground-truth trajectories; V_unc is part of the training objective, and at inference the decoder is simply sampled from Gaussian noise with alpha=0.5/0.9, so no test-side ground-truth is fed into Eq. 4 by construction. The Aggregation Head is trained with imitation loss (Eq. 5) and a map-derived boundary regularizer (Eq. 6), neither of which uses PDMS/EPDMS or trajectory-quality labels as a target; the reported metrics are measured on held-out navtest after training. The key ablation replaces learned aggregation with the arithmetic mean of the same 48 proposals, so the 4.5 PDMS gain is a controlled comparison rather than a fitted-input prediction. Self-citations (e.g., [10], [12], [15]) appear only as related-work context, baselines, or design comparisons and do not supply the load-bearing argument; no uniqueness theorem or ansatz is imported from the authors' prior work. The underspecified construction of the unconditional prior at inference (Section IV-A) is a reproducibility/coverage concern, but it is a correctness risk rather than a circular step because the evaluation result is not an input to the model.

Axiom & Free-Parameter Ledger

6 free parameters · 4 axioms · 0 invented entities

The central claim depends on several chosen hyperparameters (drift temperatures, boundary margins, loss weights) whose values are not reported, and on the availability of a map-compatible training-trajectory library for the unconditional drift prior. No free physical parameters or invented entities appear.

free parameters (6)
  • PCA latent dimension = 8
    16-dim position vectors projected to 8-dim latent space; chosen by hand (Section IV-B).
  • Sinkhorn temperatures τ_s = not reported
    Drift field summed over multiple temperatures (Section III-C); values not given.
  • Boundary loss margins and penalty = not reported (m_safe, m_soft, C_out, ε)
    Piecewise loss in Eq. 6; no values reported.
  • Loss weights λ2, λ3 = 0.1, 0.2 (stage 2)
    Given in Section IV-B; chosen.
  • Proposal count and α values = 32 at α=0.5, 16 at α=0.9
    Inference-time fixed set; chosen by hand.
  • Arc-length smoothing γ = not reported
    Used in Eq. 5 to reweight samples.
axioms (4)
  • domain assumption NAVSIM map annotations (drivable polygon, road boundaries) are accurate enough for point-in-polygon and distance queries
    Boundary loss and unconditional prior filtering rely on map accuracy (Sections IV-A, III-E).
  • domain assumption Front-view camera history plus navigation command and ego state is sufficient input to predict the expert trajectory
    The architecture omits LiDAR and BEV sensors; relies on V-JEPA features from front-view frames (Section III-A).
  • domain assumption PCA fitted on navtrain ground-truth trajectories captures the behavioral diversity needed for proposals
    Proposal generation is supervised in PCA space (Eq. 2); the fixed basis is never updated (Section IV-B).
  • domain assumption The expert trajectory is an appropriate positive sample for the drift objective
    Conditional drift attracts generated samples to the expert trajectory (Eq. 3); inherited from Drifting Models.

pith-pipeline@v1.3.0-alltime-deepseek · 11423 in / 14320 out tokens · 128474 ms · 2026-08-02T01:52:28.368034+00:00 · methodology

0 comments
read the original abstract

End-to-end trajectory planners need to represent multiple plausible driving behaviors while producing a single executable trajectory under real-time constraints. Proposal-based approaches address this ambiguity by generating multiple candidates, but converting the proposal set into a final plan remains a key design problem. We present DRIFT, a fixed-depth planner that combines one-step drifting in a compact trajectory latent space with scene-aware proposal aggregation. Conditioned on features from a pretrained visual encoder, the DRIFT Decoder generates 48 proposal features in a single batched pass, with 32 samples at alpha=0.5 and 16 samples at alpha=0.9. A lightweight Aggregation Head integrates these features with scene, navigation, and ego-state information and directly predicts the final trajectory without requiring trajectory-level quality labels for aggregation. Its output is trained with expert-trajectory imitation and a map-derived boundary regularizer that penalizes waypoints outside the drivable polygon and inside waypoints near its boundary. On NAVSIM navtest, DRIFT achieves 89.6 PDMS and 90.4 EPDMS, with strong drivable-area compliance and ego progress among the methods compared. The proposal-generation and aggregation module runs in 10.82 ms on an NVIDIA RTX 4090, while full-model inference including the visual backbone takes 66.43 ms. These results show that one-step latent proposal generation and direct aggregation provide an efficient design for multi-hypothesis motion planning.

Figures

Figures reproduced from arXiv: 2607.14507 by Jianqiang Wang, Wenhao Yu, Yining Xing, Zehong Ke, Zhiyuan Liu.

Figure 1
Figure 1. Figure 1: Overview of the DRIFT architecture. A V-JEPA pretrained ViT encodes front-view camera frames into shared scene features. The DRIFT Decoder [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Unconditional trajectory samples for two representative scenes. [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Planning results across five representative driving scenarios. The top row shows the front-view camera image. The bottom row shows the BEV [PITH_FULL_IMAGE:figures/full_fig_p007_3.png] 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

29 extracted references · 6 linked inside Pith

  1. [1]

    Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline,

    P. Wu, X. Jia, L. Chen, J. Yan, H. Li, and Y . Qiao, “Trajectory-guided control prediction for end-to-end autonomous driving: A simple yet strong baseline,”Advances in Neural Information Processing Systems, vol. 35, pp. 6119–6132, 2022

    2022

  2. [2]

    Planning-oriented autonomous driving,

    Y . Hu, J. Yang, L. Chen, K. Li, C. Sima, X. Zhu, S. Chai, S. Du, T. Lin, W. Wanget al., “Planning-oriented autonomous driving,” in Proceedings of the IEEE/CVF conference on computer vision and pattern recognition, 2023, pp. 17 853–17 862

    2023

  3. [3]

    Parting with misconceptions about learning-based vehicle motion planning,

    D. Dauner, M. Hallgarten, A. Geiger, and K. Chitta, “Parting with misconceptions about learning-based vehicle motion planning,” in Conference on Robot Learning. PMLR, 2023, pp. 1268–1281

    2023

  4. [4]

    End-to-end autonomous driving: Challenges and frontiers,

    L. Chen, P. Wu, K. Chitta, B. Jaeger, A. Geiger, and H. Li, “End-to-end autonomous driving: Challenges and frontiers,”IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 46, no. 12, pp. 10 164– 10 183, 2024

    2024

  5. [5]

    Trajectory prediction for autonomous driving: Progress, limitations, and future directions,

    N. A. Madjid, A. Ahmad, M. Mebrahtu, Y . Babaa, A. Nasser, S. Malik, B. Hassan, N. Werghi, J. Dias, and M. Khonji, “Trajectory prediction for autonomous driving: Progress, limitations, and future directions,” Information Fusion, p. 103588, 2025

    2025

  6. [6]

    Detra: A unified model for object detection and trajectory forecast- ing,

    S. Casas, B. Agro, J. Mao, T. Gilles, A. Cui, T. Li, and R. Urtasun, “Detra: A unified model for object detection and trajectory forecast- ing,” inEuropean Conference on Computer Vision. Springer, 2024, pp. 326–342

    2024

  7. [7]

    Diffusiondrive: Truncated diffusion model for end-to-end autonomous driving,

    B. Liao, S. Chen, H. Yin, B. Jiang, C. Wang, S. Yan, X. Zhang, X. Li, Y . Zhang, Q. Zhanget al., “Diffusiondrive: Truncated diffusion model for end-to-end autonomous driving,” inProceedings of the Computer Vision and Pattern Recognition Conference, 2025, pp. 12 037–12 047

    2025

  8. [8]

    Diffusion-based planning for autonomous driv- ing with flexible guidance,

    Y . Zheng, R. Liang, K. Zheng, J. Zheng, L. Mao, J. Li, W. Gu, R. Ai, S. Li, X. Zhanet al., “Diffusion-based planning for autonomous driv- ing with flexible guidance,” inInternational Conference on Learning Representations, vol. 2025, 2025, pp. 37 207–37 227

    2025

  9. [9]

    Genad: Generalized predictive model for autonomous driving,

    J. Yang, S. Gao, Y . Qiu, L. Chen, T. Li, B. Dai, K. Chitta, P. Wu, J. Zeng, P. Luoet al., “Genad: Generalized predictive model for autonomous driving,”arXiv preprint arXiv:2403.09630, 2024

    Pith/arXiv arXiv 2024

  10. [10]

    Meanfuser: Fast one-step multi-modal trajectory generation and adaptive reconstruction via meanflow for end-to-end autonomous driving,

    J. Wang, Y . Zheng, X. Liu, Z. Xing, P. Li, K. Ma, H. Ye, G. Chen, G. Li, L. Chenet al., “Meanfuser: Fast one-step multi-modal trajectory generation and adaptive reconstruction via meanflow for end-to-end autonomous driving,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 17 884–17 893

    2026

  11. [11]

    Generalized trajectory scoring for end-to- end multimodal planning,

    Z. Li, W. Yao, Z. Wang, X. Sun, J. Chen, N. Chang, M. Shen, Z. Wu, S. Lan, and J. M. Alvarez, “Generalized trajectory scoring for end-to- end multimodal planning,”arXiv preprint arXiv:2506.06664, 2025

    Pith/arXiv arXiv 2025

  12. [12]

    Grade: Guiding realistic autonomous driving with adaptive trajectory evolution,

    Z. Ke, Z. Liu, Y . Wang, J. Li, J. Jiang, Y . Jiang, Z. Xu, and J. Wang, “Grade: Guiding realistic autonomous driving with adaptive trajectory evolution,” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2026, pp. 1029–1038

    2026

  13. [13]

    Drive-jepa: Video jepa meets multi- modal trajectory distillation for end-to-end driving,

    L. Wang, Z. Yang, C. Bai, G. Zhang, X. Liu, X. Zheng, X.-X. Long, C.-T. Lu, and C. Lu, “Drive-jepa: Video jepa meets multi- modal trajectory distillation for end-to-end driving,”arXiv preprint arXiv:2601.22032, 2026

    Pith/arXiv arXiv 2026

  14. [14]

    Driveadapter: Breaking the coupling barrier of perception and planning in end-to-end autonomous driving,

    X. Jia, Y . Gao, L. Chen, J. Yan, P. L. Liu, and H. Li, “Driveadapter: Breaking the coupling barrier of perception and planning in end-to-end autonomous driving,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2023, pp. 7953–7963

    2023

  15. [15]

    Comdrive: Comfort-oriented end-to- end autonomous driving,

    J. Wang, X. Zhang, Z. Xing, S. Gu, X. Guo, Y . Hu, Z. Song, Q. Zhang, X. Long, and W. Yin, “Comdrive: Comfort-oriented end-to- end autonomous driving,” in2025 IEEE/RSJ International Conference on Intelligent Robots and Systems (IROS). IEEE, 2025, pp. 2682– 2689

    2025

  16. [16]

    Revisiting feature prediction for learning visual representations from video,

    A. Bardes, Q. Garrido, J. Ponce, X. Chen, M. Rabbat, Y . LeCun, M. Assran, and N. Ballas, “Revisiting feature prediction for learning visual representations from video,”arXiv preprint arXiv:2404.08471, 2024

    Pith/arXiv arXiv 2024

  17. [17]

    Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking,

    D. Dauner, M. Hallgarten, T. Li, X. Weng, Z. Huang, Z. Yang, H. Li, I. Gilitschenski, B. Ivanovic, M. Pavone, A. Geiger, and K. Chitta, “Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking,” inAdvances in Neural Information Processing Systems (NeurIPS), 2024

    2024

  18. [18]

    Transfuser: Imitation with transformer-based sensor fusion for au- tonomous driving,

    K. Chitta, A. Prakash, B. Jaeger, Z. Yu, K. Renz, and A. Geiger, “Transfuser: Imitation with transformer-based sensor fusion for au- tonomous driving,”IEEE transactions on pattern analysis and machine intelligence, vol. 45, no. 11, pp. 12 878–12 895, 2022

    2022

  19. [19]

    Drivetransformer: Unified transformer for scalable end-to-end autonomous driving,

    X. Jia, J. You, Z. Zhang, and J. Yan, “Drivetransformer: Unified transformer for scalable end-to-end autonomous driving,” inThe Thirteenth International Conference on Learning Representations

  20. [20]

    Gameformer: Game-theoretic modeling and learning of transformer-based interactive prediction and planning for autonomous driving,

    Z. Huang, H. Liu, and C. Lv, “Gameformer: Game-theoretic modeling and learning of transformer-based interactive prediction and planning for autonomous driving,” inProceedings of the IEEE/CVF Interna- tional Conference on Computer Vision, 2023, pp. 3903–3913

    2023

  21. [21]

    Diffrefiner: Coarse to fine trajectory planning via diffusion refinement with semantic interaction for end to end autonomous driving,

    L. Yin, R. Ju, G. Guo, and E. Cheng, “Diffrefiner: Coarse to fine trajectory planning via diffusion refinement with semantic interaction for end to end autonomous driving,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 14, 2026, pp. 12 009– 12 017

    2026

  22. [22]

    Hipro-ad: Sparse trajectory transformer for end-to-end autonomous driving with hybrid spatiotemporal attention,

    B. Chen, G. Wang, J. Yang, S. Huang, X. Qian, B. Huang, and G. Guo, “Hipro-ad: Sparse trajectory transformer for end-to-end autonomous driving with hybrid spatiotemporal attention,”Sensors (Basel, Switzer- land), vol. 26, no. 1, p. 185, 2025

    2025

  23. [23]

    Generative modeling via drifting,

    M. Deng, H. Li, T. Li, Y . Du, and K. He, “Generative modeling via drifting,”arXiv preprint arXiv:2602.04770, 2026

    Pith/arXiv arXiv 2026

  24. [24]

    End-to-end driving with online trajectory evaluation via bev world model,

    Y . Li, Y . Wang, Y . Liu, J. He, L. Fan, and Z. Zhang, “End-to-end driving with online trajectory evaluation via bev world model,” in Proceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 27 137–27 146

    2025

  25. [25]

    World4drive: End-to-end autonomous driving via intention-aware physical latent world model,

    Y . Zheng, P. Yang, Z. Xing, Q. Zhang, Y . Zheng, Y . Gao, P. Li, T. Zhang, Z. Xia, P. Jiaet al., “World4drive: End-to-end autonomous driving via intention-aware physical latent world model,” inProceed- ings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 28 632–28 642

    2025

  26. [26]

    Epona: Autoregressive diffusion world model for autonomous driving,

    K. Zhang, Z. Tang, X. Hu, X. Pan, X. Guo, Y . Liu, J. Huang, L. Yuan, Q. Zhang, X.-X. Longet al., “Epona: Autoregressive diffusion world model for autonomous driving,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 27 220– 27 230

    2025

  27. [27]

    Is ego status all you need for open-loop end-to-end autonomous driving?

    Z. Li, Z. Yu, S. Lan, J. Li, J. Kautz, T. Lu, and J. M. Alvarez, “Is ego status all you need for open-loop end-to-end autonomous driving?” inProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 14 864–14 873

    2024

  28. [28]

    Hydra-mdp++: Advancing end-to-end driving via expert- guided hydra-distillation,

    K. Li, Z. Li, S. Lan, Y . Xie, Z. Zhang, J. Liu, Z. Wu, Z. Yu, and J. M. Alvarez, “Hydra-mdp++: Advancing end-to-end driving via expert- guided hydra-distillation,”arXiv preprint arXiv:2503.12820, 2025

    Pith/arXiv arXiv 2025

  29. [29]

    Drivesuprim: Towards precise trajectory selection for end- to-end planning,

    W. Yao, Z. Li, S. Lan, Z. Wang, X. Sun, J. M. Alvarez, and Z. Wu, “Drivesuprim: Towards precise trajectory selection for end- to-end planning,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 14, 2026, pp. 11 910–11 918

    2026