pith. sign in

arxiv: 2606.04884 · v1 · pith:WBQJTPQBnew · submitted 2026-06-03 · 💻 cs.RO

D³-MoE:Dual Disentangled Diffusion Mixture-of-Experts for Style-Controllable End-to-End Autonomous Driving

Renju Feng , Rukang Wang , Ning Xi , Jianguo Yu , Liping Lu , Pan Zhou , Duanfeng Chu This is my paper

Pith reviewed 2026-06-28 05:59 UTC · model grok-4.3

classification 💻 cs.RO
keywords autonomous drivingdiffusion modelsmixture of expertsstyle controllabilitytrajectory planningdisentangled representationsend-to-end learningNAVSIM benchmark
0
0 comments X

The pith

D³-MoE disentangles behavioral style from physical long-lat axes in a diffusion mixture-of-experts so that multi-style trajectories can be generated and selected without averaging human demonstrations.

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

The paper aims to fix the style-averaging problem that turns varied human driving data into bland, sometimes unsafe end-to-end policies. It does so by splitting the modeling task along two axes: a behavioral axis that uses a style-conditioned diffusion process to create multiple candidate trajectories in parallel, and a physical axis that routes longitudinal and lateral motion through separate expert modules. Those experts are diffusion transformers trained only on self-supervised kinematic targets and later reassembled into one coherent path. A downstream selector then picks the trajectory that matches a user preference or score. If the separation works, the result is both higher planning accuracy and explicit control over driving style on the same scene.

Core claim

By decoupling generation from selection on the behavioral axis and longitudinal from lateral dynamics on the physical axis, the dual disentangled diffusion mixture-of-experts synthesizes multi-style candidate trajectories via style-conditioned diffusion while independent expert routers, trained on orthogonal ground-truth kinematics, predict their respective states before reassembly into kinematically coherent output, reaching 88.2 PDMS and 84.3 EPDMS on NAVSIM by default and higher with Best-of-Three ensemble.

What carries the argument

Dual disentanglement separating behavioral generation-selection from physical longitudinal-lateral routing inside style-conditioned Diffusion Transformer experts.

If this is right

  • Multiple style-conditioned trajectories are generated in one forward pass for a single scene and can be chosen downstream by preference or score.
  • Longitudinal and lateral experts activate separately at inference and each predicts only its own physical component.
  • Style conditioning is injected through AdaLN and asymmetric cross-attention without requiring manual style labels.
  • Default performance reaches 88.2 PDMS and 84.3 EPDMS on NAVSIM, rising to 91.3 PDMS and 87.5 EPDMS under Best-of-Three ensemble.

Where Pith is reading between the lines

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

  • The same behavioral-physical split might let other generative planners offer user-selectable modes without retraining the whole model.
  • Measuring how often the reassembled paths violate basic kinematic constraints would directly test whether the orthogonal-router assumption holds in practice.
  • If the self-supervised kinematic targets generalize across datasets, the approach could be applied to driving logs collected in new cities or weather conditions with little extra labeling.

Load-bearing premise

Self-supervised targets taken from orthogonal ground-truth kinematics are enough to train independent longitudinal and lateral routers whose separate outputs can be reassembled into safe, coherent trajectories.

What would settle it

Reassembled trajectories from the independent routers showing measurably higher rates of kinematic violation, jerk, or collision than a jointly trained baseline on the NAVSIM validation set.

Figures

Figures reproduced from arXiv: 2606.04884 by Duanfeng Chu, Jianguo Yu, Liping Lu, Ning Xi, Pan Zhou, Renju Feng, Rukang Wang.

Figure 1
Figure 1. Figure 1: Comparison of MoE architectures. (a) MoE models with implicit routing supervision [6], [19]–[21] inherently lack physical interpretability. (b) MoE models relying on explicit routing via manual scenario labels [22], [23] struggle to generalize across open-world environments. (c) Our proposed D3 - MoE framework compresses unbounded scenarios into combinations of finite, orthogonal physical primitives by exp… view at source ↗
Figure 2
Figure 2. Figure 2: Methodological pipeline of D3 -MoE. Left: Offline multi-style expert reference trajectory augmentation. Middle: The core denoising architecture with dynamic routing and independent lateral/longitudinal DiT experts. Right: The behavioral decoupling paradigm for parallel multi-style trajectory generation and their subsequent downstream selection, which features a user-centric mode driven by language instruct… view at source ↗
Figure 3
Figure 3. Figure 3: Distributions of the stylized reference trajectories. (a) Spatial density heatmaps of the different stylic trajectories. (b) Probability density distributions for speed, acceleration, and jerk. Aggressive trajectories shift toward higher speed, acceleration, and jerk, while Conservative ones concen￾trate at lower magnitudes, confirming kinematically distinct styles. (TTC), Driveable Area Compliance (DAC), … view at source ↗
Figure 4
Figure 4. Figure 4: Trajectory features across three styles on the [PITH_FULL_IMAGE:figures/full_fig_p006_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Visualization of style-controllable trajectories. Each urban scenario pairs a BEV map (left) with a corresponding front-view image (right). [PITH_FULL_IMAGE:figures/full_fig_p007_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Visualization of all longitudinal–lateral expert pairings at an intersection. [PITH_FULL_IMAGE:figures/full_fig_p007_6.png] view at source ↗
read the original abstract

Traditional end-to-end autonomous driving frameworks frequently suffer from the "style-averaging" dilemma when trained on high-variance human demonstrations, yielding homogenized, style-uncontrollable, and even kinematically unsafe policies. To overcome this limitation, we present D$^3$-MoE (Dual Disentangled Diffusion Mixture-of-Experts), which disentangles trajectory modeling along two complementary axes. On the behavioral axis, generation is decoupled from selection: a style-conditioned diffusion process synthesizes multi-style candidate trajectories in parallel within a single scene, allowing a downstream module to select the optimal trajectory based on user preference or an evaluation score. On the physical axis, decoupled longitudinal and lateral routers activate their respective experts during inference time, trained without manual labels using self-supervised targets from orthogonal ground-truth kinematics. These activated experts, architected as Diffusion Transformers (DiT) and equipped with style-conditioned AdaLN and asymmetric lateral-fusion cross-attention, independently predict their corresponding physical state before being reassembled into a unified, kinematically coherent trajectory. Extensive evaluations on the challenging NAVSIM benchmark demonstrate that D$^3$-MoE achieves state-of-the-art planning performance, reaching 88.2 PDMS and 84.3 EPDMS by default. Moreover, our Best-of-Three ensemble strategy effectively broadens the multi-modal solution space, raising performance to 91.3 PDMS and 87.5 EPDMS. Both quantitative and qualitative analyses jointly confirm the framework's advantages in planning quality and style controllability.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper presents D³-MoE, a Dual Disentangled Diffusion Mixture-of-Experts architecture for end-to-end autonomous driving. It disentangles modeling along a behavioral axis (style-conditioned diffusion generating multi-style candidate trajectories for downstream selection) and a physical axis (self-supervised decoupled longitudinal and lateral routers activating DiT experts with style-conditioned AdaLN and asymmetric cross-attention, whose independent predictions are reassembled into a single trajectory). The central claims are resolution of style-averaging, style controllability without manual labels, and SOTA planning performance on NAVSIM (88.2 PDMS / 84.3 EPDMS default; 91.3 / 87.5 with Best-of-Three ensemble).

Significance. If the reassembly produces kinematically coherent and safe trajectories and the reported gains are reproducible, the dual-disentanglement approach could meaningfully advance controllable multi-modal planning by separating style variation from physical axes in a parameter-efficient MoE diffusion framework. The self-supervised router training from orthogonal kinematics is a notable design choice that avoids manual labeling.

major comments (2)
  1. [Abstract / physical-axis description] Abstract / physical-axis description: the claim that independently generated longitudinal and lateral diffusion trajectories can be reassembled into 'kinematically coherent' outputs rests on the unstated assumption that separate diffusion processes plus router gating are sufficient. No consistency loss, kinematic projection, or joint refinement step is described; this assumption is load-bearing for both the PDMS/EPDMS claims and the safety of the resulting policies.
  2. [Experimental evaluation] Experimental evaluation: the abstract states specific benchmark numbers (88.2 PDMS, 84.3 EPDMS, etc.) and 'extensive evaluations,' yet the manuscript supplies no experimental protocol, baseline comparisons, ablation studies, or error analysis, rendering the central performance claims unevaluable from the provided text.
minor comments (1)
  1. Define all acronyms (PDMS, EPDMS, DiT, AdaLN, MoE) on first use and ensure consistent notation for the two disentanglement axes throughout.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments on the physical-axis reassembly and experimental details. We address each point below with honest clarifications based on the manuscript content and indicate planned revisions.

read point-by-point responses

  1. Referee: [Abstract / physical-axis description] Abstract / physical-axis description: the claim that independently generated longitudinal and lateral diffusion trajectories can be reassembled into 'kinematically coherent' outputs rests on the unstated assumption that separate diffusion processes plus router gating are sufficient. No consistency loss, kinematic projection, or joint refinement step is described; this assumption is load-bearing for both the PDMS/EPDMS claims and the safety of the resulting policies.

    Authors: The manuscript relies on the self-supervised router training from orthogonal ground-truth kinematics (longitudinal vs. lateral) to ensure the independent DiT expert predictions remain compatible when reassembled by direct concatenation of states. No consistency loss, projection, or joint refinement is present because the orthogonal decomposition and asymmetric cross-attention are designed to preserve physical consistency without additional terms. We agree the reassembly step is described too briefly in the abstract and method overview. We will expand the physical-axis section with an explicit reassembly procedure, a diagram of state concatenation, and a short discussion of why kinematic coherence holds under the orthogonal training regime. revision: yes

  2. Referee: [Experimental evaluation] Experimental evaluation: the abstract states specific benchmark numbers (88.2 PDMS, 84.3 EPDMS, etc.) and 'extensive evaluations,' yet the manuscript supplies no experimental protocol, baseline comparisons, ablation studies, or error analysis, rendering the central performance claims unevaluable from the provided text.

    Authors: The full manuscript contains Section 4 (Experiments) that specifies the NAVSIM evaluation protocol, data preprocessing, metric definitions, full baseline tables with comparisons to prior methods, ablation studies isolating the dual disentanglement and MoE routers, and both quantitative error breakdowns and qualitative trajectory visualizations. The abstract only summarizes headline numbers. If the review text was limited to the abstract excerpt, the complete paper already supplies the requested protocol and analyses. We will add a one-paragraph summary of the experimental setup to the abstract and ensure all tables are cross-referenced in the main text for easier navigation. revision: partial

Circularity Check

0 steps flagged

No significant circularity in claimed derivation chain

full rationale

The paper describes an architectural method (dual disentanglement along behavioral and physical axes, self-supervised routers from orthogonal kinematics, DiT experts with AdaLN and cross-attention, reassembly into trajectories) whose performance is reported via benchmark evaluation on NAVSIM rather than any first-principles derivation or prediction. No equations, uniqueness theorems, or fitted-parameter renamings appear that would reduce the reported PDMS/EPDMS scores or style controllability to inputs by construction. The self-supervised targets and reassembly are presented as design choices whose validity is assessed empirically, not as a closed loop equivalent to the inputs. This is the normal case of an empirical ML systems paper whose central claims remain independent of the listed circularity patterns.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract supplies no explicit free parameters, axioms, or invented entities; insufficient detail to populate the ledger.

pith-pipeline@v0.9.1-grok · 5831 in / 1039 out tokens · 27121 ms · 2026-06-28T05:59:33.513334+00:00 · methodology

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

39 extracted references · 21 canonical work pages · 8 internal anchors

  1. [1]

    Bevformer: learning bird’s-eye-view representation from lidar-camera via spatiotemporal transformers,

    Z. Li, W. Wang, H. Li, E. Xie, C. Sima, T. Lu, Q. Yu, and J. Dai, “Bevformer: learning bird’s-eye-view representation from lidar-camera via spatiotemporal transformers,”IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024

    2024

  2. [2]

    Multi-agent tra- jectory prediction with difficulty-guided feature enhancement network,

    G. Xin, D. Chu, L. Lu, Z. Deng, Y . Lu, and X. Wu, “Multi-agent tra- jectory prediction with difficulty-guided feature enhancement network,” IEEE Robotics and Automation Letters, 2025

    2025

  3. [3]

    Safe and stylized trajectory planning for autonomous driving via diffusion model,

    S. Pei, Y . Wang, Y . Zhu, C. Sun, Q. Li, Y . Zhao, and H. Tan, “Safe and stylized trajectory planning for autonomous driving via diffusion model,”arXiv preprint arXiv:2602.04329, 2026

    work page arXiv 2026

  4. [4]

    Planning-oriented autonomous driving,

    Y . Hu, J. Yang, L. Chen, K. Li, C. Sima, X. Zhu, S. Chai, S. Du, T. Lin, W. Wang,et 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

  5. [5]

    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

  6. [6]

    Artemis: Autoregressive end-to-end trajectory planning with mixture of experts for autonomous driving,

    R. Feng, N. Xi, D. Chu, R. Wang, Z. Deng, A. Wang, L. Lu, J. Wang, and Y . Huang, “Artemis: Autoregressive end-to-end trajectory planning with mixture of experts for autonomous driving,”IEEE Robotics and Automation Letters, vol. 11, no. 1, pp. 226–233, 2025

    2025

  7. [7]

    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

  8. [8]

    A unified candidate set with scene-adaptive refinement via diffusion for end-to-end autonomous driving,

    Z. Wu, S. Pan, S. Chen, S. Yang, and Y . Huang, “A unified candidate set with scene-adaptive refinement via diffusion for end-to-end autonomous driving,”arXiv preprint arXiv:2602.03112, 2026

    work page arXiv 2026

  9. [9]

    Unleashing the Potential of Diffusion Models for End-to-End Autonomous Driving

    Y . Zheng, T. Tan, B. Huang, E. Liu, R. Liang, J. Zhang, J. Cui, G. Chen, K. Ma, H. Ye,et al., “Unleashing the potential of diffusion models for end-to-end autonomous driving,”arXiv preprint arXiv:2602.22801, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  10. [10]

    Implicit behavioral cloning,

    P. Florence, C. Lynch, A. Zeng, O. A. Ramirez, A. Wahid, L. Downs, A. Wong, J. Lee, I. Mordatch, and J. Tompson, “Implicit behavioral cloning,” inConference on robot learning. PMLR, 2022, pp. 158–168

    2022

  11. [11]

    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

  12. [12]

    Drivefine: Refining-augmented masked diffusion vla for precise and robust driving,

    C. Dang, S. Ang, Y . Li, H. Tian, J. Wang, G. Li, H. Ye, J. Ma, L. Chen, and Y . Wang, “Drivefine: Refining-augmented masked diffusion vla for precise and robust driving,”arXiv preprint arXiv:2602.14577, 2026

    work page arXiv 2026

  13. [13]

    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. Zhang,et al., “Diffusiondrive: Truncated diffusion model for end-to-end autonomous driving,”arXiv preprint arXiv:2411.15139, 2024

    work page arXiv 2024

  14. [14]

    FeaXDrive: Feasibility-aware Trajectory-Centric Diffusion Planning for End-to-End Autonomous Driving

    B. Wang, Z. Li, M. Liu, X. Zhang, B. Leng, and L. Xiong, “Feaxdrive: Feasibility-aware trajectory-centric diffusion planning for end-to-end autonomous driving,”arXiv preprint arXiv:2604.12656, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  15. [15]

    Diffusiondrivev2: Reinforcement learning-constrained truncated diffusion modeling in end-to-end autonomous driving.arXiv preprint arXiv:2512.07745, 2025

    J. Zou, S. Chen, B. Liao, Z. Zheng, Y . Song, L. Zhang, Q. Zhang, W. Liu, and X. Wang, “Diffusiondrivev2: Reinforcement learning-constrained truncated diffusion modeling in end-to-end autonomous driving,”arXiv preprint arXiv:2512.07745, 2025

    work page arXiv 2025

  16. [16]

    Guideflow: Constraint-guided flow matching for planning in end-to-end autonomous driving,

    L. Liu, C. Jia, G. Yu, Z. Song, J. Li, F. Jia, P. Wu, X. Hao, and Y . Luo, “Guideflow: Constraint-guided flow matching for planning in end-to-end autonomous driving,”arXiv preprint arXiv:2511.18729, 2025

    work page arXiv 2025

  17. [17]

    Trajdiff: End-to-end autonomous driving without perception annotation,

    X. Gui, J. Zhao, W. Han, J. Wang, J. Gong, F. Tan, C.-z. Xu, and J. Shen, “Trajdiff: End-to-end autonomous driving without perception annotation,”arXiv preprint arXiv:2512.00723, 2025

    work page arXiv 2025

  18. [18]

    DIVER: Reinforced Diffusion Breaks Imitation Bottlenecks in End-to-End Autonomous Driving

    Z. Song, L. Liu, H. Pan, B. Liao, M. Guo, L. Yang, Y . Zhang, S. Xu, C. Jia, and Y . Luo, “Diver: Reinforced diffusion breaks imi- tation bottlenecks in end-to-end autonomous driving,”arXiv preprint arXiv:2507.04049, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  19. [19]

    Pie: Perception and interaction enhanced end-to-end motion planning for autonomous driving,

    C. Yuan, Z. Lu, Z. Zhang, Y . Zhao, Z. Huang, S. Sun, J. Sun, J. Li, C. D. W. Lee, D. Li,et al., “Pie: Perception and interaction enhanced end-to-end motion planning for autonomous driving,”arXiv preprint arXiv:2509.18609, 2025

    work page arXiv 2025

  20. [20]

    Generalizing motion planners with mixture of experts for autonomous driving,

    Q. Sun, H. Wang, J. Zhan, F. Nie, X. Wen, L. Xu, K. Zhan, P. Jia, X. Lang, and H. Zhao, “Generalizing motion planners with mixture of experts for autonomous driving,” in2025 IEEE International Conference on Robotics and Automation (ICRA). IEEE, 2025, pp. 6033–6039

    2025

  21. [21]

    Trajmoe: Scene-adaptive trajectory planning with mixture of experts and reinforcement learning,

    Z. Xing, P. Yang, L. Wang, Y . Zhang, Y . Hu, Y . Zheng, J. Wang, Y . Gao, G. Li, K. Ma,et al., “Trajmoe: Scene-adaptive trajectory planning with mixture of experts and reinforcement learning,”arXiv preprint arXiv:2512.07135, 2025

    work page arXiv 2025

  22. [22]

    Geminus: Dual-aware global and scene-adaptive mixture-of-experts for end-to-end autonomous driving,

    C. Wan, Y . Cui, J. Du, S. Yang, Y . Bai, P. Yi, N. Li, and Y . Huang, “Geminus: Dual-aware global and scene-adaptive mixture-of-experts for end-to-end autonomous driving,”arXiv preprint arXiv:2507.14456, 2025

    work page arXiv 2025

  23. [23]

    Mose: Skill-by-skill mixture-of-experts learning for embodied autonomous machines,

    L. Xu, J. Yu, X. Peng, Y . Chen, W. Li, J. Yoo, S. Chunag, D. Lee, D. Ji, and C. Zhang, “Mose: Skill-by-skill mixture-of-experts learning for embodied autonomous machines,”arXiv preprint arXiv:2507.07818, 2025

    work page arXiv 2025

  24. [24]

    Expertad: Enhancing autonomous driving systems with mixture of experts,

    H. Jiang, X. Huang, Y . Lu, D. Wang, Y . Cao, C. Sha, B. Chen, K. Chen, and X. Peng, “Expertad: Enhancing autonomous driving systems with mixture of experts,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 7, 2026, pp. 5378–5387

    2026

  25. [25]

    Feature pyramid networks for object detection,

    T.-Y . Lin, P. Doll´ar, R. Girshick, K. He, B. Hariharan, and S. Belongie, “Feature pyramid networks for object detection,” inProceedings of the IEEE conference on computer vision and pattern recognition, 2017, pp. 2117–2125

    2017

  26. [26]

    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,et al., “Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking,”Ad- vances in Neural Information Processing Systems, vol. 37, pp. 28 706– 28 719, 2024

    2024

  27. [27]

    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

    work page arXiv 2025

  28. [28]

    VADv2: End-to-End Vectorized Autonomous Driving via Probabilistic Planning

    S. Chen, B. Jiang, H. Gao, B. Liao, Q. Xu, Q. Zhang, C. Huang, W. Liu, and X. Wang, “Vadv2: End-to-end vectorized autonomous driving via probabilistic planning,”arXiv preprint arXiv:2402.13243, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  29. [29]

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

    X. Weng, B. Ivanovic, Y . Wang, Y . Wang, and M. Pavone, “Para- drive: Parallelized architecture for real-time autonomous driving,” in Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2024, pp. 15 449–15 458

    2024

  30. [30]

    Drivex: Omni scene modeling for learning generalizable world knowledge in autonomous driving,

    C. Shi, S. Shi, K. Sheng, B. Zhang, and L. Jiang, “Drivex: Omni scene modeling for learning generalizable world knowledge in autonomous driving,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 28 599–28 609

    2025

  31. [31]

    Enhancing End-to-End Autonomous Driving with Latent World Model

    Y . Li, L. Fan, J. He, Y . Wang, Y . Chen, Z. Zhang, and T. Tan, “Enhancing end-to-end autonomous driving with latent world model,”arXiv preprint arXiv:2406.08481, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  32. [32]

    FutureSightDrive: Thinking Visually with Spatio-Temporal CoT for Autonomous Driving

    S. Zeng, X. Chang, M. Xie, X. Liu, Y . Bai, Z. Pan, M. Xu, X. Wei, and N. Guo, “Futuresightdrive: Thinking visually with spatio-temporal cot for autonomous driving,”arXiv preprint arXiv:2505.17685, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  33. [33]

    NoRD: A Data-Efficient Vision-Language-Action Model that Drives without Reasoning

    I. Rawal, S. Gupta, Y . Hu, and W. Zhan, “Nord: A data-efficient vision- language-action model that drives without reasoning,”arXiv preprint arXiv:2602.21172, 2026

    work page internal anchor Pith review Pith/arXiv arXiv 2026

  34. [34]

    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. Long,et 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

  35. [35]

    Distilldrive: End-to- end multi-mode autonomous driving distillation by isomorphic hetero- source planning model,

    R. Yu, X. Zhang, R. Zhao, H. Yan, and M. Wang, “Distilldrive: End-to- end multi-mode autonomous driving distillation by isomorphic hetero- source planning model,” inProceedings of the IEEE/CVF International Conference on Computer Vision, 2025, pp. 26 188–26 197

    2025

  36. [36]

    Prix: Learning to plan from raw pixels for end-to-end autonomous driving,

    M. Wozniak, L. Liu, Y . Cai, and P. Jensfelt, “Prix: Learning to plan from raw pixels for end-to-end autonomous driving,”IEEE Robotics and Automation Letters, vol. 11, no. 5, pp. 6400–6407, 2026

    2026

  37. [37]

    Drivesuprim: Towards precise trajectory selection for end-to-end plan- ning,

    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 plan- ning,” inProceedings of the AAAI Conference on Artificial Intelligence, vol. 40, no. 14, 2026, pp. 11 910–11 918

    2026

  38. [38]

    ReCogDrive: A Reinforced Cognitive Framework for End-to-End Autonomous Driving

    Y . Li, K. Xiong, X. Guo, F. Li, S. Yan, G. Xu, L. Zhou, L. Chen, H. Sun, B. Wang,et al., “Recogdrive: A reinforced cognitive framework for end- to-end autonomous driving,”arXiv preprint arXiv:2506.08052, 2025

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  39. [39]

    ipad: Iterative proposal-centric end-to-end autonomous driving

    K. Guo, H. Liu, X. Wu, J. Pan, and C. Lv, “ipad: Itera- tive proposal-centric end-to-end autonomous driving,”arXiv preprint arXiv:2505.15111, 2025

    work page arXiv 2025