Pith. sign in

REVIEW 4 major objections 6 minor 50 references

DistillDrive: End-to-End Multi-Mode Autonomous Driving Distillation by Isomorphic Hetero-Source Planning Model

T0 review · 4 major / 6 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read DistillDrive claims that distilling a structured multi-mode planner into an end-to-end driving model, with RL state optimization and generative latent-space interaction, cuts collision rate by ~50% and lifts closed-loop scores by ~3 points.

desk verdict A promising combination of distillation, RL, and generative modeling for driving planning, but the headline gains are not cleanly isolated from temporal backbone and baseline choices. read the letter →

arxiv 2508.05402 v1 pith:ZKBVFRAO submitted 2025-08-07 cs.RO cs.CV

Rui Yu , Xianghang Zhang , Runkai Zhao , Huaicheng Yan , Meng Wang This is my paper
classification cs.ROcs.CV
keywords autonomousdrivingend-to-endplanningknowledgedistillationmulti-modereinforcementlearninggenerativemodelingnuScenesNAVSIM
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

The paper claims that the gap between annotation-based planners and end-to-end drivers can be closed by changing what the end-to-end model imitates: instead of one expert trajectory, it learns from a structured teacher planner's multi-mode planning instances, so it must represent several feasible behaviors at once. Two supporting ideas matter as much as the setup: a reinforcement-learning loss that explicitly ties predicted ego status to driving decisions, and a generative decoder that models expert trajectories and instance features as Gaussian distributions and supervises their interaction in latent space. If the claims are right, camera-only end-to-end driving can reach collision rates roughly half those of the previous best end-to-end baseline, and planners trained on clean annotation input can transfer their planning skill, not just their perception features, to models that see raw sensors. That matters because closed-loop safety — not just open-loop trajectory error — is the bottleneck for deploying end-to-end systems.

What carries the argument

Multi-stage distillation between two isomorphic planning models: a teacher consuming annotated vectorized scenes and a camera-only student, linked by an MLP adapter so decoder features live in the same space, with the teacher's planning vocabulary supplying diverse multi-mode targets. Around it sit two auxiliary mechanisms. First, a reinforcement-learning status optimization: predicted ego status feeds a learnable reward comparing predictions with expert status and trajectories, and a Q-learning-style target maps statuses to discrete decisions (left/right/straight), tying state space to decision space. Second, a generative interaction module: expert trajectories and agent/ego instances are e

What would settle it

On a held-out batch of nuScenes val scenes, collect the latent codes the trained generative encoder produces for expert trajectories and instance tokens, then fit and inspect their empirical distributions (a Gaussianity test or a density comparison against $N(\mu_F,\sigma_F)$ and $N(\mu_I,\sigma_I)$). If the codes are clearly multimodal, skewed, or heavy-tailed, Eq. (6) optimizes a distance that does not match the data, and the 50% collision reduction should shrink or vanish when the KL term is replaced by a distribution-agnostic divergence; a cheaper check is retraining with a non-parametric

Watch

Extended reading notes

Core claim

End-to-end planning can be supervised by a teacher's multi-mode instances — different feasible driving behaviors — instead of one expert trajectory. The teacher, trained on annotated scenes with imitation and reinforcement learning, is aligned to the camera-only student at encoder, decoder, and multi-mode classification levels. Two mechanisms carry the rest: an RL loss linking predicted ego states to discrete decisions, and a generative decoder that models expert and instance features as Gaussians, supervised by a closed-form KL loss. Reported gains: average collision rate falls from 0.13% to 0.06% on nuScenes versus SparseDrive, and closed-loop PDMS on NAVSIM rises about 2.5 points over Tra

Load-bearing premise

The load-bearing premise is that expert trajectories and agent/ego instance encodings are Gaussian in latent space — if the true distributions are multimodal or heavy-tailed, the closed-form KL loss (Eq. 6) measures the wrong geometry and the generative interaction may not confer the intended motion prior.

Editorial extensions

If this is right

  • A camera-only end-to-end planner can match or exceed the safety of LiDAR or perception-ground-truth baselines, cutting average collision rate from 0.13% to 0.06% on nuScenes.
  • Multi-mode instance distillation transfers trajectory diversity: the student produces a richer set of candidate trajectories in turning scenarios than single-trajectory imitation, as the qualitative comparisons illustrate.
  • The RL status optimization offers a concrete remedy for ego-status leakage, keeping the state-to-decision mapping explicit while reducing over-reliance on ego status features.
  • The student surpasses the teacher on some closed-loop metrics (NC, TTC), suggesting distillation can smooth away teacher imperfections rather than merely copying them.

Reading between the lines

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

  • A direct test of the architecture's core bet: estimate the empirical latent distributions of the generative encoder on held-out nuScenes scenes; if they are visibly non-Gaussian, the closed-form KL loss in Eq. (6) is fitting the wrong geometry, and the gains attributed to distribution interaction should be re-attributed to generic multi-mode supervision.
  • Because the adapter renders teacher and student decoder spaces isomorphic, the same distillation should transfer across sensor suites (e.g., LiDAR teacher to camera student, or surround-view to monocular) without retraining the teacher; this is a testable extension the paper does not run.
  • The paper's ablation weights (lambda2, lambda3 near no effect; lambda1 dominant) suggest the generative and RL modules may be secondary to multi-mode distillation; a stripped-down variant with only instance distillation would isolate how much of the 50% collision reduction actually requires the generative interaction.
  • The supplementary reports that RL optimization gave no significant closed-loop gain on NAVSIM because Transfuser's explicit ego-status input conflicts with implicit status learning — a stated boundary: the RL claim currently rests on the nuScenes ablation rather than on both benchmarks.
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 / 6 minor

Summary. The paper proposes DistillDrive, an end-to-end autonomous driving model that distills knowledge from a teacher planning model trained on structured scene annotations. The method combines three proposed components: multi-mode instance distillation (encoder, decoder, and classification losses), reinforcement-learning-based state-to-decision optimization (a weighted exponential reward and Q-learning), and a generative distribution-level interaction in latent space using KL divergence between Gaussian encodings of expert trajectories and agent/ego instances. Experiments are reported on nuScenes open-loop planning and on the NAVSIM closed-loop benchmark. The headline claims are a roughly 50% reduction in collision rate versus SparseDrive on nuScenes and a 3-point improvement in closed-loop EP and PDMS over an unnamed baseline on NAVSIM, with code and models promised publicly.

Significance. If the reported gains are robust and correctly attributed, the work would make a useful contribution to end-to-end planning, since it explicitly targets multi-modality of planning, ego-state leakage, and motion-guided latent interactions, and it provides both open- and closed-loop evaluations. The idea of using a teacher model trained on privileged annotations to supervise a camera/LiDAR student through instance-level and distribution-level distillation is well aligned with current trends in knowledge distillation for driving. The paper also includes reproducibility-oriented elements: it compares against replicated SparseDrive under the same settings and commits to releasing code and models. However, the significance is currently conditional on resolving several load-bearing experimental and formal issues, most notably the incomplete ablation section and the ambiguous identity of the closed-loop baseline.

major comments (4)
  1. [Sec. 4.4 and Supp. Sec. 7.1] The main-text ablation section is truncated: it begins to describe Tab. 5 and breaks off mid-sentence after “the proposed linear weighted reinforcement learn-”. The referenced Tab. 5 is not present, and the supplementary Tab. 9, which is said to show the NAVSIM ablations, is not included either. Moreover, no ablation reports a control student trained with only the imitation loss L_IL and none of the three proposed components (no distillation, no distribution loss, no RL loss) on the same backbone. Without this control, the reported gains cannot be attributed to the proposed modules. This is a central issue, since the paper’s thesis is that these modules cause the improvements. The authors should provide complete ablation tables for both nuScenes and NAVSIM, including a no-proposed-components baseline.
  2. [Abstract and Sec. 4.3, Table 2] The abstract claims a “3-point improvement in closed-loop performance” but does not name the baseline. Against the Transfuser baseline in Table 2, DistillDrive improves EP by 2.4 and PDMS by 2.7 absolute points, not 3 points. Against VADv2 or Hydra-MDP the improvement is larger than 3 points, but the abstract and Sec. 4.3 do not specify which comparison is meant. The text in Sec. 4.3 also states a “2.5%” PDMS improvement over Transfuser, which does not match the table value of 2.7 absolute points (or 3.2% relative). The baseline used for the headline claim must be named unambiguously, and the numbers should be reconciled.
  3. [Sec. 4.2, Sec. 3.3, Supp. Sec. 6.1] There is a potential temporal-modeling confound in the NAVSIM comparison. Sec. 4.2 says the model uses Transfuser settings “without temporal training to ensure a fair comparison,” but Sec. 3.3 states that the student retains SparseDrive’s fully sparse temporal representation, and the teacher uses a temporal decoder and memory bank (Sec. 3.2 and Supp. Sec. 6.1). If the student’s NAVSIM implementation includes temporal fusion while the Transfuser baseline does not, the closed-loop gains could stem from temporal information rather than the proposed distillation, distribution interaction, or RL optimization. This must be clarified: specify exactly which temporal modules are present in the student at NAVSIM test time, and what “without temporal training” means for the student versus the baseline.
  4. [Eq. (6)] The distribution loss LDS is not consistent with a standard KL divergence between two Gaussians. For N(μ_F,σ_F) and N(μ_I,σ_I), the KL divergence contains log(σ_I/σ_F) and quadratic terms divided by the appropriate variance. Eq. (6) instead writes σ_I − σ_F − 0.5 + e^{2σ_F} + (μ_F−μ_I)^2/(2e^{2σ_I}), which is not a standard KL for σ interpreted either as a standard deviation or as a log-standard-deviation. Since LDS is one of the three core proposed losses, this point needs a corrected formula or a clear notational explanation. In addition, the paper provides no empirical check of the Gaussian-latent assumption on which Eq. (6) and the generative interaction in Sec. 3.3 rely.
minor comments (6)
  1. [Abstract] Typo: “prcocess” should be “process.”
  2. [Fig. 1] “Preception” should be “Perception” in the figure caption.
  3. [Throughout] The dataset name is inconsistently typeset as “NAVSIM” and “NA VSIM”; choose one form.
  4. [Eqs. (4) and (24)] The sum notation “\bar{a}X” is unclear; it should be written as a sum over actions, e.g., ∑_{i=1}^{N_a} or ∑_{\bar{a}}, so that the dimension of the sum is explicit.
  5. [References] Several reference entries contain unusual trailing page numbers (e.g., “5, 6, 3” for Caesar et al., “6, 2, 4” for Dauner et al.); these appear to be formatting errors and should be cleaned.
  6. [Sec. 4.4 / Supp. Sec. 7.1] The text refers to Tab. 5, Tab. 9, Fig. 10, Fig. 11, and Fig. 12, but these items are not included in the provided manuscript. The final version must include all cited tables and figures.

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: DistillDrive's claims are empirical measurements against external baselines; no load-bearing step reduces to its own input or to a load-bearing self-citation.

full rationale

I walked the derivation chain from the teacher model (Sec. 3.2) through the student model (Sec. 3.3), the distillation losses (Sec. 3.4), and the total objective (Sec. 3.5), and then checked the experimental tables and appendices. The paper's central claims — ~50% collision reduction and ~3-point closed-loop EP/PDMS gains — are closed/open-loop benchmark measurements against SparseDrive, Transfuser, and other external baselines. None of the reported evaluation quantities is a fitted parameter renamed as a prediction: the distillation, distribution, and RL losses (Eqs. 6-11) are training objectives, and they do not by construction equal the L2, collision, EP, or PDMS metrics used for evaluation. The teacher-student setup is standard knowledge distillation; the fact that the teacher is trained on the same expert trajectories that the student imitates is a property of the method, not a definitional reduction of the measured outcome to the training signal. The self-citations in the paper (refs. 25, 39, 40, 43) appear only in related-work context and are not load-bearing for the main claim. The NAVSIM comparison may be confounded because the student retains a temporal backbone while the Transfuser baseline is deliberately trained without temporal modeling, and the supplementary text (Sec. 7.1) even admits that RL optimization did not yield significant gains on NAVSIM. That is a causal-attribution or experimental-validity concern, not circularity under the specific definitions in scope. I therefore find no circular step that can be exhibited as Eq. X = Eq. Y by construction or as a fitted parameter called a prediction, and assign score 0.

Assumptions & free parameters 5 free parameters · 4 assumptions · 0 invented entities

The central claim rests on several unvalidated modeling assumptions, chiefly the Gaussian latent-space assumption and the transferability of annotation-based teacher features to a camera-based student. The hyperparameters λ1, λ2, λ3, γ, ζ and the learnable reward weights add flexibility but are not derived. No new physical entities are introduced.

free parameters (5)
  • λ1, λ2, λ3 = 0.5, 1, 1
    Loss weights in Eq. (11); sensitivity shown in Fig. 10 but no theoretical basis.
  • γ = 0.95
    Discount factor in Eq. (3); set by default, not ablated.
  • ζ = 0.5
    Coefficient for Gaussian noise sampling in Sec. 3.3.
  • ω_i = learned
    Learnable weights in reward Eq. (23); no explicit training objective given.
  • speed threshold = 20
    Threshold in espeed (Eq. 21) to decide out-of-bounds speed.
assumptions (4)
  • domain assumption Expert trajectories and learned instance encodings are approximately Gaussian in latent space.
    Stated in Sec. 3.3; underpins the KL-divergence distribution loss (Eq. 6) and the generative interaction, but no evidence is given that the data are Gaussian.
  • domain assumption The teacher model's ground-truth instance annotations provide a supervision signal that can be transferred to a camera-based student without loss of planning relevance.
    Assumed throughout the distillation design (Sec. 3.4); the adapter ψ is intended to bridge domain shift but its sufficiency is not proven.
  • domain assumption Open-loop collision rate on nuScenes and closed-loop metrics on NAVSIM reflect the safety of the planner in real driving.
    The paper relies on these metrics as evidence of safety (Tab. 1, 2), which is known to be controversial, as the paper itself cites [27] on ego-status leakage.
  • standard math Standard math: KL divergence, attention, PointNet, GRU operations are used correctly.
    These are standard and not contested.

how reviews work

0 comments
Cite this review

Pith. "Pith review of DistillDrive: End-to-End Multi-Mode Autonomous Driving Distillation by Isomorphic Hetero-Source Planning Model." pith.science (2026). https://pith.science/paper/ZKBVFRAO

@misc{pith2026250805402,
  author       = {Pith},
  title        = {Pith review of: DistillDrive: End-to-End Multi-Mode Autonomous Driving Distillation by Isomorphic Hetero-Source Planning Model},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/ZKBVFRAO}},
  note         = {Machine review of arXiv:2508.05402}
}
read the original abstract

End-to-end autonomous driving has been recently seen rapid development, exerting a profound influence on both industry and academia. However, the existing work places excessive focus on ego-vehicle status as their sole learning objectives and lacks of planning-oriented understanding, which limits the robustness of the overall decision-making prcocess. In this work, we introduce DistillDrive, an end-to-end knowledge distillation-based autonomous driving model that leverages diversified instance imitation to enhance multi-mode motion feature learning. Specifically, we employ a planning model based on structured scene representations as the teacher model, leveraging its diversified planning instances as multi-objective learning targets for the end-to-end model. Moreover, we incorporate reinforcement learning to enhance the optimization of state-to-decision mappings, while utilizing generative modeling to construct planning-oriented instances, fostering intricate interactions within the latent space. We validate our model on the nuScenes and NAVSIM datasets, achieving a 50\% reduction in collision rate and a 3-point improvement in closed-loop performance compared to the baseline model. Code and model are publicly available at https://github.com/YuruiAI/DistillDrive

Discussion (0). Sign in to comment.

Reference graph

Works this paper leans on

50 extracted references · 37 canonical work pages

  1. [1]

    nuscenes: A multi- modal dataset for autonomous driving

    Holger Caesar, Varun Bankiti, Alex H Lang, Sourabh V ora, Venice Erin Liong, Qiang Xu, Anush Krishnan, Yu Pan, Gi- ancarlo Baldan, and Oscar Beijbom. nuscenes: A multi- modal dataset for autonomous driving. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 11621–11631, 2020. 5, 6, 3

    work page 2020

  2. [2]

    nuplan: A closed-loop ml-based plan- ning benchmark for autonomous vehicles

    Holger Caesar, Juraj Kabzan, Kok Seang Tan, Whye Kit Fong, Eric Wolff, Alex Lang, Luke Fletcher, Oscar Beijbom, and Sammy Omari. nuplan: A closed-loop ml-based plan- ning benchmark for autonomous vehicles. arXiv preprint arXiv:2106.11810, 2021. 5

    arXiv 2021

  3. [3]

    End-to-end autonomous driving: Challenges and frontiers

    Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, An- dreas Geiger, and Hongyang Li. End-to-end autonomous driving: Challenges and frontiers. IEEE Transactions on Pat- tern Analysis and Machine Intelligence, 2024. 2

    work page 2024

  4. [4]

    Vadv2: End-to-end vectorized autonomous driving via probabilistic planning

    Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized autonomous driving via probabilistic planning. arXiv preprint arXiv:2402.13243,

    arXiv

  5. [5]

    Pluto: Push- ing the limit of imitation learning-based planning for au- tonomous driving

    Jie Cheng, Yingbing Chen, and Qifeng Chen. Pluto: Push- ing the limit of imitation learning-based planning for au- tonomous driving. arXiv preprint arXiv:2404.14327, 2024. 1

    arXiv 2024

  6. [6]

    Transfuser: Imitation with transformer-based sensor fusion for autonomous driv- ing

    Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imitation with transformer-based sensor fusion for autonomous driv- ing. IEEE Transactions on Pattern Analysis and Machine Intelligence, 45(11):12878–12895, 2022. 2, 5, 6, 3

    work page 2022

  7. [7]

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

    Dauner, et al. Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking. Advances in Neural Information Processing Systems, 37:28706–28719, 2025. 5, 6, 2, 4

    work page 2025

  8. [8]

    Carla: An open urban driving simulator

    Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. In Conference on Robot Learning , pages 1–16. PMLR, 2017. 2

    work page 2017

Show all 50 references
  1. [9]

    On the road to portability: Compressing end- to-end motion planner for autonomous driving

    Kaituo Feng, Changsheng Li, Dongchun Ren, Ye Yuan, and Guoren Wang. On the road to portability: Compressing end- to-end motion planner for autonomous driving. In Proceed- ings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 15099–15108, 2024. 1, 2

    2024

  2. [10]

    Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor

    Tuomas Haarnoja, Aurick Zhou, Pieter Abbeel, and Sergey Levine. Soft actor-critic: Off-policy maximum entropy deep reinforcement learning with a stochastic actor. In Interna- tional Conference on Machine Learning , pages 1861–1870. Pmlr, 2018. 2

    2018

  3. [11]

    One-for-all: Bridge the gap between heterogeneous architectures in knowledge distilla- tion

    Zhiwei Hao, Jianyuan Guo, Kai Han, Yehui Tang, Han Hu, Yunhe Wang, and Chang Xu. One-for-all: Bridge the gap between heterogeneous architectures in knowledge distilla- tion. Advances in Neural Information Processing Systems , 36, 2024. 2

    2024

  4. [12]

    Model-based imitation learning for urban driving

    Anthony Hu, Gianluca Corrado, Nicolas Griffiths, Zachary Murez, Corina Gurau, Hudson Yeo, Alex Kendall, Roberto Cipolla, and Jamie Shotton. Model-based imitation learning for urban driving. Advances in Neural Information Process- ing Systems, 35:20703–20716, 2022. 2

    2022

  5. [13]

    Safe local motion planning with self- supervised freespace forecasting

    Peiyun Hu, Aaron Huang, John Dolan, David Held, and Deva Ramanan. Safe local motion planning with self- supervised freespace forecasting. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 12732–12741, 2021. 6

    2021

  6. [14]

    St-p3: End-to-end vision-based au- tonomous driving via spatial-temporal feature learning

    Shengchao Hu, Li Chen, Penghao Wu, Hongyang Li, Junchi Yan, and Dacheng Tao. St-p3: End-to-end vision-based au- tonomous driving via spatial-temporal feature learning. In European Conference on Computer Vision, pages 533–549. Springer, 2022. 1, 2, 6

    2022

  7. [15]

    Planning-oriented autonomous driving

    Yihan Hu, Jiazhi Yang, Li Chen, Keyu Li, Chonghao Sima, Xizhou Zhu, Siqi Chai, Senyao Du, Tianwei Lin, Wenhai Wang, et al. Planning-oriented autonomous driving. In Pro- ceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853–17862, 2023. 1, 2, 6

    2023

  8. [16]

    Driving behavior modeling using naturalistic human driving data with inverse reinforcement learning

    Zhiyu Huang, Jingda Wu, and Chen Lv. Driving behavior modeling using naturalistic human driving data with inverse reinforcement learning. IEEE Transactions on Intelligent Transportation Systems, 23(8):10239–10251, 2021. 2, 4

    2021

  9. [17]

    Con- ditional predictive behavior planning with inverse reinforce- ment learning for human-like autonomous driving

    Zhiyu Huang, Haochen Liu, Jingda Wu, and Chen Lv. Con- ditional predictive behavior planning with inverse reinforce- ment learning for human-like autonomous driving. IEEE Transactions on Intelligent Transportation Systems , 24(7): 7244–7258, 2023. 2

    2023

  10. [18]

    Imitation learning: A survey of learning methods

    Ahmed Hussein, Mohamed Medhat Gaber, Eyad Elyan, and Chrisina Jayne. Imitation learning: A survey of learning methods. ACM Computing Surveys , 50(2):1–35, 2017. 1, 4

    2017

  11. [19]

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

    Xiaosong Jia, Yulu Gao, Li Chen, Junchi Yan, Patrick Langechuan Liu, and Hongyang Li. Driveadapter: Breaking the coupling barrier of perception and planning in end-to-end autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 795...

    2023

  12. [20]

    Vad: Vectorized scene representa- tion for efficient autonomous driving

    Bo Jiang, Shaoyu Chen, Qing Xu, Bencheng Liao, Jiajie Chen, Helong Zhou, Qian Zhang, Wenyu Liu, Chang Huang, and Xinggang Wang. Vad: Vectorized scene representa- tion for efficient autonomous driving. In Proceedings of the IEEE/CVF International Conference on Computer Vision ,...

    2023

  13. [21]

    Learning to drive in a day

    Alex Kendall, Jeffrey Hawke, David Janz, Przemyslaw Mazur, Daniele Reda, John-Mark Allen, Vinh-Dieu Lam, Alex Bewley, and Amar Shah. Learning to drive in a day. In 2019 International Conference on Robotics and Automa- tion, pages 8248–8254. IEEE, 2019. 2

    2019

  14. [22]

    Differentiable raycasting for self-supervised occupancy forecasting

    Tarasha Khurana, Peiyun Hu, Achal Dave, Jason Ziglar, David Held, and Deva Ramanan. Differentiable raycasting for self-supervised occupancy forecasting. InEuropean Con- ference on Computer Vision, pages 353–369. Springer, 2022. 6

    2022

  15. [23]

    The hungarian method for the assignment problem

    Harold W Kuhn. The hungarian method for the assignment problem. Naval Research Logistics Quarterly, 2(1-2):83–97,

  16. [24]

    Conservative q-learning for offline reinforcement learning

    Aviral Kumar, Aurick Zhou, George Tucker, and Sergey Levine. Conservative q-learning for offline reinforcement learning. Advances in Neural Information Processing Sys- tems, 33:1179–1191, 2020. 2

    2020

  17. [25]

    Unimodal-driven distillation in multimodal emotion recogni- tion with dynamic fusion

    Jiagen Li, Rui Yu, Huihao Huang, and Huaicheng Yan. Unimodal-driven distillation in multimodal emotion recogni- tion with dynamic fusion. arXiv preprint arXiv:2503.23721,

    arXiv

  18. [26]

    Hydra-mdp: End-to-end multimodal planning with multi- target hydra-distillation

    Zhenxin Li, Kailin Li, Shihao Wang, Shiyi Lan, Zhiding Yu, Yishen Ji, Zhiqi Li, Ziyue Zhu, Jan Kautz, Zuxuan Wu, et al. Hydra-mdp: End-to-end multimodal planning with multi- target hydra-distillation. arXiv preprint arXiv:2406.06978 ,

    arXiv

  19. [27]

    Zhiqi Li, Zhiding Yu, Shiyi Lan, Jiahan Li, Jan Kautz, Tong Lu, and Jose M Alvarez. Is ego status all you need for open- loop end-to-end autonomous driving? In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 14864–14873, 2024. 2, 4

    2024

  20. [28]

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

    Bencheng Liao, Shaoyu Chen, Haoran Yin, Bo Jiang, Cheng Wang, Sixu Yan, Xinbang Zhang, Xiangyu Li, Ying Zhang, Qian Zhang, et al. Diffusiondrive: Truncated diffusion model for end-to-end autonomous driving. arXiv preprint arXiv:2411.15139, 2024. 1

    2024 arXiv

  21. [29]

    Sparse4d: Multi-view 3d object detec- tion with sparse spatial-temporal fusion

    Xuewu Lin, Tianwei Lin, Zixiang Pei, Lichao Huang, and Zhizhong Su. Sparse4d: Multi-view 3d object detec- tion with sparse spatial-temporal fusion. arXiv preprint arXiv:2211.10581, 2022. 1

    2022 arXiv

  22. [30]

    Ca-w3d: Leveraging context-aware knowledge for weakly supervised monocular 3d detection

    Chupeng Liu, Runkai Zhao, and Weidong Cai. Ca-w3d: Leveraging context-aware knowledge for weakly supervised monocular 3d detection. IEEE/RSJ International Conference on Intelligent Robots and Systems, 2025. 1

    2025

  23. [31]

    Imitation is not enough: Robustifying imitation with reinforcement learning for challenging driving scenarios

    Yiren Lu, Justin Fu, George Tucker, Xinlei Pan, Eli Bron- stein, Rebecca Roelofs, Benjamin Sapp, Brandyn White, Aleksandra Faust, Shimon Whiteson, et al. Imitation is not enough: Robustifying imitation with reinforcement learning for challenging driving scenarios. In 2023 IEEE...

    2023

  24. [32]

    Msd: Multi-self-distillation learning via multi-classifiers within deep neural networks

    Yunteng Luan, Hanyu Zhao, Zhi Yang, and Yafei Dai. Msd: Multi-self-distillation learning via multi-classifiers within deep neural networks. arXiv preprint arXiv:1911.09418 ,

    1911 arXiv

  25. [33]

    Human-level control through deep reinforcement learn- ing

    V olodymyr Mnih, Koray Kavukcuoglu, David Silver, An- drei A Rusu, Joel Veness, Marc G Bellemare, Alex Graves, Martin Riedmiller, Andreas K Fidjeland, Georg Ostrovski, et al. Human-level control through deep reinforcement learn- ing. nature, 518(7540):529–533, 2015. 4, 2

    2015

  26. [34]

    Asynchronous methods for deep reinforcement learning

    V olodymyr Mnih, Adria Puigdomenech Badia, Mehdi Mirza, Alex Graves, Timothy Lillicrap, Tim Harley, David Silver, and Koray Kavukcuoglu. Asynchronous methods for deep reinforcement learning. In International Conference on Ma- chine Learning, pages 1928–1937. PmLR, 2016. 2

    1928

  27. [35]

    Pointnet: Deep learning on point sets for 3d classification and segmentation

    Charles R Qi, Hao Su, Kaichun Mo, and Leonidas J Guibas. Pointnet: Deep learning on point sets for 3d classification and segmentation. In Proceedings of the IEEE/CVF Con- ference on Computer Vision and Pattern Recognition, pages 652–660, 2017. 3

    2017

  28. [36]

    Sparsedrive: End-to-end au- tonomous driving via sparse scene representation

    Wenchao Sun, Xuewu Lin, Yining Shi, Chuang Zhang, Hao- ran Wu, and Sifa Zheng. Sparsedrive: End-to-end au- tonomous driving via sparse scene representation. arXiv preprint arXiv:2405.19620, 2024. 1, 2, 4, 5, 6, 7, 3

    2024 arXiv

  29. [37]

    Motion planning for autonomous driv- ing: The state of the art and future perspectives.IEEE Trans- actions on Intelligent Vehicles, 8(6):3692–3711, 2023

    Siyu Teng, Xuemin Hu, Peng Deng, Bai Li, Yuchen Li, Yunfeng Ai, Dongsheng Yang, Lingxi Li, Zhe Xuanyuan, Fenghua Zhu, et al. Motion planning for autonomous driv- ing: The state of the art and future perspectives.IEEE Trans- actions on Intelligent Vehicles, 8(6):3692–3711, 2023. 2

    2023

  30. [38]

    A reinforcement learning based approach for automated lane change maneuvers

    Pin Wang, Ching-Yao Chan, and Arnaud de La Fortelle. A reinforcement learning based approach for automated lane change maneuvers. In 2018 IEEE Intelligent Vehicles Sym- posium, pages 1379–1384. IEEE, 2018. 2

    2018

  31. [39]

    Unleashing the po- tential of mamba: Boosting a lidar 3d sparse detector by using cross-model knowledge distillation

    Rui Yu, Runkai Zhao, Jiagen Li, Qingsong Zhao, Songhao Zhu, HuaiCheng Yan, and Meng Wang. Unleashing the po- tential of mamba: Boosting a lidar 3d sparse detector by using cross-model knowledge distillation. arXiv preprint arXiv:2409.11018, 2024. 2

    2024

  32. [40]

    Future does matter: Boosting 3d ob- ject detection with temporal motion estimation in point cloud sequences

    Rui Yu, Runkai Zhao, Cong Nie, Heng Wang, HuaiCheng Yan, and Meng Wang. Future does matter: Boosting 3d ob- ject detection with temporal motion estimation in point cloud sequences. British Machine Vision Conference, 2024. 1

    2024

  33. [41]

    Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes

    Jiang-Tian Zhai, Ze Feng, Jinhao Du, Yongqiang Mao, Jiang-Jiang Liu, Zichang Tan, Yifu Zhang, Xiaoqing Ye, and Jingdong Wang. Rethinking the open-loop evaluation of end-to-end autonomous driving in nuscenes. arXiv preprint arXiv:2305.10430, 2023. 2

    2023 arXiv

  34. [42]

    End-to-end urban driving by imitat- ing a reinforcement learning coach

    Zhejun Zhang, Alexander Liniger, Dengxin Dai, Fisher Yu, and Luc Van Gool. End-to-end urban driving by imitat- ing a reinforcement learning coach. In Proceedings of the IEEE/CVF International Conference on Computer Vision , pages 15222–15232, 2021. 2

    2021

  35. [43]

    Lanecmkt: Boosting monocular 3d lane detection with cross-modal knowledge transfer

    Runkai Zhao, Heng Wang, and Weidong Cai. Lanecmkt: Boosting monocular 3d lane detection with cross-modal knowledge transfer. In ACM Multimedia 2024. 2

    2024

  36. [44]

    Genad: Generative end-to-end au- tonomous driving

    Wenzhao Zheng, Ruiqi Song, Xianda Guo, Chenming Zhang, and Long Chen. Genad: Generative end-to-end au- tonomous driving. In European Conference on Computer Vision, pages 87–104. Springer, 2024. 2

    2024

  37. [45]

    Diffusion-based planning for autonomous driving with flexible guidance

    Yinan Zheng, Ruiming Liang, Kexin Zheng, Jinliang Zheng, Liyuan Mao, Jianxiong Li, Weihao Gu, Rui Ai, Shengbo Eben Li, Xianyuan Zhan, et al. Diffusion-based planning for autonomous driving with flexible guidance. arXiv preprint arXiv:2501.15564, 2025. 1

    2025 arXiv

  38. [46]

    Knowledge distillation by on-the-fly native ensemble

    Xiatian Zhu, Shaogang Gong, et al. Knowledge distillation by on-the-fly native ensemble. Advances in Neural Informa- tion Processing Systems, 31, 2018. 2

    2018

  39. [47]

    Deformable detr: Deformable trans- formers for end-to-end object detection

    Xizhou Zhu, Weijie Su, Lewei Lu, Bin Li, Xiaogang Wang, and Jifeng Dai. Deformable detr: Deformable trans- formers for end-to-end object detection. arXiv preprint arXiv:2010.04159, 2020. 2 DistillDrive: End-to-End Multi-Mode Autonomous Driving Distillation by Isomorphic Hetero...

    2010 arXiv

  40. [48]

    Supplementary Methodology In this section, we present additional content that is not cov- ered in detail in the paper to further support the theoretical framework of the argumentative article. 6.1. IRL-based Teacher Model Position Embedding. For the temporal feature IA, we ap-...

  41. [49]

    otherwise. (21) Additionally, for state predictions ˆs and planning tra- jectories ˆptraj, we evaluate their consistency error to en- courage the model to establish associations between learned states and planning as following: econsist = ˆst1 vx × ∆t − ˆpt1 traj . (22) Meanwh...

  42. [50]

    Supplementary Experiments Here, we provide a complementary account of ablation ex- periments and visual analyses to illustrate the improvements in the model’s performance. 7.1. Ablation Studies Ablation Study in Closed-Loop Evaluation While the paper focused on ablation studie...

Pith tools

Reviewed August 5, 2026 · model on record in the stance chip above.