Pith. sign in

REVIEW 5 major objections 6 minor 7 cited by

Generalized Trajectory Scoring for End-to-end Multimodal Planning

T0 review · 5 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

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

Pith's one-line read The paper's central claim is that a camera-only trajectory scorer can nearly match a privileged ground-truth planner by learning to score both static and dynamically generated trajectory candidates.

desk verdict GTRS is a solid benchmark win: the vocabulary-dropout scoring idea is genuinely new and the ablations hold together, but the paper oversells 'approaching PDM-Closed' and should report per-stage EPDMS before that headline is trusted. read the letter →

arxiv 2506.06664 v1 pith:FOEICHWP submitted 2025-06-07 cs.RO cs.CV

Zhenxin Li , Wenhao Yao , Zi Wang , Xinglong Sun , Joshua Chen , Nadine Chang , Maying Shen , Zuxuan Wu
show 2 more authors
Shiyi Lan Jose M. Alvarez
This is my paper
classification cs.ROcs.CV
keywords autonomousdrivingend-to-endmultimodalplanningtrajectoryscoringdiffusionpolicyvocabularygeneralizationdropoutsensoraugmentationNavsimbenchmark
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 a trajectory scorer for end-to-end driving can learn to generalize across both static and dynamically generated trajectory candidates if it is trained on a very dense fixed vocabulary under deliberate dropout, then combined at inference with a diffusion-based proposal generator. The authors build GTRS around three components: a diffusion policy that proposes fine-grained trajectories, a scorer trained on 16,384 trajectories but asked to score smaller subsets, and sensor augmentation with refinement training to survive viewpoint shifts and subtle trajectory differences. Their sensor-only system, an ensemble of six models, reaches an extended PDM score (EPDMS) of 49.4 on the Navhard benchmark, within 1.9 points of the privileged planner PDM-Closed, which uses ground-truth perception and scores 51.3. The payoff, if this holds, is that camera-only planners can approach the quality of planners with perfect perception by improving how candidates are ranked rather than by adding sensors.

What carries the argument

The load-bearing mechanism is the vocabulary mismatch: training on the super-dense trajectory set VXL (16,384 candidates) with random dropout, then evaluating on smaller subsets (VL with 8,192 candidates plus 100 diffusion proposals). This deliberately creates a train/inference vocabulary gap that forces the scorer to learn generalizable trajectory features rather than memorized candidates. The diffusion-based generator supplies fine-grained, context-aware proposals at inference only, while the sensor augmentation (2D horizontal rotations of input views) and the refinement decoder with EMA soft targets are the supporting mechanisms that make the scorer reliable under distribution shift and subtle candidate differences.

What would settle it

Evaluate the same ensemble on a non-synthetic, real-world driving benchmark: if its score relative to the ground-truth-perception planner drops substantially below the reported 49.4 versus 51.3, the near-parity claim is specific to Navhard rather than general.

Watch

Extended reading notes

Core claim

GTRS's central claim is that trajectory scoring generalization is a vocabulary problem. Instead of choosing between a large static vocabulary and a small set of dynamic proposals, GTRS trains its scorer on a super-dense vocabulary VXL of 16,384 trajectories, randomly drops half of the trajectories per batch during training, and at inference scores the union of a smaller 8,192-entry vocabulary VL and 100 diffusion-generated proposals. Because the model never sees the inference-time subset during training, it must learn trajectory-quality features rather than memorize candidates. On top of this, rotation-based sensor augmentation and an EMA-guided refinement decoder allow the scorer to keep working under viewpoint shifts and to separate near-identical trajectories. The paper reports that the six-model ensemble GTRS-E obtains EPDMS 49.4 on Navhard, approaching PDM-Closed's 51.3 despite PDM-Closed's access to ground-truth perception.

Load-bearing premise

The central claim rests on Navhard, including its computer-generated synthetic driving scenes, being a valid stand-in for real-world driving quality; the paper itself notes those synthetic images contain artifacts such as distortion and blurring.

Editorial extensions

If this is right

  • A camera-only trajectory scorer can come within about two EPDMS points of a privileged planner that uses ground-truth perception, on the Navhard benchmark.
  • A scorer trained only on a static vocabulary can rank unseen diffusion-generated trajectories without retraining, so dynamic proposals can be appended at inference time.
  • Deliberately training on a super-dense vocabulary with random dropout improves out-of-domain scoring compared with matching the training and inference vocabularies.
  • Rotation-based sensor augmentation and refinement training allow the scorer to maintain performance under viewpoint changes and to separate near-identical trajectory candidates.

Reading between the lines

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

  • If the vocabulary-mismatch mechanism is the active ingredient, the same training-on-dense/inference-on-sparse recipe could extend to other candidate-ranking tasks, such as motion forecasting or selecting from large language-model response sets, where the scorer must rank candidates not seen during training.
  • The paper's own observation that the 3DGS synthetic continuations contain artifacts leaves open whether the 1.9-point gap to PDM-Closed shrinks or grows on fully realistic sensor data; a natural next test is to evaluate GTRS on real-world closed-loop tracks.
  • One can isolate the contribution of vocabulary dropout by sweeping the inference vocabulary size downward from 8,192; the claim predicts a slow degradation rather than a cliff, since the whole point is generalization to smaller unseen subsets.
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

5 major / 6 minor

Summary. The paper presents GTRS, a trajectory scoring framework for camera-based end-to-end planning evaluated on the Navsim v2 Navhard benchmark. It combines three ingredients: (i) a diffusion-policy generator that produces 100 dynamic trajectory proposals; (ii) a transformer-based scorer trained on a super-dense static vocabulary of 16,384 trajectories with vocabulary dropout, then evaluated on a smaller vocabulary augmented with dynamic proposals; and (iii) sensor-rotation augmentation plus EMA-refinement training. The best single model achieves EPDMS 45.3, while the six-model ensemble GTRS-E achieves 49.4 against 51.3 for the privileged PDM-Closed planner. The paper claims that a sensor-based approach can 'approach' a ground-truth-perception planner despite degraded synthetic sensor inputs.

Significance. GTRS is the winning entry of the Navsim v2 Challenge, and if the reported numbers hold, it represents a substantial step: a camera-only scorer nearly matching a privileged planner, with a simple vocabulary-dropout recipe that appears to transfer from static to dynamic proposals. The paper is honest about synthetic-data artifacts in Section 4.1 and plans code release, which supports reproducibility. However, the headline comparison is currently under-specified: per-stage EPDMS values are not reported, the training-target generation for the super-dense vocabulary is not described, and the ablations have no variance estimates. These issues do not invalidate the technical ingredients, but they must be resolved before the central generalization claim can be fully assessed.

major comments (5)
  1. [Sec. 4.1, Table 2] Table 2 reports per-metric stage-wise numbers but not the official EPDMS for Stage 1 and Stage 2 separately. Since EPDMS is a nonlinear aggregate and PDM-Closed is insensitive to sensor corruption, the 49.4 versus 51.3 gap could be dominated by Stage 2, whose synthetic inputs are, by the authors' own admission, artifact-laden. Please report official Stage 1 and Stage 2 EPDMS for every method and state whether the 'approaches PDM-Closed' conclusion holds on real-world Stage 1 data alone.
  2. [Sec. 2.2, Sec. 4.1] The process that assigns scores to the 16,384-token super-dense vocabulary VXL is not specified. If these scores are produced by the PDM/EPDMS rule-based metric, then the scorer is trained against the same function used for evaluation; this is not a logical circularity, but it is a concern about how much of the reported improvement is metric alignment rather than planning quality. Please state the supervision protocol for VXL (e.g., PDM Score, expert distillation, or one-hot labels) and, if rule-based scores are used, discuss the implications for interpreting the gains.
  3. [Sec. 4.3, Table 1] The roadmap ablations are presented without error bars or significance tests, and several load-bearing differences are small: adding Vdp to VXL improves EPDMS by 1.1 (39.7 to 40.8), switching from VXL to VL improves by 1.2 (40.8 to 42.0), and vocabulary dropout adds 1.4 (42.0 to 43.4). With a single seed and an opaque aggregate metric, these differences are not distinguishable from noise. Please provide at least three seeds with standard deviations, or otherwise justify that the EPDMS differences exceed the benchmark's run-to-run variability.
  4. [Sec. 1, Table 2] The abstract and Section 1 claim that 'our sensor-based GTRS ... approaches' PDM-Closed, but this is true only for the six-model ensemble GTRS-E (49.4 versus 51.3); the best single model is 45.3. Please report the ensemble details (which components, any weight averaging, and inference cost) and clarify that the claim refers to an ensemble, or temper the claim so that it matches the single-model result.
  5. [Sec. 4.3] The sentence 'Vdp ∪ VL outperforms Vdp ∪ VXL, likely because the reduced vocabulary complexity leads to better generalization in out-of-domain synthetic data' is a post hoc explanation for an experiment that was not designed to test it. Because this observation motivates the central design choice of training on VXL and inferring on VL, please add a controlled comparison (e.g., holding vocabulary construction fixed while varying size, or analyzing per-stage EPDMS) that directly supports the explanation.
minor comments (6)
  1. [Sec. 4.3] The first paragraph refers to 'Tab. 4.3' but the table is numbered Table 1; please correct the cross-reference.
  2. [Sec. 2.3, Sec. 4.2] Sensor augmentation and refinement hyperparameters are missing: the rotation angle distribution, the clipping delta in Eq. (1), the top-k used for refinement, the EMA decay, and the number of refinement iterations are not given. Please add these for reproducibility.
  3. [Eq. (1)] The symbols s^m_{i,teacher} and delta^m are not defined in the text, and it is unclear whether y-hat^m_i is the raw metric score or a normalized target. Please clarify.
  4. [Fig. 1] The 'Q, K, V' annotations in the figure are not explained in the caption; the reader cannot tell which decoder is being illustrated in each pillar.
  5. [Sec. 4.1] The Navhard split and the two-stage scoring pipeline are described only by a footnote and reference [2]; please give the exact EPDMS aggregation formula or cite a versioned challenge document.
  6. [Abstract, Sec. 4.1] The abstract says 'superior performance even with sub-optimal sensor inputs,' while Section 4.1 concedes that the synthetic inputs 'may impair the performance of sensor-based planners.' Please make the claim consistent, for example, 'comparative performance despite degraded inputs.'

Circularity Check

0 steps flagged · score 0.0 of 10

No significant circularity: GTRS's headline result is an externally evaluated benchmark comparison, not a quantity re-derived from its own fitted targets.

full rationale

The paper's central claim is an empirical benchmark statement: a sensor-based scorer reaches an EPDMS of 49.4 on the external Navsim v2 Navhard test split, approaching PDM-Closed's 51.3, with training performed on Navtrain and no use of Navhard for training. The evaluation metric EPDMS is a rule-based, externally documented scoring pipeline, not a quantity defined by the model or by the paper's own equations. The three technical contributions (diffusion proposals, vocabulary dropout generalization, and sensor augmentation with EMA-refined targets) are described as training-time mechanisms; Eq. (1) is a self-distillation refinement target, not the evaluation metric. The paper builds on Hydra-MDP [18], which is prior work with overlapping authorship, and cites Hydra-MDP++ [16] for EPDMS; these are self-citations, but they are not load-bearing in the sense of making the result true by construction. The reported improvements over the LTF baseline and over the paper's own ablations are empirical and are not fitted to the evaluation metric on the test split. The paper's own caveat that the 3DGS synthetic continuation data contain artifacts that may impair sensor-based planners is a generalization limitation, not a circularity: it does not make the EPDMS comparison equivalent to the input. No circular step could be exhibited with a specific equation or construction reducing the prediction to the input.

Assumptions & free parameters 8 free parameters · 5 assumptions · 0 invented entities

The central claim rests on benchmark validity, metric consistency, and a set of hand-chosen hyperparameters (vocabulary sizes, dropout ratio, number of proposals, augmentation angle, refinement clipping). No new physical or computational entities are introduced; the framework is a combination of existing components.

free parameters (8)
  • Super-dense training vocabulary size = 16,384
    Chosen by hand; central to the vocabulary generalization claim (Sec. 2.2).
  • Inference vocabulary size = 8,192
    Chosen by hand; the paper reports better EPDMS with VL than VXL at inference (Table 1).
  • Vocabulary dropout ratio = 50% (half removed per batch)
    Chosen by hand; dropout ratio is not ablated.
  • Number of diffusion proposals = 100
    Used at inference for Vdp; no sensitivity analysis (Sec. 4.2).
  • Sensor rotation perturbation magnitude = not specified numerically
    The 2D horizontal view rotation angle is not quantified, so the augmentation strength is a hidden tuning choice (Sec. 2.3).
  • Refinement clipping delta = not specified numerically
    The clip bound delta_m in Eq. 1 is central to the refinement target but its value is not reported.
  • Top-k for refinement = not specified
    The refinement decoder operates on top-k candidates; k is not stated (Sec. 2.3).
  • Number of denoising steps = 100
    DDPM scheduler with 100 steps (Sec. 4.2); not ablated.
assumptions (5)
  • domain assumption The Navhard benchmark, including its 3DGS synthetic continuations, is a valid proxy for out-of-domain planning quality.
    The generalization claims are evaluated solely on this benchmark; the paper admits synthetic artifacts (Sec. 4.1).
  • domain assumption EPDMS, an extension of the PDM Score, is a faithful measure of planning performance.
    Used as the target metric; if it misrepresents real-world driving, the headline numbers are not meaningful (Sec. 4.1).
  • domain assumption 2D horizontal image rotation with corresponding ground truth transformation preserves label consistency for a multi-camera BEV pipeline.
    The augmentation assumes a planar rotation approximates viewpoint change without breaking camera geometry (Sec. 2.3).
  • standard math DDPM denoising correctly models multi-modal trajectory distributions under first-order waypoint differentiation.
    The generator follows Diffusion Policy [4] and DDPM [10]; the mathematical correctness is taken from those references.
  • ad hoc to paper Training supervision for the super-dense vocabulary is consistent with the evaluation metric.
    The paper builds on Hydra-MDP, whose multi-target distillation uses PDM-style scores, and evaluates on EPDMS; this coupling is assumed beneficial but promotes metric overfitting.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Generalized Trajectory Scoring for End-to-end Multimodal Planning." pith.science (2026). https://pith.science/paper/FOEICHWP

@misc{pith2026250606664,
  author       = {Pith},
  title        = {Pith review of: Generalized Trajectory Scoring for End-to-end Multimodal Planning},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/FOEICHWP}},
  note         = {Machine review of arXiv:2506.06664}
}
read the original abstract

End-to-end multi-modal planning is a promising paradigm in autonomous driving, enabling decision-making with diverse trajectory candidates. A key component is a robust trajectory scorer capable of selecting the optimal trajectory from these candidates. While recent trajectory scorers focus on scoring either large sets of static trajectories or small sets of dynamically generated ones, both approaches face significant limitations in generalization. Static vocabularies provide effective coarse discretization but struggle to make fine-grained adaptation, while dynamic proposals offer detailed precision but fail to capture broader trajectory distributions. To overcome these challenges, we propose GTRS (Generalized Trajectory Scoring), a unified framework for end-to-end multi-modal planning that combines coarse and fine-grained trajectory evaluation. GTRS consists of three complementary innovations: (1) a diffusion-based trajectory generator that produces diverse fine-grained proposals; (2) a vocabulary generalization technique that trains a scorer on super-dense trajectory sets with dropout regularization, enabling its robust inference on smaller subsets; and (3) a sensor augmentation strategy that enhances out-of-domain generalization while incorporating refinement training for critical trajectory discrimination. As the winning solution of the Navsim v2 Challenge, GTRS demonstrates superior performance even with sub-optimal sensor inputs, approaching privileged methods that rely on ground-truth perception. Code will be available at https://github.com/NVlabs/GTRS.

Figures

Figures reproduced from arXiv: 2506.06664 by the authors.

Figure 1
Figure 1. The Three Pillars of GTRS. bution shifts in the form of viewpoint changes, we intro￾duce a data augmentation strategy by applying rotation perturbations to sensor inputs, dramatically improving ro￾bustness to out-of-domain environments. Further, a refine￾ment training mechanism enables the model to distinguish between subtly different trajectory options. (Sec. 2.3) GTRS demonstrates strong trajectory scoring abiliti… view at source ↗
Figure 2
Figure 2. The Inference Pipeline of GTRS. the trajectories in each batch. This serves multiple purposes: (1) it aligns the number of trajectory tokens during training and inference, (2) it creates intentional distribution shifts that improve robustness, and (3) it acts as an effective regularizer against overfitting to specific trajectory patterns. This vocabulary generalization technique enables our model to effectively scor… view at source ↗

Discussion (0). Sign in to comment.

Forward citations

Cited by 7 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score. Full citation record

  1. MOJITO: Modal Joint Learning for Unified End-to-End Autonomous Driving

    cs.CV 2026-07 conditional novelty 6.0 of 10

    Block-wise Modal Joint Attention over image, LiDAR, and diffusion action tokens yields 88.9 PDMS / 88.4 EPDMS on NAVSIM without anchors or auxiliary supervision.

  2. DRIFT: Drift and Aggregation for Motion Planning

    cs.RO 2026-07 conditional novelty 6.0 of 10

    DRIFT achieves 89.6 PDMS and 90.4 EPDMS on NAVSIM navtest by generating proposal features via one-step latent drift and aggregating them label-free.

  3. Self-Imitated Diffusion Policy for Efficient and Robust Visual Navigation

    cs.RO 2026-01 conditional novelty 6.0 of 10

    SIDP trains a diffusion policy for visual navigation by reward-weighting its own sampled trajectories, improving success rate and cutting inference latency.

  4. IRL-VLA: Training an Vision-Language-Action Policy via Reward World Model

    cs.AI 2025-08 conditional novelty 6.0 of 10

    IRL-VLA fine-tunes a vision-language-action driving policy with PPO against a learned reward world model trained on NAVSIM's EPDMS metrics, reaching 74.9 EPDMS on navhard-real.

  5. HAD: Combining Hierarchical Diffusion with Metric-Decoupled RL for End-to-End Driving

    cs.RO 2026-04 conditional novelty 5.5 of 10

    Hierarchical diffusion plus polar structure-preserving expansion and metric-decoupled RL yields SOTA open- and closed-loop planning scores on NAVSIM and HUGSIM.

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

    cs.CV 2025-07 unverdicted novelty 5.0 of 10

    DIVER uses RL-guided diffusion to produce diverse feasible trajectories from one ground-truth path, addressing mode collapse in imitation learning for autonomous driving.

  7. A Survey on Vision-Language-Action Models for Autonomous Driving

    cs.CV 2025-06 conditional novelty 4.0 of 10

    A survey organizes vision-language-action models for autonomous driving into four stages, compares over 20 systems, and catalogs datasets, benchmarks, and open challenges.

Reference graph

Works this paper leans on

33 extracted references · 17 canonical work pages · cited by 7 Pith papers

  1. [1]

    Quad: Query-based inter- pretable neural motion planning for autonomous driving

    Sourav Biswas, Sergio Casas, Quinlan Sykora, Ben Agro, Abbas Sadat, and Raquel Urtasun. Quad: Query-based inter- pretable neural motion planning for autonomous driving. In IEEE International Conference on Robotics and Automation, pages 14236–14243. IEEE, 2024. 1

    work page 2024

  2. [2]

    Pseudo-simulation for autonomous driving.arXiv preprint arXiv:2506.04218, 2025

    Wei Cao, Marcel Hallgarten, Tianyu Li, Daniel Dauner, Xunjiang Gu, Caojun Wang, Yakov Miron, Marco Aiello, Hongyang Li, Igor Gilitschenski, et al. Pseudo-simulation for autonomous driving.arXiv preprint arXiv:2506.04218, 2025. 2

    arXiv 2025

  3. [3]

    Vadv2: End-to-end vectorized autonomous driving via probabilistic planning.arXiv preprint arXiv:2402.13243,

    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

  4. [4]

    Diffusion policy: Visuomotor policy learning via action diffu- sion.The International Journal of Robotics Research, page 02783649241273668, 2023

    Cheng Chi, Zhenjia Xu, Siyuan Feng, Eric Cousineau, Yilun Du, Benjamin Burchfiel, Russ Tedrake, and Shuran Song. Diffusion policy: Visuomotor policy learning via action diffu- sion.The International Journal of Robotics Research, page 02783649241273668, 2023. 1, 2

    work page 2023

  5. [5]

    Transfuser: Imita- tion with transformer-based sensor fusion for autonomous driving.IEEE TPAMI, 2022

    Kashyap Chitta, Aditya Prakash, Bernhard Jaeger, Zehao Yu, Katrin Renz, and Andreas Geiger. Transfuser: Imita- tion with transformer-based sensor fusion for autonomous driving.IEEE TPAMI, 2022. 1, 2, 4

    work page 2022

  6. [6]

    Parting with misconceptions about learning- based vehicle motion planning

    Daniel Dauner, Marcel Hallgarten, Andreas Geiger, and Kashyap Chitta. Parting with misconceptions about learning- based vehicle motion planning. InConference on Robot Learning, pages 1268–1281. PMLR, 2023. 1, 2, 4

    work page 2023

  7. [7]

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

    Daniel Dauner, Marcel Hallgarten, Tianyu Li, Xinshuo Weng, Zhiyu Huang, Zetong Yang, Hongyang Li, Igor Gilitschenski, Boris Ivanovic, Marco Pavone, Andreas Geiger, and Kashyap Chitta. Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking. InNeurIPS, 2024. 3, 4

    work page 2024

  8. [8]

    An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Syl- vain Gelly, et al. An image is worth 16x16 words: Trans- formers for image recognition at scale.arXiv preprint arXiv:2010.11929, 2020. 3, 4

    arXiv 2010

Show all 33 references
  1. [9]

    Eva-02: A visual representation for neon genesis.arXiv preprint arXiv:2303.11331, 2023

    Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xin- long Wang, and Yue Cao. Eva-02: A visual representation for neon genesis.arXiv preprint arXiv:2303.11331, 2023. 3, 4

    2023 arXiv

  2. [10]

    Denoising diffu- sion probabilistic models.NeurIPS, 33:6840–6851, 2020

    Jonathan Ho, Ajay Jain, and Pieter Abbeel. Denoising diffu- sion probabilistic models.NeurIPS, 33:6840–6851, 2020. 2, 4

    2020

  3. [11]

    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. InCVPR, pages 17853–17862, 2023. 1

    2023

  4. [12]

    Vad: Vectorized scene representation 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 representation for efficient autonomous driving. InICCV, pages 8340–8350,

  5. [13]

    Motiondiffuser: Con- trollable multi-agent motion prediction using diffusion

    Chiyu Jiang, Andre Cornman, Cheolho Park, Benjamin Sapp, Yin Zhou, Dragomir Anguelov, et al. Motiondiffuser: Con- trollable multi-agent motion prediction using diffusion. In CVPR, pages 9644–9653, 2023. 2

    2023

  6. [14]

    An energy and gpu-computation efficient backbone network for real-time object detection

    Youngwan Lee, Joong-won Hwang, Sangrok Lee, Yuseok Bae, and Jongyoul Park. An energy and gpu-computation efficient backbone network for real-time object detection. In CVPRW, pages 0–0, 2019. 3

    2019

  7. [15]

    Finetuning generative trajectory model with reinforcement learning from human feedback.arXiv preprint arXiv:2503.10434, 2025

    Derun Li, Jianwei Ren, Yue Wang, Xin Wen, Pengxiang Li, Leimeng Xu, Kun Zhan, Zhongpu Xia, Peng Jia, Xianpeng Lang, et al. Finetuning generative trajectory model with reinforcement learning from human feedback.arXiv preprint arXiv:2503.10434, 2025. 1, 2

    2025

  8. [16]

    Hydra-mdp++: Advancing end-to-end driving via expert- guided hydra-distillation.arXiv preprint arXiv:2503.12820,

    Kailin Li, Zhenxin Li, Shiyi Lan, Yuan Xie, Zhizhong Zhang, Jiayi Liu, Zuxuan Wu, Zhiding Yu, and Jose M Alvarez. Hydra-mdp++: Advancing end-to-end driving via expert- guided hydra-distillation.arXiv preprint arXiv:2503.12820,

    arXiv

  9. [17]

    Is ego status all you need for open-loop end-to-end autonomous driving?arXiv preprint arXiv:2312.03031, 2023

    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?arXiv preprint arXiv:2312.03031, 2023. 3

    2023 arXiv

  10. [18]

    Hydra-mdp: End-to-end multimodal planning with multi- target hydra-distillation.arXiv preprint arXiv:2406.06978,

    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

  11. [19]

    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. InCVPR, 2025. 1, 2

    2025

  12. [20]

    Is pseudo-lidar needed for monocular 3d object detection? InICCV, pages 3142–3152, 2021

    Dennis Park, Rares Ambrus, Vitor Guizilini, Jie Li, and Adrien Gaidon. Is pseudo-lidar needed for monocular 3d object detection? InICCV, pages 3142–3152, 2021. 3

    2021

  13. [21]

    Carllava: Vision language mod- els for camera-only closed-loop driving.arXiv preprint arXiv:2406.10165, 2024

    Katrin Renz, Long Chen, Ana-Maria Marcu, Jan H¨unermann, Benoit Hanotte, Alice Karnsund, Jamie Shotton, Elahe Arani, and Oleg Sinavski. Carllava: Vision language mod- els for camera-only closed-loop driving.arXiv preprint arXiv:2406.10165, 2024. 1

    2024 arXiv

  14. [22]

    Simlingo: Vision-only closed-loop autonomous driving with language-action alignment.arXiv preprint arXiv:2503.09594, 2025

    Katrin Renz, Long Chen, Elahe Arani, and Oleg Sinavski. Simlingo: Vision-only closed-loop autonomous driving with language-action alignment.arXiv preprint arXiv:2503.09594, 2025

    2025 arXiv

  15. [23]

    Drivelm: Driving with graph visual question answering

    Chonghao Sima, Katrin Renz, Kashyap Chitta, Li Chen, Hanxue Zhang, Chengen Xie, Jens Beißwenger, Ping Luo, Andreas Geiger, and Hongyang Li. Drivelm: Driving with graph visual question answering. InECCV, pages 256–274. Springer, 2024. 1

    2024

  16. [24]

    Centaur: Robust end-to-end autonomous driving with test- time training.arXiv preprint arXiv:2503.11650, 2025

    Chonghao Sima, Kashyap Chitta, Zhiding Yu, Shiyi Lan, Ping Luo, Andreas Geiger, Hongyang Li, and Jose M Alvarez. Centaur: Robust end-to-end autonomous driving with test- time training.arXiv preprint arXiv:2503.11650, 2025. 1

    2025 arXiv

  17. [25]

    Hip-ad: Hierarchical and multi-granularity planning with de- formable attention for autonomous driving in a single decoder

    Yingqi Tang, Zhuoran Xu, Zhaotie Meng, and Erkang Cheng. Hip-ad: Hierarchical and multi-granularity planning with de- formable attention for autonomous driving in a single decoder. arXiv preprint arXiv:2503.08612, 2025. 1

    2025 arXiv

  18. [26]

    Attention is all you need.NeurIPS, 30, 2017

    Ashish Vaswani, Noam Shazeer, Niki Parmar, Jakob Uszko- reit, Llion Jones, Aidan N Gomez, Łukasz Kaiser, and Illia Polosukhin. Attention is all you need.NeurIPS, 30, 2017. 2 5

    2017

  19. [27]

    Exploring object-centric temporal modeling for efficient multi-view 3d object detection

    Shihao Wang, Yingfei Liu, Tiancai Wang, Ying Li, and Xi- angyu Zhang. Exploring object-centric temporal modeling for efficient multi-view 3d object detection. InICCV, pages 3621–3631, 2023. 3

    2023

  20. [28]

    Omnidrive: A holistic llm-agent framework for autonomous driving with 3d perception, reasoning and planning.arXiv preprint arXiv:2405.01533, 2024

    Shihao Wang, Zhiding Yu, Xiaohui Jiang, Shiyi Lan, Min Shi, Nadine Chang, Jan Kautz, Ying Li, and Jose M Alvarez. Omnidrive: A holistic llm-agent framework for autonomous driving with 3d perception, reasoning and planning.arXiv preprint arXiv:2405.01533, 2024. 1

    2024 arXiv

  21. [29]

    Diffusion- based environment-aware trajectory prediction.arXiv preprint arXiv:2403.11643, 2024

    Theodor Westny, Bj¨orn Olofsson, and Erik Frisk. Diffusion- based environment-aware trajectory prediction.arXiv preprint arXiv:2403.11643, 2024. 2

    2024 arXiv

  22. [30]

    Goalflow: Goal- driven flow matching for multimodal trajectories generation in end-to-end autonomous driving

    Zebin Xing, Xingyu Zhang, Yang Hu, Bo Jiang, Tong He, Qian Zhang, Xiaoxiao Long, and Wei Yin. Goalflow: Goal- driven flow matching for multimodal trajectories generation in end-to-end autonomous driving. InCVPR, 2025. 1

    2025

  23. [31]

    Diffusion-es: Gradient-free planning with diffusion for au- tonomous and instruction-guided driving

    Brian Yang, Huangyuan Su, Nikolaos Gkanatsios, Tsung-Wei Ke, Ayush Jain, Jeff Schneider, and Katerina Fragkiadaki. Diffusion-es: Gradient-free planning with diffusion for au- tonomous and instruction-guided driving. InCVPR, pages 15342–15353, 2024. 2

    2024

  24. [32]

    Depth anything: Unleash- ing the power of large-scale unlabeled data.arXiv preprint arXiv:2401.10891, 2024

    Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Ji- ashi Feng, and Hengshuang Zhao. Depth anything: Unleash- ing the power of large-scale unlabeled data.arXiv preprint arXiv:2401.10891, 2024. 3

    2024 arXiv

  25. [33]

    Diffusion-based planning for au- tonomous driving with flexible guidance.arXiv preprint arXiv:2501.15564, 2025

    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 au- tonomous driving with flexible guidance.arXiv preprint arXiv:2501.15564, 2025. 2 6

    2025 arXiv

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.