pith. machine review for the scientific record. sign in

arxiv: 2605.14201 · v1 · submitted 2026-05-13 · 💻 cs.RO · cs.CV

Recognition: 2 theorem links

· Lean Theorem

MAPLE: Latent Multi-Agent Play for End-to-End Autonomous Driving

Rajeev Yasarla , Deepti Hegde , Hsin-Pai Cheng , Shizhong Han , Yunxiao Shi , Meysam Sadeghigooghari , Hanno Ackermann , Litian Liu
show 4 more authors
Pranav Desai Fatih Porikli Mohammad Ghavamzadeh Hong Cai
Authors on Pith no claims yet

Pith reviewed 2026-05-15 04:38 UTC · model grok-4.3

classification 💻 cs.RO cs.CV
keywords autonomous drivingvision-language-action modelsmulti-agent systemsclosed-loop traininglatent spacereinforcement learningend-to-end planning
0
0 comments X

The pith

MAPLE trains end-to-end driving models in closed loop using latent multi-agent rollouts without external simulators.

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

The paper introduces MAPLE to address the brittleness of vision-language-action models in closed-loop autonomous driving evaluations. It does this by enabling reactive multi-agent interactions directly in the model's latent space over multi-step horizons. Training proceeds in two stages: supervised fine-tuning on ground-truth latent trajectories and then reinforcement learning with rewards for safety, progress, interaction realism, and behavioral diversity. A reader would care because this avoids the scalability issues and limited fidelity of traditional simulators while improving robustness in dynamic traffic scenarios.

Core claim

MAPLE performs independent control of the ego vehicle and nearby traffic agents in the latent space of a vision-language-action model, allowing them to react to each other over multiple time steps. This latent rollout supports closed-loop supervision through an initial supervised fine-tuning stage on ground-truth data followed by reinforcement learning that incorporates global and agent-specific rewards. The resulting model achieves state-of-the-art performance on the Bench2Drive benchmark by learning more realistic and diverse driving behaviors.

What carries the argument

latent multi-agent rollout which enables independent yet reactive control of multiple agents in the VLA model's latent space to simulate closed-loop dynamics for training

If this is right

  • The model can handle reactive environments better than standard imitation learning approaches.
  • Training scales without the need for external simulators or high visual fidelity requirements.
  • Diversity rewards allow the generation of planning behaviors absent from logged data.
  • Global and agent-specific rewards promote safety, progress, and realistic interactions in multi-agent scenes.

Where Pith is reading between the lines

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

  • This latent-space approach might reduce the domain gap when transferring to real-world driving compared to simulator-based methods.
  • Extending the framework to longer horizons or more agents could further enhance its applicability to complex urban scenarios.
  • Combining MAPLE with online adaptation during deployment might address remaining distribution shifts.

Load-bearing premise

Latent space rollouts of the VLA model can accurately capture the reactive dynamics between the ego vehicle and other agents without external simulators or extra visual fidelity losses.

What would settle it

If evaluations on Bench2Drive or similar closed-loop tests show no improvement over baseline imitation learning methods, or if the generated rollouts fail to produce appropriate reactions to changes in other agents' behaviors.

Figures

Figures reproduced from arXiv: 2605.14201 by Deepti Hegde, Fatih Porikli, Hanno Ackermann, Hong Cai, Hsin-Pai Cheng, Litian Liu, Meysam Sadeghigooghari, Mohammad Ghavamzadeh, Pranav Desai, Rajeev Yasarla, Shizhong Han, Yunxiao Shi.

Figure 1
Figure 1. Figure 1: MAPLE pretraining and future state prediction. Left: Pretraining the VLA backbone with auxiliary supervision (e.g., map learning, detection, and motion prediction). Right: State-transition pretraining that predicts next-step ego/agent states over a horizon T to stabilize the token space. Multi-agent simulation and self-play. Trajectory forecasting methods [12, 35, 49] model joint agent futures from fixed o… view at source ↗
Figure 2
Figure 2. Figure 2: MAPLE supervised fine-tuning (SFT) stage. Left: Single-step supervision and inference. The VLA backbone encodes multi-view images (and map features) into ego and agent tokens, which are decoded by an ego planner, reactive-agent planners, and a motion head. Right: The same model unrolled for T steps during imitation-learning-based scenario rollouts. Predicted tokens/trajectories are fed back autoregressivel… view at source ↗
Figure 3
Figure 3. Figure 3: MAPLE RL fine-tuning stage. Starting from the SFT model, we optimize multi-step rollouts over T steps using RL with safety-aware and interaction-aware rewards (e.g., collision avoidance and TTC). progress and safe driving for each controlled agent, and (iii) a diversity reward that promotes distinct behaviors across different planners/policies. At a time t, we define the total rollout reward as Rt = Gt + D… view at source ↗
Figure 4
Figure 4. Figure 4: Qualitative examples of closed-loop driving on Bench2Drive using MAPLE. We show repre￾sentative trajectories in diverse scenarios, including adverse-weather scenes with limited visibility and sudden pedestrian crossings (top row), and clear suburban traffic with dynamic agents such as cyclists and surrounding vehicles (bottom row). Blue curves denote the planned ego-vehicle trajectory, highlighting smooth … view at source ↗
Figure 5
Figure 5. Figure 5: BEV qualitative comparison on Bench2Drive (closed-loop). Bird’s-eye-view visualiza￾tion for the same route/scenario (RouteScenario_25951_rep0, HazardAtSideLaneTwoWays_1, weather_id=7). Left: ReCogDrive [27]. Right: MAPLE (ours). The planned ego trajectory is overlaid, illustrating different interaction outcomes in the same context [PITH_FULL_IMAGE:figures/full_fig_p016_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Closed-loop rollout comparison on Bench2Drive. Multi-frame qualitative roll￾outs for the same route/scenario (RouteScenario_25951_rep0, HazardAtSideLaneTwoWays_1, weather_id=7). Top row: ReCogDrive. Bottom row: MAPLE. Colored curves denote the planned ego trajectory across time, highlighting differences in closed-loop interaction behavior. containing dynamic agents (e.g., two cyclists traveling along the r… view at source ↗
Figure 7
Figure 7. Figure 7: Additional qualitative closed-loop driving examples on Bench2Drive using MAPLE. These examples includes challenging conditions, like low-light/night driving with sudden pedestrian appearances and wet-road reflections, dense fog/highway driving with reduced visibility, and urban scenes with adverse weather. Blue/cyan curves denote the planned ego trajectory [PITH_FULL_IMAGE:figures/full_fig_p017_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Additional qualitative closed-loop driving examples on Bench2Drive using MAPLE. More examples covering suburban/rural traffic with oncoming vehicles and lane curvature, as well as nighttime intersection scenarios with wet-road conditions and surrounding traffic. Blue/cyan curves denote the planned ego trajectory. gradual curvature and without abrupt corrections [PITH_FULL_IMAGE:figures/full_fig_p017_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Failure case: over-cautious avoidance leading to lane departure (Bench2Drive, closed￾loop). In this scenario, MAPLE performs an overly conservative unprotected left turn to avoid a potential collision, resulting in a brief deviation outside the route lanes for about 1.0 meters (1.29% of the full route). The car quickly moves back to the lane after this brief deviation. Blue/cyan curves denote the planned e… view at source ↗
read the original abstract

Vision-language-action (VLA) models are effective as end-to-end motion planners, but can be brittle when evaluated in closed-loop settings due to being trained under traditional imitation learning framework. Existing closed-loop supervision approaches lack scalability and fail to completely model a reactive environment. We propose MAPLE, a novel framework for reactive, multi-agent rollout of a dynamic driving scenario in the latent space of the VLA model. The ego vehicle and nearby traffic agents are independently controlled over multi-step horizons, while being reactive to other agents in the scene, enabling closed-loop training. MAPLE consists of two training stages: (1) supervised fine-tuning on the latent rollouts based on ground-truth trajectories, followed by (2) reinforcement learning with global and agent -specific rewards that encourage safety, progress, and interaction realism. We further propose diversity rewards that encourage the model to generate planning behaviors that may not be present in logged driving data. Notably, our closed-loop training framework is scalable and does not require external simulators, which can be computationally expensive to run and have limited visual fidelity to the real-world. MAPLE achieves state-of-the-art driving performance on Bench2Drive and demonstrates scalable, closed-loop multi-agent play for robust E2E autonomous driving systems.

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 / 0 minor

Summary. The manuscript proposes MAPLE, a two-stage framework for training vision-language-action (VLA) models for end-to-end autonomous driving. Stage 1 performs supervised fine-tuning on latent-space multi-agent rollouts generated from ground-truth trajectories, with the ego vehicle and nearby agents controlled independently over multi-step horizons while remaining reactive to each other. Stage 2 applies reinforcement learning using global and agent-specific rewards for safety, progress, interaction realism, and diversity. The method claims to enable scalable closed-loop training without external simulators and achieves state-of-the-art performance on the Bench2Drive benchmark.

Significance. If the latent rollouts are shown to faithfully reproduce reactive multi-agent dynamics, the approach would offer a scalable alternative to simulator-based closed-loop training for VLA models, potentially improving robustness over pure imitation learning while avoiding high computational costs and visual fidelity limitations of external simulators. The inclusion of diversity rewards to encourage behaviors beyond logged data is a positive element for exploration.

major comments (2)
  1. [Section 3] Section 3: The central claim that latent-space rollouts enable reactive multi-agent play for closed-loop RL rests on the unverified assumption that these rollouts accurately capture real-world dynamics. No quantitative validation is provided, such as prediction error metrics against ground-truth trajectories, distribution matching statistics, or ablation studies on rollout horizon length.
  2. [Section 3] Section 3: Without external grounding or visual fidelity losses, any reported SOTA on Bench2Drive could arise from reduced train-test mismatch within the model's latent biases rather than genuine reactivity gains; this requires explicit checks (e.g., closed-loop vs. open-loop performance deltas or cross-validation on held-out real trajectories) to support the scalability and robustness claims.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the thoughtful and constructive comments on our work. We agree that stronger empirical validation of the latent rollouts' fidelity would better support the central claims. We address each major comment below and will incorporate the suggested analyses in the revised manuscript.

read point-by-point responses

  1. Referee: [Section 3] Section 3: The central claim that latent-space rollouts enable reactive multi-agent play for closed-loop RL rests on the unverified assumption that these rollouts accurately capture real-world dynamics. No quantitative validation is provided, such as prediction error metrics against ground-truth trajectories, distribution matching statistics, or ablation studies on rollout horizon length.

    Authors: We acknowledge the value of direct quantitative validation. In the revision we will add (i) per-step and multi-step prediction error metrics (L2 displacement and heading error) between latent rollouts and ground-truth trajectories on held-out Bench2Drive sequences, (ii) distribution-matching statistics (e.g., Wasserstein distance on velocity and acceleration histograms), and (iii) an ablation table varying rollout horizon length (1, 3, 5, 8 steps) that reports both training stability and final closed-loop driving metrics. These additions will quantify how faithfully the latent dynamics reproduce reactive multi-agent behavior. revision: yes

  2. Referee: [Section 3] Section 3: Without external grounding or visual fidelity losses, any reported SOTA on Bench2Drive could arise from reduced train-test mismatch within the model's latent biases rather than genuine reactivity gains; this requires explicit checks (e.g., closed-loop vs. open-loop performance deltas or cross-validation on held-out real trajectories) to support the scalability and robustness claims.

    Authors: We will include two new experiments in the revision: (1) a direct closed-loop versus open-loop comparison on the full Bench2Drive test set, reporting the performance delta attributable to our multi-agent RL stage, and (2) cross-validation results on a held-out set of real-world trajectories (distinct from the training distribution) that measure both open-loop imitation accuracy and closed-loop success rate. These checks will demonstrate that the observed SOTA gains stem from improved reactivity rather than latent-space overfitting. revision: yes

Circularity Check

0 steps flagged

No significant circularity in MAPLE derivation chain

full rationale

The paper presents a two-stage pipeline of supervised fine-tuning on ground-truth latent trajectories followed by RL with hand-designed rewards for safety, progress, interaction realism, and diversity. No equations appear in the manuscript that would reduce any claimed prediction or performance gain to a fitted parameter or input by construction. The latent multi-agent rollout is introduced as a novel mechanism without invoking self-citation load-bearing uniqueness theorems or ansatzes smuggled from prior author work. The central claims rest on empirical Bench2Drive results rather than any self-referential redefinition of inputs as outputs, making the derivation self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract-only; no explicit free parameters, axioms, or invented entities are stated. Latent rollout fidelity is implicitly assumed but not formalized.

pith-pipeline@v0.9.0 · 5571 in / 1030 out tokens · 19879 ms · 2026-05-15T04:38:17.797142+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

50 extracted references · 50 canonical work pages · 6 internal anchors

  1. [1]

    Learning dexterous in-hand manipulation

    Marcin Andrychowicz, Bowen Baker, Maciek Chociej, Rafal Józefowicz, Bob McGrew, Jakub Pachocki, Arthur Petron, Matthias Plappert, Glenn Powell, Alex Ray, et al. Learning dexterous in-hand manipulation. The International Journal of Robotics Research, 39(1):3–20, 2020

    work page 2020

  2. [2]

    Qwen-VL: A Versatile Vision-Language Model for Understanding, Localization, Text Reading, and Beyond

    Jinze Bai, Shuai Bai, Shusheng Yang, Shijie Wang, Sinan Tan, Peng Wang, Junyang Lin, Chang Zhou, and Jingren Zhou. Qwen-vl: A versatile vision-language model for understanding, localization, text reading, and beyond.arXiv preprint arXiv:2308.12966, 2023

    work page internal anchor Pith review Pith/arXiv arXiv 2023

  3. [3]

    Qwen2.5-VL Technical Report

    Shuai Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, Sibo Song, Kai Dang, Peng Wang, Shijie Wang, Jun Tang, Humen Zhong, Yuanzhi Zhu, Mingkun Yang, Zhaohai Li, Jianqiang Wan, Pengfei Wang, Wei Ding, Zheren Fu, Yiheng Xu, Jiabo Ye, Xi Zhang, Tianbao Xie, Zesen Cheng, Hang Zhang, Zhibo Yang, Haiyang Xu, and Junyang Lin. Qwen2.5-vl technical report.ar...

    work page internal anchor Pith review Pith/arXiv arXiv 2025

  4. [4]

    SimNet: Learning reactive self-driving simulations from real-world observations

    Luca Bergamini, Yawei Ye, Oliver Scheel, Long Chen, Chih-Yuan Hu, Luca Delévaux, Niels Muller, and Peter Ondruska. SimNet: Learning reactive self-driving simulations from real-world observations. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2021

    work page 2021

  5. [5]

    Killian, Stuart Bowers, Ozan Sener, Philipp Kraehenbuehl, and Vladlen Koltun

    Marco Cusumano-Towner, David Hafner, Alexander Hertzberg, Brody Huval, Aleksei Petrenko, Eugene Vinitsky, Erik Wijmans, Taylor W. Killian, Stuart Bowers, Ozan Sener, Philipp Kraehenbuehl, and Vladlen Koltun. Robust autonomy emerges from self-play. InProceedings of the 42nd International Conference on Machine Learning (ICML), 2025

    work page 2025

  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

    work page 2023

  7. [7]

    Carla: An open urban driving simulator

    Alexey Dosovitskiy, German Ros, Felipe Codevilla, Antonio Lopez, and Vladlen Koltun. Carla: An open urban driving simulator. InConference on robot learning, pages 1–16. PMLR, 2017

    work page 2017

  8. [8]

    Eva: Exploring the limits of masked visual representation learning at scale

    Yuxin Fang, Wen Wang, Binhui Xie, Quan Sun, Ledell Wu, Xinggang Wang, Tiejun Huang, Xinlong Wang, and Yue Cao. Eva: Exploring the limits of masked visual representation learning at scale. InProceedings of the IEEE/CVF conference on computer vision and pattern recognition, pages 19358–19369, 2023

    work page 2023

  9. [9]

    Orion: A holistic end-to-end autonomous driving framework by vision-language instructed action generation.arXiv preprint arXiv:2503.19755,

    Haoyu Fu, Diankun Zhang, Zongchuang Zhao, Jianfeng Cui, Dingkang Liang, Chong Zhang, Dingyuan Zhang, Hongwei Xie, Bing Wang, and Xiang Bai. Orion: A holistic end-to-end autonomous driving framework by vision-language instructed action generation.arXiv preprint arXiv:2503.19755, 2025

    work page arXiv 2025

  10. [10]

    Co-Reyes, Rishabh Agarwal, Rebecca Roelofs, Yao Lu, Nico Montali, Paul Mougin, Zoey Yang, Brandyn White, Aleksandra Faust, Rowan McAllister, Dragomir Anguelov, and Benjamin Sapp

    Cole Gulino, Justin Fu, Wenjie Luo, George Tucker, Eli Bronstein, Yiren Lu, Jean Harb, Xinlei Pan, Yan Wang, Xiangyu Chen, John D. Co-Reyes, Rishabh Agarwal, Rebecca Roelofs, Yao Lu, Nico Montali, Paul Mougin, Zoey Yang, Brandyn White, Aleksandra Faust, Rowan McAllister, Dragomir Anguelov, and Benjamin Sapp. Waymax: An accelerated, data-driven simulator f...

    work page 2023

  11. [11]

    Tan et al

    K. Tan et al. H. Caesar, J. Kabzan. Nuplan: A closed-loop ml-based planning benchmark for autonomous vehicles. InCVPR ADP3 workshop, 2021

    work page 2021

  12. [12]

    Social force model for pedestrian dynamics.Physical Review E, 51(5): 4282–4286, 1995

    Dirk Helbing and Péter Molnár. Social force model for pedestrian dynamics.Physical Review E, 51(5): 4282–4286, 1995

    work page 1995

  13. [13]

    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. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17853–17862, 2023

    work page 2023

  14. [14]

    Gen-drive: Enhancing diffusion generative driving policies with reward modeling and reinforcement learning fine-tuning

    Zhiyu Huang, Xinshuo Weng, Maximilian Igl, Yuxiao Chen, Yulong Cao, Boris Ivanovic, Marco Pavone, and Chen Lv. Gen-drive: Enhancing diffusion generative driving policies with reward modeling and reinforcement learning fine-tuning. In2025 IEEE International Conference on Robotics and Automation (ICRA), pages 3445–3451. IEEE, 2025

    work page 2025

  15. [15]

    EMMA: End-to-End Multimodal Model for Autonomous Driving

    Jyh-Jing Hwang, Runsheng Xu, Hubert Lin, Wei-Chih Hung, Jingwei Ji, Kristy Choi, Di Huang, Tong He, Paul Covington, Benjamin Sapp, Yin Zhou, James Guo, Dragomir Anguelov, and Mingxing Tan. Emma: End-to-end multimodal model for autonomous driving.arXiv preprint arXiv:2410.23262, 2024

    work page internal anchor Pith review arXiv 2024

  16. [16]

    Carl: Learning scalable planning policies with simple rewards.arXiv preprint arXiv:2504.17838, 2025

    Bernhard Jaeger, Daniel Dauner, Jens Beißwenger, Simon Gerstenecker, Kashyap Chitta, and Andreas Geiger. Carl: Learning scalable planning policies with simple rewards.arXiv preprint arXiv:2504.17838, 2025. 10

    work page arXiv 2025

  17. [17]

    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. InICCV, 2023

    work page 2023

  18. [18]

    Think twice before driving: Towards scalable decoders for end-to-end autonomous driving

    Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think twice before driving: Towards scalable decoders for end-to-end autonomous driving. InCVPR, 2023

    work page 2023

  19. [19]

    Bench2drive: Towards multi- ability benchmarking of closed-loop end-to-end autonomous driving

    Xiaosong Jia, Zhenjie Yang, Qifeng Li, Zhiyuan Zhang, and Junchi Yan. Bench2drive: Towards multi- ability benchmarking of closed-loop end-to-end autonomous driving. InNeurIPS 2024 Datasets and Benchmarks Track, 2024

    work page 2024

  20. [20]

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

    Xiaosong Jia, Junqi You, Zhiyuan Zhang, and Junchi Yan. Drivetransformer: Unified transformer for scalable end-to-end autonomous driving. InThe Thirteenth International Conference on Learning Representations, 2025

    work page 2025

  21. [21]

    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. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 8340–8350, 2023

    work page 2023

  22. [22]

    Senna: Bridging large vision-language models and end-to-end autonomous driving.arXiv preprint arXiv:2410.22313, 2024a

    Bo Jiang, Shaoyu Chen, Bencheng Liao, Xingyu Zhang, Wei Yin, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Senna: Bridging large vision-language models and end-to-end autonomous driving. arXiv preprint arXiv:2410.22313, 2024

    work page arXiv 2024

  23. [23]

    Alphadrive: Unleashing the power of vlms in autonomous driving via reinforcement learning and reasoning.arXiv preprint arXiv:2503.07608, 2025

    Bo Jiang, Shaoyu Chen, Qian Zhang, Wenyu Liu, and Xinggang Wang. Alphadrive: Unleashing the power of vlms in autonomous driving via reinforcement learning and reasoning.arXiv preprint arXiv:2503.07608, 2025

    work page arXiv 2025

  24. [24]

    Beyond behavior cloning in autonomous driving: a survey of closed-loop training techniques

    Peter Karkus, Maximilian Igl, Yuxiao Chen, Kashyap Chitta, Boris Ivanovic, and Marco Pavone. Beyond behavior cloning in autonomous driving: a survey of closed-loop training techniques. Technical report, NVIDIA Research, 2025

    work page 2025

  25. [25]

    A survey of generalisation in deep reinforcement learning.arXiv preprint arXiv:2111.09794, 2023

    Roberta Kirk, Amy Zhang, Edward Grefenstette, and Tim Rocktäschel. A survey of generalisation in deep reinforcement learning.arXiv preprint arXiv:2111.09794, 2023

    work page arXiv 2023

  26. [26]

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

    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

    work page arXiv 2025

  27. [27]

    Recogdrive: A reinforced cognitive framework for end-to-end autonomous driving

    Yongkang Li, Kaixin Xiong, Xiangyu Guo, Fang Li, Sixu Yan, Gangwei Xu, Lijun Zhou, Long Chen, Haiyang Sun, Bing Wang, et al. Recogdrive: A reinforced cognitive framework for end-to-end autonomous driving. InInternational Conference on Learning Representations (ICLR), 2026

    work page 2026

  28. [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. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 12037–12047, 2025

    work page 2025

  29. [29]

    Reinforced refinement with self-aware expansion for end-to-end autonomous driving.arXiv preprint arXiv:2506.09800,

    Haochen Liu, Tianyu Li, Haohan Yang, Li Chen, Caojun Wang, Ke Guo, Haochen Tian, Hongchen Li, Hongyang Li, and Chen Lv. Reinforced refinement with self-aware expansion for end-to-end autonomous driving.arXiv preprint arXiv:2506.09800, 2025

    work page arXiv 2025

  30. [30]

    Decoupled weight decay regularization

    Ilya Loshchilov and Frank Hutter. Decoupled weight decay regularization. InInternational Conference on Learning Representations, 2019

    work page 2019

  31. [31]

    Gpt-driver: Learning to drive with gpt.arXiv preprint arXiv:2310.01415, 2023a

    Jiageng Mao, Yuxi Qian, Junjie Ye, Hang Zhao, and Yue Wang. Gpt-driver: Learning to drive with gpt. arXiv preprint arXiv:2310.01415, 2023

    work page arXiv 2023

  32. [32]

    Generating useful accident- prone driving scenarios via a learned traffic prior

    Davis Rempe, Jonah Philion, Leonidas J Guibas, Sanja Fidler, and Or Litany. Generating useful accident- prone driving scenarios via a learned traffic prior. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2022

    work page 2022

  33. [33]

    Simlingo: Vision-only closed-loop autonomous driving with language-action alignment

    Katrin Renz, Long Chen, Elahe Arani, and Oleg Sinavski. Simlingo: Vision-only closed-loop autonomous driving with language-action alignment. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 11993–12003, 2025

    work page 2025

  34. [34]

    DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models

    Zhihong Shao, Peiyi Wang, Qihao Zhu, Runxin Xu, Junxiao Song, Xiao Bi, Haowei Zhang, Mingchuan Zhang, YK Li, Yang Wu, et al. Deepseekmath: Pushing the limits of mathematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024. 11

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  35. [35]

    Motion transformer with global intention localization and local movement refinement.arXiv preprint arXiv:2209.13508, 2022

    Shaoshuai Shi, Li Jiang, Dengxin Dai, and Bernt Schiele. Motion transformer with global intention localization and local movement refinement.arXiv preprint arXiv:2209.13508, 2022

    work page arXiv 2022

  36. [36]

    Mastering the game of Go without human knowledge.Nature, 550:354–359, 2017

    David Silver, Julian Schrittwieser, Karen Simonyan, Ioannis Antonoglou, Aja Huang, Arthur Guez, Thomas Hubert, Lucas Baker, Matthew Lai, Adrian Bolton, et al. Mastering the game of Go without human knowledge.Nature, 550:354–359, 2017

    work page 2017

  37. [37]

    Don’t shake the wheel: Momentum-aware planning in end-to-end autonomous driving

    Ziying Song, Caiyan Jia, Lin Liu, Hongyu Pan, Yongchang Zhang, Junming Wang, Xingyu Zhang, Shaoqing Xu, Lei Yang, and Yadan Luo. Don’t shake the wheel: Momentum-aware planning in end-to-end autonomous driving. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 22432–22441, 2025

    work page 2025

  38. [38]

    TrafficSim: Learning to simulate realistic multi-agent behaviors

    Simon Suo, Sebastian Regalado, Sergio Casas, and Raquel Urtasun. TrafficSim: Learning to simulate realistic multi-agent behaviors. InProceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 10400–10409, 2021

    work page 2021

  39. [39]

    Hip-ad: Hierarchical and multi-granularity planning with deformable 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 deformable attention for autonomous driving in a single decoder. InProceedings of the IEEE/CVF International Conference on Computer Vision, pages 25605–25615, 2025

    work page 2025

  40. [40]

    DriveVLM: The Convergence of Autonomous Driving and Large Vision-Language Models

    Xiaoyu Tian, Junru Gu, Bailin Li, Yicheng Liu, Chenxu Hu, Yang Wang, Kun Zhan, Peng Jia, Xianpeng Lang, and Hang Zhao. Drivevlm: The convergence of autonomous driving and large vision-language models.arXiv preprint arXiv:2402.12289, 2024

    work page internal anchor Pith review Pith/arXiv arXiv 2024

  41. [41]

    Grandmaster level in StarCraft II using multi-agent reinforcement learning.Nature, 575:350–354, 2019

    Oriol Vinyals, Igor Babuschkin, Wojciech M Czarnecki, Michaël Mathieu, Andrew Dudzik, Junyoung Chung, David H Choi, Richard Powell, Timo Ewalds, et al. Grandmaster level in StarCraft II using multi-agent reinforcement learning.Nature, 575:350–354, 2019

    work page 2019

  42. [42]

    Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    Jason Wei, Xuezhi Wang, Dale Schuurmans, Maarten Bosma, Fei Xia, Ed Chi, Quoc V Le, Denny Zhou, et al. Chain-of-thought prompting elicits reasoning in large language models.Advances in neural information processing systems, 35:24824–24837, 2022

    work page 2022

  43. [43]

    Trajectory-guided con- trol prediction for end-to-end autonomous driving: A simple yet strong baseline.Advances in Neural Information Processing Systems, 35:6119–6132, 2022

    Penghao Wu, Xiaosong Jia, Li Chen, Junchi Yan, Hongyang Li, and Yu Qiao. Trajectory-guided con- trol prediction for end-to-end autonomous driving: A simple yet strong baseline.Advances in Neural Information Processing Systems, 35:6119–6132, 2022

    work page 2022

  44. [44]

    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. InProceedings of the Computer Vision and Pattern Recognition Conference, pages 1602–1611, 2025

    work page 2025

  45. [45]

    Generative scenario rollouts for end-to-end autonomous driving.arXiv preprint arXiv:2601.11475, 2026

    Rajeev Yasarla, Deepti Hegde, Shizhong Han, Hsin-Pai Cheng, Yunxiao Shi, Meysam Sadeghigooghari, Shweta Mahajan, Apratim Bhattacharyya, Litian Liu, Risheek Garrepalli, et al. Generative scenario rollouts for end-to-end autonomous driving.arXiv preprint arXiv:2601.11475, 2026

    work page arXiv 2026

  46. [46]

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

    Liuhan Yin, Runkun Ju, Guodong Guo, and Erkang Cheng. Diffrefiner: Coarse to fine trajectory planning via diffusion refinement with semantic interaction for end to end autonomous driving. InProceedings of the AAAI Conference on Artificial Intelligence, pages 12009–12017, 2026

    work page 2026

  47. [47]

    CAT: Closed-loop adversarial training for safe end-to-end driving

    Linrui Zhang, Zhenghao Peng, Quanyi Li, and Bolei Zhou. CAT: Closed-loop adversarial training for safe end-to-end driving. InConference on Robot Learning, 2023

    work page 2023

  48. [48]

    Diffusion-based planning for autonomous driving with flexible guidance.arXiv preprint arXiv:2501.15564,

    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

    work page arXiv 2025

  49. [49]

    Query-centric trajectory prediction

    Zikang Zhou, Jianping Wang, Yung-Hui Li, and Yu-Kai Huang. Query-centric trajectory prediction. In Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, pages 17863– 17873, 2023

    work page 2023

  50. [50]

    AutoVLA: A Vision-Language-Action Model for End-to-End Autonomous Driving with Adaptive Reasoning and Reinforcement Fine-Tuning

    Zewei Zhou, Tianhui Cai, Seth Z Zhao, Yun Zhang, Zhiyu Huang, Bolei Zhou, and Jiaqi Ma. Au- tovla: A vision-language-action model for end-to-end autonomous driving with adaptive reasoning and reinforcement fine-tuning.arXiv preprint arXiv:2506.13757, 2025. 12 A Ablation Study A.1 Number of Reactive Agents Agent Distribution in Bench2Drive.To contextualize...

    work page internal anchor Pith review Pith/arXiv arXiv 2025