Pith. sign in

REVIEW 4 major objections 4 minor 10 cited by

Treating driving as a choice among 13 linguistic meta-actions lets a 0.5B vision-language-action model learn online by trial and error, reaching 78.04 Driving Score and 55.09% Success Rate on Bench2Drive.

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

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

T0 review · deepseek-v4-flash

2026-08-03 16:23 UTC pith:F3W7Q7Q4

load-bearing objection Solid architecture, overclaimed headline: RL training routes are inside the evaluation set, and the true held-out gain is about half the reported one. the 4 major comments →

arxiv 2512.13636 v4 pith:F3W7Q7Q4 submitted 2025-12-15 cs.CV cs.RO

MindDrive: A Vision-Language-Action Model for Autonomous Driving via Online Reinforcement Learning

Haoyu Fu , Diankun Zhang , Zongchuang Zhao , Jianfeng Cui , Hongwei Xie , Bing Wang , Guang Chen , Hangjun Ye
show 2 more authors
Dingkang Liang Xiang Bai
This is my paper
classification cs.CV cs.RO
keywords vision-language-action modelonline reinforcement learningautonomous drivingmeta-actionsBench2DriveCARLAPPOLoRA
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

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

MindDrive argues that online reinforcement learning becomes practical for vision-language-action (VLA) driving models if the continuous trajectory space is compressed into a small set of linguistic meta-actions. A single small language model, split by two LoRA adapters into a Decision Expert and an Action Expert, first learns from imitation to map each meta-action to a concrete trajectory, then is fine-tuned online with PPO in the CARLA simulator using only sparse success/failure rewards. On the Bench2Drive benchmark this raises Driving Score from 75.85 to 78.04 and Success Rate from 49.30% to 55.09% over the same model trained by imitation alone, with a 0.5B-parameter LLM. The paper presents this as the first demonstration that online RL can improve a VLA autonomous-driving model.

Core claim

The central claim is that trajectory-level reward can be back-propagated into a language model's reasoning by treating driving as a discrete decision problem: the Decision Expert emits one of 7 speed meta-actions and 6 path meta-actions, the Action Expert converts that choice into speed and path waypoints through a shared LLM with separate LoRA weights, and PPO updates only the Decision Expert's policy. The authors show that RL fine-tuning strengthens causal reasoning in interactive scenarios such as overtaking and give-way, and that the split architecture prevents the action decoder from collapsing when trained with sparse rewards. Under the paper's evaluation, the resulting 0.5B model outp

What carries the argument

The load-bearing mechanism is the language-action loop: a base LLM (Qwen2-0.5B) carries two LoRA adapters, one for a Decision Expert emitting discrete meta-actions (speed: stop, speed up, slow down, maintain slow/moderate/fast; path: turn, lane change, straight, lane follow), and one for an Action Expert that maps the chosen meta-actions into 6 speed waypoints and 20 path waypoints through a small VAE/GRU decoder. The meta-action token space makes exploration discrete and cheap, while the Action Expert keeps trajectories smooth and human-like. The closed-loop training pipeline precomputes vision-encoder state embeddings, stores them in a buffer, and applies PPO with generalized advantage est

Load-bearing premise

The paper's quantitative case rests on the assumption that its online-RL results are evaluated fairly: the 44 routes used for RL rollout are a subset of the 220 evaluation routes, and the main table reports aggregate scores without holding them out, so part of the reported 2.19 DS / 5.79% SR gain may be re-measuring routes the model was trained on.

What would settle it

Re-run the Bench2Drive evaluation restricted to the 176 routes that were not used in RL rollout and compare MindDrive-RL, MindDrive-IL, and the 0.5B imitation baseline under identical conditions. If the DS/SR gap between MindDrive-RL and MindDrive-IL shrinks to near zero on those held-out routes, the paper's central claim of generalizable online-RL improvement would be falsified; if the gap persists, the claim stands.

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

If this is right

  • Online RL over discrete meta-actions can be combined with any LLM-based decision backbone; the framework is not tied to a particular trajectory decoder.
  • Because only the Decision Expert's LoRA is updated, RL improves high-level reasoning without destroying low-level action quality, addressing catastrophic forgetting observed in single-expert RL.
  • Sparse success/failure rewards are sufficient to produce measurable gains, suggesting handcrafted dense reward engineering is not required for this paradigm.
  • The 5.8-point success-rate gain on the benchmark's full 220 routes, if it holds under held-out evaluation, would make online RL a practical complement to imitation learning in closed-loop autonomous-driving pipelines.

Where Pith is reading between the lines

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

  • [Editorial inference] The headline comparison would be strengthened by reporting RL gains on the 176 routes that were never used for rollout; the appendix's split shows a 1.06 DS / 4.7% SR gain on non-rollout routes versus 5.29 DS / 9.09% on rollout routes, so the generalizable portion is smaller than the headline delta.
  • [Editorial inference] The discrete meta-action vocabulary (7 speed × 6 path) is likely the bottleneck for richer behaviors such as negotiation at unprotected turns; a testable extension is to add meta-actions for yield/wait/creep and check whether online RL can learn to use them safely.
  • [Editorial inference] The same architecture should transfer to other closed-loop simulators or to real-world shadow-mode data collection, where the sparse reward would be replaced by human-disengagement events; the paper's claim about online RL would then be testable outside CARLA.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 4 minor

Summary. The paper proposes MindDrive, a vision-language-action (VLA) driving model that decouples decision-making from trajectory generation. A single Qwen2-0.5B LLM is equipped with two LoRA adapters: a Decision Expert that outputs discrete linguistic meta-actions (7 speed actions, 6 path actions) and an Action Expert that maps these meta-actions to continuous trajectories via a VAE/GRU decoder. The model is first trained with imitation learning (IL) on the Bench2Drive base set, then fine-tuned with online PPO in the CARLA simulator using sparse rewards (+1 for arrival, -1 for collisions/red-light/route-deviation/stop violations). The main claim is that this online-RL scheme improves closed-loop driving beyond IL at the same model scale, reporting DS 78.04 and SR 55.09 on the 220-route Bench2Drive benchmark, outperforming MindDrive-IL by 2.19 DS and 5.79 SR. The paper also claims to be the first to demonstrate online RL for a VLA model in autonomous driving.

Significance. If the central claim is validated on a properly held-out evaluation, the paper makes a meaningful contribution: it shows that online RL over a discrete language of driving decisions, rather than over continuous trajectories, can improve a lightweight VLA policy's closed-loop performance. The decoupling of reasoning (Decision Expert) from trajectory generation (Action Expert) is a sensible design that addresses exploration efficiency, and the use of cached visual embeddings to make RL data collection tractable is practically useful. The ablation studies on penalty events, rollout rounds, and policy regularization are extensive and internally consistent, and the qualitative results support the intended mechanism. The main risk is that the headline comparison is not an apples-to-apples held-out comparison because the RL training routes overlap with the evaluation set; the authors' own Appendix D.2 shows the held-out gains are substantially smaller. The contribution is therefore plausible but currently overstated.

major comments (4)
  1. [Sec. 4.1, Appendix C, Appendix D.2 (Tab. A2)] The headline comparison in Table 1 (MindDrive vs MindDrive-IL: +2.19 DS, +5.79 SR) is contaminated by overlap between the RL training routes and the evaluation set. Sec. 4.1 states that 44 routes were selected for RL, and Appendix C confirms these are 'routes that the model is able to successfully complete through sampled actions.' Appendix D.2 then explicitly reports results on 'Rollout Routes' and 'Other Routes,' confirming that the 44 rollout routes are a subset of the 220-route Bench2Drive benchmark. Table A2 shows the gain on Rollout Routes is +5.29 DS and +9.09 SR, while on Other Routes it is only +1.06 DS and +4.7 SR. Thus the abstract and Sec. 4.2's claim of effectiveness rests partly on re-measuring routes on which the RL policy was trained, while the IL baseline never received RL training on them. The evaluation must be reported on the held-out 'Other Routes' as the primary met
  2. [Tables 1, 3, Fig. 4] All reported closed-loop metrics appear to come from a single evaluation run, with no standard deviation, confidence intervals, or multiple seeds. CARLA and the Bench2Drive evaluation are stochastic; the 2.19 DS and 5.79 SR gains over MindDrive-IL, as well as the differences between PPO variants in Table 3, could be within run-to-run variability. The authors should provide variance estimates over at least three seeds, or route-level bootstrap confidence intervals, particularly for the central RL-vs-IL comparison and for the rollout-round ablation in Fig. 4. Without this, the quantitative claims are not fully supported.
  3. [Sec. 4.3 vs Appendix C] There is a direct contradiction in the number of rollouts per route. Sec. 4.3 states 'two online RL rollouts are performed per route unless specified otherwise' and Fig. 4 is presented as an ablation on 'the number of Rollout Rounds in RL,' with the default set to 2. Appendix C, however, states 'each routes rollout five times' for the main data collection. This is not a minor wording issue: it determines what the reported results actually used, and the Fig. 4 ablation is uninterpretable if the default was 5 rather than 2. Please state clearly the number of rollouts per route used for the main results and for each ablation row.
  4. [Sec. 3.3, Eq. (15)] The sentence 'During training, only the parameters of the MLP head within the value net are updated' is inconsistent with Eq. (15), where the total RL loss includes L_ppo, an objective over the policy π_d's parameters. If the policy network is not updated during RL, the reported improvements cannot be explained by PPO; if it is updated, the sentence is incorrect and should be reworded. Please clarify which parameters are updated for the policy and which for the value network.
minor comments (4)
  1. [Eq. (3)] The factorization π_c(a) = π_d(a|V,L) · π_g(a|V,L) is presented as an equality, but it is a design choice rather than a formal decomposition; the arg max over the product is not in general equivalent to the arg max of a scoring function. Clarify that this is a modeling choice for decoupling selection and generation.
  2. [Abstract / Intro] The claim 'this is the first work to demonstrate the effectiveness of online reinforcement learning for the VLA model in autonomous driving' is too strong given the related work (e.g., AlphaDrive) and the route-overlap caveat. Please qualify the novelty claim to 'first, to our knowledge, within this specific architecture/evaluation setup' or soften it in light of the needed held-out verification.
  3. [References] Several references contain the typo 'Porc.' instead of 'Proc.' (e.g., references [8], [10], [14], [25], [33], [61]). Please correct throughout.
  4. [Fig. A3 / Sec. D.2] The route IDs in Fig. A3 are hard to read and it is not clear how they map to the 220 benchmark routes. The caption should state whether these are official Bench2Drive route IDs and whether they are a subset of the evaluation set.

Circularity Check

1 steps flagged

Evaluation set includes the 44 online-RL rollout routes, so the headline RL-vs-IL gain is partly a train-set measurement; the held-out gain is smaller.

specific steps
  1. fitted input called prediction [Sec. 4.1 (Experimental Settings) / Appendix C / Appendix D.2, Table A2]
    "For the online reinforcement learning (RL), we selected 44 routes that the model successfully completed through action sampling (see the Appendix for details). ... We select 44 routes that the model is able to successfully complete through sampled actions ... To evaluate the effectiveness of reinforcement learning, we analyze the model’s driving performance over both the rollout routes and the remaining routes. ... On the remaining routes, our model still achieves improvements of 1.06 in driving score and 4.7% in success rate over imitation learning."

    The 44 routes used for PPO rollouts are part of the 220-route Bench2Drive evaluation on which the headline 78.04 DS / 55.09% SR and the 2.19 DS / 5.79 SR gain over MindDrive-IL are computed. Appendix D.2 confirms that the benchmark is split into 'Rollout Routes' and 'Other Routes.' Therefore the main comparison includes routes on which the RL policy was explicitly trained, while the IL baseline was not RL-trained on them. The true held-out gain (Other Routes) is only 1.06 DS and 4.7 SR, so the headline RL effectiveness claim is partly a re-measurement of training routes rather than an independent prediction.

full rationale

No equation in the paper reduces a derived quantity to a fitted constant, and the architecture is not self-definitional. The load-bearing comparison is empirical: online RL is claimed to improve VLA driving over IL. That comparison is compromised, but not destroyed, by the fact that the 44 rollout routes are a subset of the 220-route evaluation benchmark and the main table aggregates both rollout and held-out routes. Because the paper itself discloses the split in Appendix D.2 and shows a positive (if much smaller) gain on held-out routes, the central claim has independent content. Self-citations to ORION [10] are used as a component source and baseline, but the baseline is reproduced from official code, so they are not load-bearing circularity. Score 3 reflects one evaluation-loop overlap that partially conflates training and test routes without making the whole derivation circular.

Axiom & Free-Parameter Ledger

6 free parameters · 6 axioms · 1 invented entities

The reported system is largely an empirical ML pipeline. The main free choices are the reward/penalty set, the rollout-round count, the 44-route RL training set, and the hand-built meta-action vocabulary. The most consequential is the 44-route selection, which overlaps the evaluation benchmark and is described inconsistently in the main text and appendix.

free parameters (6)
  • RL rollout rounds = 2
    Authors test 1-5 rollout rounds and select 2 because DS/SR peak (Fig. 4); the reported configuration is tuned to the evaluation benchmark.
  • Penalty event set = Collision, Traffic Light, Route Deviation, Stop
    Added progressively during ablation (Tab. 2); the final set is the one with best DS/SR, so reward design is benchmark-fitted.
  • PPO hyperparameters (gamma, lambda, clip, batch, KL weight, value loss weight) = 0.99, 1.0, 0.2, 32, 0.5, 0.5
    Set by hand (Tab. A1) with no sensitivity analysis; needed for the policy/value update.
  • Number of RL training routes = 44
    Selected from the benchmark because the IL model could complete them via action sampling (Appendix C); this controls training data and overlaps the evaluation set.
  • Meta-action vocabulary = 7 speed + 6 path actions
    Manually defined and filtered; the size of the discrete decision space affects exploration and expressiveness.
  • LoRA rank/alpha = 16/16
    Standard hyperparameter pair, no ablation reported.
axioms (6)
  • domain assumption Bench2Drive/CARLA closed-loop metrics (DS, SR, Multi-Ability) measure driving competence.
    All conclusions treat simulator scores as the objective; no real-world validation exists (Sec. 5 Limitations).
  • ad hoc to paper Sparse reward (+1 arrival, -1 for collision/red light/route deviation/stop violation, 0 otherwise) is a sufficient training signal.
    Eq. 9 defines the reward; no analysis of reward noise or shaping alternatives.
  • domain assumption Meta-actions generated by Qwen2VL-72B and manually filtered form a one-to-one language-action mapping.
    Appendix B describes LLM generation plus manual review; manual filtering is not reproducible from the paper alone.
  • ad hoc to paper The 44 rollout routes are representative enough that PPO training on them transfers to the remaining benchmark routes.
    Appendix C selects routes where action sampling succeeds; Tab. A2 shows weaker gains on non-rollout routes.
  • ad hoc to paper Eq. 3's factorization pi_c = pi_d * pi_g is a valid decomposition of trajectory selection.
    The paper assumes optimal trajectory selection can be modeled as a product of a language selection policy and a generation policy; this is architectural, not proven.
  • domain assumption The frozen Action Expert continues to generate feasible trajectories under the Decision Expert's shifted meta-action distribution after RL.
    Only the Decision Expert is updated by PPO; credit assignment assumes decisions map consistently to trajectories.
invented entities (1)
  • Discrete meta-action vocabulary (7 speed actions, 6 path actions) no independent evidence
    purpose: Serves as the action space for online RL; the Decision Expert chooses among them and the Action Expert maps them to trajectories.
    No external evidence that these 13 actions are the right granularity; validity is shown only through aggregate benchmark scores.

pith-pipeline@v1.3.0-alltime-deepseek · 18511 in / 15252 out tokens · 133524 ms · 2026-08-03T16:23:25.751872+00:00 · methodology

0 comments
read the original abstract

Current Vision-Language-Action (VLA) paradigms in autonomous driving primarily rely on Imitation Learning (IL), which introduces inherent challenges such as distribution shift and causal confusion. Online Reinforcement Learning offers a promising pathway to address these issues through trial-and-error learning. However, applying online reinforcement learning to VLA models in autonomous driving is hindered by inefficient exploration in continuous action spaces. To overcome this limitation, we propose MindDrive, a VLA framework comprising a large language model (LLM) with two distinct sets of LoRA parameters. The one LLM serves as a Decision Expert for scenario reasoning and driving decision-making, while the other acts as an Action Expert that dynamically maps linguistic decisions into feasible trajectories. By feeding trajectory-level rewards back into the reasoning space, MindDrive enables trial-and-error learning over a finite set of discrete linguistic driving decisions, instead of operating directly in a continuous action space. This approach effectively balances optimal decision-making in complex scenarios, human-like driving behavior, and efficient exploration in online reinforcement learning. Using the lightweight Qwen-0.5B LLM, MindDrive achieves Driving Score (DS) of 78.04 and Success Rate (SR) of 55.09% on the challenging Bench2Drive benchmark. To the best of our knowledge, this is the first work to demonstrate the effectiveness of online reinforcement learning for the VLA model in autonomous driving.

Figures

Figures reproduced from arXiv: 2512.13636 by Bing Wang, Diankun Zhang, Dingkang Liang, Guang Chen, Hangjun Ye, Haoyu Fu, Hongwei Xie, Jianfeng Cui, Xiang Bai, Zongchuang Zhao.

Figure 1
Figure 1. Figure 1: The comparison of different VLA for reinforcement [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of MindDrive and its training pipeline. MindDrive consists of two experts, both of which utilize the same base LLM [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Online reinforcement learning framework. The RL [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Ablation on the rollout rounds number. DS and SR [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Qualitative results of MindDrive after IL and RL on the Bench2Drive closed-loop evaluation set. The [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗

discussion (0)

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

Forward citations

Cited by 10 Pith papers

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

  1. nuReasoning: A Reasoning-Centric Dataset and Benchmark for Long-Tail Autonomous Driving

    cs.CV 2026-05 unverdicted novelty 6.0

    nuReasoning is a new real-world dataset and benchmark extending nuScenes/nuPlan with 20k clips and multi-type reasoning annotations to evaluate and improve reasoning in long-tail autonomous driving.

  2. HERMES++: Toward a Unified Driving World Model for 3D Scene Understanding and Generation

    cs.CV 2026-04 unverdicted novelty 6.0

    HERMES++ unifies 3D scene understanding and future geometry prediction in driving scenes via BEV representations, LLM-enhanced queries, a temporal link, and joint geometric optimization.

  3. Xiaomi OneVL: One-Step Latent Reasoning and Planning with Vision-Language Explanation

    cs.CV 2026-04 unverdicted novelty 6.0

    OneVL is the first latent CoT method to exceed explicit CoT accuracy on four driving benchmarks while running at answer-only speed, by supervising latent tokens with a visual world model decoder.

  4. Xiaomi OneVL: One-Step Latent Reasoning and Planning with Vision-Language Explanation

    cs.CV 2026-04 unverdicted novelty 6.0

    OneVL achieves superior accuracy to explicit chain-of-thought reasoning at answer-only latency by supervising latent tokens with a visual world model decoder that predicts future frames.

  5. Orion-Lite: Distilling LLM Reasoning into Efficient Vision-Only Driving Models

    cs.CV 2026-04 unverdicted novelty 6.0

    Orion-Lite uses latent feature distillation and trajectory supervision to create a vision-only model that surpasses its LLM-based teacher on closed-loop Bench2Drive evaluation, achieving a new SOTA driving score of 80.6.

  6. PointTPA: Dynamic Network Parameter Adaptation for 3D Scene Understanding

    cs.CV 2026-04 unverdicted novelty 6.0

    PointTPA uses serialization-based neighborhood grouping and a dynamic parameter projector to adapt network weights per scene patch, reaching 78.4% mIoU on ScanNet with under 2% added parameters.

  7. LVDrive: Latent Visual Representation Enhanced Vision-Language-Action Autonomous Driving Model

    cs.CV 2026-05 unverdicted novelty 5.0

    LVDrive improves closed-loop driving on Bench2Drive by adding latent future scene prediction to VLA models via unified embedding space processing and two-stage trajectory decoding.

  8. CRAFT: Counterfactual-to-Interactive Reinforcement Fine-Tuning for Driving Policies

    cs.LG 2026-05 unverdicted novelty 5.0

    CRAFT is an on-policy RL fine-tuning framework that decomposes closed-loop policy gradients into a group-normalized counterfactual proxy plus residual correction from interaction events, achieving top closed-loop perf...

  9. SpanVLA: Efficient Action Bridging and Learning from Negative-Recovery Samples for Vision-Language-Action Model

    cs.CV 2026-04 unverdicted novelty 5.0

    SpanVLA reduces action generation latency via flow-matching conditioned on history and improves robustness by training on negative-recovery samples with GRPO and a dedicated reasoning dataset.

  10. RAD-2: Scaling Reinforcement Learning in a Generator-Discriminator Framework

    cs.CV 2026-04 unverdicted novelty 5.0

    RAD-2 uses a diffusion generator and RL discriminator to cut collision rates by 56% in closed-loop autonomous driving planning.

Reference graph

Works this paper leans on

64 extracted references · 22 linked inside Pith · cited by 9 Pith papers

  1. [1]

    Gpt-4 technical report.arXiv preprint arXiv:2303.08774,

    Josh Achiam, Steven Adler, Sandhini Agarwal, Lama Ah- mad, Ilge Akkaya, Florencia Leoni Aleman, Diogo Almeida, Janko Altenschmidt, Sam Altman, Shyamal Anadkat, et al. Gpt-4 technical report.arXiv preprint arXiv:2303.08774,

    Pith/arXiv arXiv

  2. [2]

    Gemini: A family of highly capable multimodal models.arXiv preprint arXiv:2312.11805, 1, 2023

    Rohan Anil, Sebastian Borgeaud, Yonghui Wu, Jean- Baptiste Alayrac, Jiahui Yu, Radu Soricut, Johan Schalkwyk, Andrew M Dai, Anja Hauth, Katie Millican, et al. Gemini: A family of highly capable multimodal models.arXiv preprint arXiv:2312.11805, 1, 2023. 1

    Pith/arXiv arXiv 2023

  3. [3]

    Kevin Black, Noah Brown, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, Lachy Groom, Karol Hausman, Brian Ichter, et al.π 0: A vision-language- action flow model for general robot control.arXiv preprint arXiv:2410.24164, 2024. 3

    Pith/arXiv arXiv 2024

  4. [4]

    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. InProc. of IEEE Intl. Conf. on Computer Vision and Pattern Recognition, pages 11621–11631, 2020. 3

    2020

  5. [5]

    Rethinking imitation-based planners for autonomous driving

    Jie Cheng, Yingbing Chen, Xiaodong Mei, Bowen Yang, Bo Li, and Ming Liu. Rethinking imitation-based planners for autonomous driving. InProc. of the IEEE Int. Conf. on Robotics and Automation, pages 14123–14130, 2024. 2

    2024

  6. [6]

    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, et al. Navsim: Data-driven non-reactive autonomous vehicle simulation and benchmarking. InProc. of Advances in Neural Information Processing Systems, pages 28706–28719, 2024. 2

    2024

  7. [7]

    Causal confusion in imitation learning.Proc

    Pim De Haan, Dinesh Jayaraman, and Sergey Levine. Causal confusion in imitation learning.Proc. of Advances in Neural Information Processing Systems, 32, 2019. 2

    2019

  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. InPorc. of IEEE Intl. Conf. on Robot Learning, pages 1–16, 2017. 2, 5, 6, 9

    2017

  9. [9]

    Eva-02: A visual representation for neon genesis.Image and Vision Computing, 149:105171,

    Yuxin Fang, Quan Sun, Xinggang Wang, Tiejun Huang, Xin- long Wang, and Yue Cao. Eva-02: A visual representation for neon genesis.Image and Vision Computing, 149:105171,

  10. [10]

    Orion: A holistic end-to- end autonomous driving framework by vision-language in- structed action generation

    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 in- structed action generation. InPorc. of IEEE Intl. Conf. on Computer Vision, 2025. 1, 2, 3, 5, 6, 7

    2025

  11. [11]

    Rad: Training an end-to-end driv- ing policy via large-scale 3dgs-based reinforcement learning

    Hao Gao, Shaoyu Chen, Bo Jiang, Bencheng Liao, Yiang Shi, Xiaoyang Guo, Yuechuan Pu, Haoran Yin, Xiangyu Li, Xinbang Zhang, et al. Rad: Training an end-to-end driv- ing policy via large-scale 3dgs-based reinforcement learning. InProc. of Advances in Neural Information Processing Sys- tems, 2025. 3

    2025

  12. [12]

    Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025

    Daya Guo, Dejian Yang, Haowei Zhang, Junxiao Song, Ruoyu Zhang, Runxin Xu, Qihao Zhu, Shirong Ma, Peiyi Wang, Xiao Bi, et al. Deepseek-r1: Incentivizing reasoning capability in llms via reinforcement learning.arXiv preprint arXiv:2501.12948, 2025. 2, 3

    Pith/arXiv arXiv 2025

  13. [13]

    Lora: Low-rank adaptation of large language models

    Edward J Hu, Phillip Wallis, Zeyuan Allen-Zhu, Yuanzhi Li, Shean Wang, Lu Wang, Weizhu Chen, et al. Lora: Low-rank adaptation of large language models. InProc. of Intl. Conf. on Learning Representations, 2021. 2, 3, 6

    2021

  14. [14]

    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. InProc. of IEEE Intl. Conf. on Computer Vision and Pattern Recog- nition, pages 17853–17862, 2023. 1, 2, 7

    2023

  15. [15]

    Robotron- drive: All-in-one large multimodal model for autonomous driving

    Zhijian Huang, Chengjian Feng, Feng Yan, Baihui Xiao, Ze- qun Jie, Yujie Zhong, Xiaodan Liang, and Lin Ma. Robotron- drive: All-in-one large multimodal model for autonomous driving. InPorc. of IEEE Intl. Conf. on Computer Vision, pages 8011–8021, 2025. 3

    2025

  16. [16]

    Zilin Huang, Zihao Sheng, Yansong Qu, Junwei You, and Sikai Chen. Vlm-rl: A unified vision language models and reinforcement learning framework for safe autonomous driv- ing.Transportation Research Part C: Emerging Technolo- gies, 180:105321, 2025. 2

    2025

  17. [17]

    Emma: End-to-end multimodal model for autonomous driving.arXiv preprint arXiv:2410.23262,

    Jyh-Jing Hwang, Runsheng Xu, Hubert Lin, Wei-Chih Hung, Jingwei Ji, Kristy Choi, Di Huang, Tong He, Paul Covington, Benjamin Sapp, et al. Emma: End-to-end multimodal model for autonomous driving.arXiv preprint arXiv:2410.23262,

    Pith/arXiv arXiv

  18. [18]

    Physical Intelligence, Kevin Black, Noah Brown, James Darpinian, Karan Dhabalia, Danny Driess, Adnan Esmail, Michael Equi, Chelsea Finn, Niccolo Fusai, et al.π 0.5: a vision-language-action model with open-world generaliza- tion.arXiv preprint arXiv:2504.16054, 2025. 3

    Pith/arXiv arXiv 2025

  19. [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. InPorc. of IEEE Intl. Conf. on Computer Vision, 2023. 7

    2023

  20. [20]

    Think twice before driv- ing: Towards scalable decoders for end-to-end autonomous 9 driving

    Xiaosong Jia, Penghao Wu, Li Chen, Jiangwei Xie, Conghui He, Junchi Yan, and Hongyang Li. Think twice before driv- ing: Towards scalable decoders for end-to-end autonomous 9 driving. InProc. of IEEE Intl. Conf. on Computer Vision and Pattern Recognition, 2023. 7

    2023

  21. [21]

    Amp: Autoregressive mo- tion prediction revisited with next token prediction for au- tonomous driving.arXiv preprint arXiv:2403.13331, 2024

    Xiaosong Jia, Shaoshuai Shi, Zijun Chen, Li Jiang, Wenlong Liao, Tao He, and Junchi Yan. Amp: Autoregressive mo- tion prediction revisited with next token prediction for au- tonomous driving.arXiv preprint arXiv:2403.13331, 2024. 2

    Pith/arXiv arXiv 2024

  22. [22]

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

    Xiaosong Jia, Zhenjie Yang, Qifeng Li, Zhiyuan Zhang, and Junchi Yan. Bench2drive: Towards multi-ability bench- marking of closed-loop end-to-end autonomous driving. In Proc. of Advances in Neural Information Processing Sys- tems, 2024. 2, 3, 6, 1

    2024

  23. [23]

    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. InProc. of Intl. Conf. on Learning Representations, 2025. 3, 7

    2025

  24. [24]

    Irl-vla: Training an vision-language- action policy via reward world model.arXiv preprint arXiv:2508.06571, 2025

    Anqing Jiang, Yu Gao, Yiru Wang, Zhigang Sun, Shuo Wang, Yuwen Heng, Hao Sun, Shichen Tang, Lijuan Zhu, Jinhao Chai, et al. Irl-vla: Training an vision-language- action policy via reward world model.arXiv preprint arXiv:2508.06571, 2025. 2

    Pith/arXiv arXiv 2025

  25. [25]

    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. InPorc. of IEEE Intl. Conf. on Computer Vision, pages 8340–8350, 2023. 1, 2, 7

    2023

  26. [26]

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

    Bo Jiang, Shaoyu Chen, Qian Zhang, Wenyu Liu, and Xing- gang Wang. Alphadrive: Unleashing the power of vlms in autonomous driving via reinforcement learning and reason- ing.arXiv preprint arXiv:2503.07608, 2025. 2, 3

    Pith/arXiv arXiv 2025

  27. [27]

    Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024

    Moo Jin Kim, Karl Pertsch, Siddharth Karamcheti, Ted Xiao, Ashwin Balakrishna, Suraj Nair, Rafael Rafailov, Ethan Foster, Grace Lam, Pannag Sanketi, et al. Openvla: An open-source vision-language-action model.arXiv preprint arXiv:2406.09246, 2024. 3

    Pith/arXiv arXiv 2024

  28. [28]

    Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114, 2013

    Diederik P Kingma. Auto-encoding variational bayes.arXiv preprint arXiv:1312.6114, 2013. 5

    Pith/arXiv arXiv 2013

  29. [29]

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

    Yingyan Li, Yuqi Wang, Yang Liu, Jiawei He, Lue Fan, and Zhaoxiang Zhang. End-to-end driving with online trajectory evaluation via bev world model. InPorc. of IEEE Intl. Conf. on Computer Vision, 2025. 7

    2025

  30. [30]

    Recogdrive: A reinforced cognitive frame- work for end-to-end autonomous driving.arXiv preprint arXiv:2506.08052, 2025

    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 frame- work for end-to-end autonomous driving.arXiv preprint arXiv:2506.08052, 2025. 2, 3, 6, 7

    Pith/arXiv arXiv 2025

  31. [31]

    A comprehensive review of multi-agent reinforcement learning in video games.IEEE Transactions on Games, 2025

    Zhengyang Li, Qijin Ji, Xinghong Ling, and Quan Liu. A comprehensive review of multi-agent reinforcement learning in video games.IEEE Transactions on Games, 2025. 3

    2025

  32. [32]

    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. InProc. of IEEE Intl. Conf. on Computer Vision and Pattern Recognition, pages 12037–12047, 2025. 3

    2025

  33. [33]

    Reinforced refinement with self-aware ex- pansion for end-to-end autonomous driving.arXiv preprint arXiv:2506.09800, 2025

    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 ex- pansion for end-to-end autonomous driving.arXiv preprint arXiv:2506.09800, 2025. 3

    Pith/arXiv arXiv 2025

  34. [34]

    Laformer: Trajectory prediction for autonomous driving with lane-aware scene constraints

    Mengmeng Liu, Hao Cheng, Lin Chen, Hellward Broszio, Jiangtao Li, Runjiang Zhao, Monika Sester, and Michael Ying Yang. Laformer: Trajectory prediction for autonomous driving with lane-aware scene constraints. InProc. of IEEE Intl. Conf. on Computer Vision and Pattern Recognition, pages 2039–2049, 2024. 2

    2039

  35. [35]

    Pillarnest: Embracing backbone scaling and pretraining for pillar-based 3d object detection.IEEE Trans- actions on Intelligent Vehicles, 2024

    Weixin Mao, Tiancai Wang, Diankun Zhang, Junjie Yan, and Osamu Yoshie. Pillarnest: Embracing backbone scaling and pretraining for pillar-based 3d object detection.IEEE Trans- actions on Intelligent Vehicles, 2024. 2

    2024

  36. [36]

    Training language models to follow instructions with human feedback.Proc

    Long Ouyang, Jeffrey Wu, Xu Jiang, Diogo Almeida, Car- roll Wainwright, Pamela Mishkin, Chong Zhang, Sandhini Agarwal, Katarina Slama, Alex Ray, et al. Training language models to follow instructions with human feedback.Proc. of Advances in Neural Information Processing Systems, 35: 27730–27744, 2022. 2

    2022

  37. [37]

    Agile autonomous driving using end-to-end deep imitation learning.Robotics: Science and Systems, 2018

    Yunpeng Pan, Ching-An Cheng, Kamil Saigol, Keuntaek Lee, Xinyan Yan, Evangelos Theodorou, and Byron Boots. Agile autonomous driving using end-to-end deep imitation learning.Robotics: Science and Systems, 2018. 2

    2018

  38. [38]

    Markov decision processes.Handbooks in operations research and management science, 2:331–434,

    Martin L Puterman. Markov decision processes.Handbooks in operations research and management science, 2:331–434,

  39. [39]

    Direct preference optimization: Your language model is secretly a reward model.Proc

    Rafael Rafailov, Archit Sharma, Eric Mitchell, Christo- pher D Manning, Stefano Ermon, and Chelsea Finn. Direct preference optimization: Your language model is secretly a reward model.Proc. of Advances in Neural Information Pro- cessing Systems, 36:53728–53741, 2023. 2

    2023

  40. [40]

    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. InProc. of IEEE Intl. Conf. on Computer Vision and Pattern Recognition, pages 11993– 12003, 2025. 1, 4

    2025

  41. [41]

    A re- duction of imitation learning and structured prediction to no- regret online learning

    St ´ephane Ross, Geoffrey Gordon, and Drew Bagnell. A re- duction of imitation learning and structured prediction to no- regret online learning. InProc. of Intl. Conf. artificial intel- ligence and statistics, pages 627–635, 2011. 2

    2011

  42. [42]

    Proximal policy optimization algo- rithms.arXiv preprint arXiv:1707.06347, 2017

    John Schulman, Filip Wolski, Prafulla Dhariwal, Alec Rad- ford, and Oleg Klimov. Proximal policy optimization algo- rithms.arXiv preprint arXiv:1707.06347, 2017. 5, 2

    Pith/arXiv arXiv 2017

  43. [43]

    Drivedpo: Policy learning via safety dpo for end-to-end autonomous driving

    Shuyao Shang, Yuntao Chen, Yuqi Wang, Yingyan Li, and Zhaoxiang Zhang. Drivedpo: Policy learning via safety dpo for end-to-end autonomous driving. InProc. of Advances in Neural Information Processing Systems, 2025. 7

    2025

  44. [44]

    Deepseekmath: Pushing the limits of math- ematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024

    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 math- ematical reasoning in open language models.arXiv preprint arXiv:2402.03300, 2024. 9

    Pith/arXiv arXiv 2024

  45. [45]

    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. InProc. of IEEE Intl. Conf. on Computer Vision and Pattern Recog- nition, 2025. 3, 7 10

    2025

  46. [46]

    Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024

    Peng Wang, Shuai Bai, Sinan Tan, Shijie Wang, Zhihao Fan, Jinze Bai, Keqin Chen, Xuejing Liu, Jialin Wang, Wenbin Ge, et al. Qwen2-vl: Enhancing vision-language model’s perception of the world at any resolution.arXiv preprint arXiv:2409.12191, 2024. 1

    Pith/arXiv arXiv 2024

  47. [47]

    Omnidrive: A holistic vision-language dataset for au- tonomous driving with counterfactual reasoning

    Shihao Wang, Zhiding Yu, Xiaohui Jiang, Shiyi Lan, Min Shi, Nadine Chang, Jan Kautz, Ying Li, and Jose M Al- varez. Omnidrive: A holistic vision-language dataset for au- tonomous driving with counterfactual reasoning. InProc. of IEEE Intl. Conf. on Computer Vision and Pattern Recogni- tion, pages 22442–22452, 2025. 3

    2025

  48. [48]

    Diffad: A unified diffusion mod- eling approach for autonomous driving.arXiv preprint arXiv:2503.12170, 2025

    Tao Wang, Cong Zhang, Xingguang Qu, Kun Li, Weiwei Liu, and Chang Huang. Diffad: A unified diffusion mod- eling approach for autonomous driving.arXiv preprint arXiv:2503.12170, 2025. 3, 6, 7

    Pith/arXiv arXiv 2025

  49. [49]

    Chain-of-thought prompting elicits reasoning in large lan- guage models.Proc

    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 lan- guage models.Proc. of Advances in Neural Information Pro- cessing Systems, 35:24824–24837, 2022. 3

    2022

  50. [50]

    Openemma: Open-source multimodal model for end-to-end autonomous driving

    Shuo Xing, Chengyuan Qian, Yuping Wang, Hongyuan Hua, Kexin Tian, Yang Zhou, and Zhengzhong Tu. Openemma: Open-source multimodal model for end-to-end autonomous driving. InProc. of IEEE Winter Conf. on Applications of Computer Vision, pages 1001–1009, 2025. 3

    2025

  51. [51]

    Drivegpt4: Interpretable end-to-end autonomous driving via large language model.IEEE Robotics and Automation Let- ters, 2024

    Zhenhua Xu, Yujia Zhang, Enze Xie, Zhen Zhao, Yong Guo, Kwan-Yee K Wong, Zhenguo Li, and Hengshuang Zhao. Drivegpt4: Interpretable end-to-end autonomous driving via large language model.IEEE Robotics and Automation Let- ters, 2024. 1, 3

    2024

  52. [52]

    Drivegpt4-v2: Harnessing large language model capabilities for enhanced closed-loop autonomous driving

    Zhenhua Xu, Yan Bai, Yujia Zhang, Zhuoling Li, Fei Xia, Kwan-Yee K Wong, Jianqiang Wang, and Hengshuang Zhao. Drivegpt4-v2: Harnessing large language model capabilities for enhanced closed-loop autonomous driving. InProc. of IEEE Intl. Conf. on Computer Vision and Pattern Recogni- tion, pages 17261–17270, 2025. 1, 4

    2025

  53. [53]

    Qwen2 technical report.arXiv preprint arXiv:2407.10671, 2024

    An Yang, Baosong Yang, Binyuan Hui, Bo Zheng, Bowen Yu, Chang Zhou, Chengpeng Li, Chengyuan Li, Dayi- heng Liu, et al. Qwen2 technical report.arXiv preprint arXiv:2407.10671, 2024. 6

    Pith/arXiv arXiv 2024

  54. [54]

    Drivemoe: Mixture-of-experts for vision-language-action model in end-to-end autonomous driving.arXiv preprint arXiv:2505.16278, 2025

    Zhenjie Yang, Yilin Chai, Xiaosong Jia, Qifeng Li, Yuqian Shao, Xuekai Zhu, Haisheng Su, and Junchi Yan. Drivemoe: Mixture-of-experts for vision-language-action model in end-to-end autonomous driving.arXiv preprint arXiv:2505.16278, 2025. 6, 7

    Pith/arXiv arXiv 2025

  55. [55]

    Raw2drive: Reinforcement learning with aligned world models for end-to-end autonomous driv- ing (in carla v2)

    Zhenjie Yang, Xiaosong Jia, Qifeng Li, Xue Yang, Maoqing Yao, and Junchi Yan. Raw2drive: Reinforcement learning with aligned world models for end-to-end autonomous driv- ing (in carla v2). InProc. of Advances in Neural Information Processing Systems, 2025. 3, 4, 6, 7

    2025

  56. [56]

    Autodrive-r 2: Incentivizing reason- ing and self-reflection capacity for vla model in autonomous driving.arXiv preprint arXiv:2509.01944, 2025

    Zhenlong Yuan, Jing Tang, Jinguo Luo, Rui Chen, Chengx- uan Qian, Lei Sun, Xiangxiang Chu, Yujun Cai, Dapeng Zhang, and Shuo Li. Autodrive-r 2: Incentivizing reason- ing and self-reflection capacity for vla model in autonomous driving.arXiv preprint arXiv:2509.01944, 2025. 2

    Pith/arXiv arXiv 2025

  57. [57]

    Fully sparse transformer 3-d detector for lidar point cloud.IEEE Transactions on Geoscience and Re- mote Sensing, 61:1–12, 2023

    Diankun Zhang, Zhijie Zheng, Haoyu Niu, Xueqing Wang, and Xiaojun Liu. Fully sparse transformer 3-d detector for lidar point cloud.IEEE Transactions on Geoscience and Re- mote Sensing, 61:1–12, 2023. 2

    2023

  58. [58]

    Sparsead: Sparse query-centric paradigm for efficient end-to-end autonomous driving.arXiv preprint arXiv:2404.06892, 2024

    Diankun Zhang, Guoan Wang, Runwen Zhu, Jianbo Zhao, Xiwu Chen, Siyu Zhang, Jiahao Gong, Qibin Zhou, Wenyuan Zhang, Ningzi Wang, et al. Sparsead: Sparse query-centric paradigm for efficient end-to-end autonomous driving.arXiv preprint arXiv:2404.06892, 2024. 1, 3

    Pith/arXiv arXiv 2024

  59. [59]

    Carplanner: Consistent auto-regressive trajectory planning for large-scale reinforcement learning in autonomous driving

    Dongkun Zhang, Jiaming Liang, Ke Guo, Sha Lu, Qi Wang, Rong Xiong, Zhenwei Miao, and Yue Wang. Carplanner: Consistent auto-regressive trajectory planning for large-scale reinforcement learning in autonomous driving. InProc. of IEEE Intl. Conf. on Computer Vision and Pattern Recogni- tion, pages 17239–17248, 2025. 3, 4

    2025

  60. [60]

    Extending large vision-language model for diverse interactive tasks in autonomous driving.arXiv preprint arXiv:2505.08725, 2025

    Zongchuang Zhao, Haoyu Fu, Dingkang Liang, Xin Zhou, Dingyuan Zhang, Hongwei Xie, Bing Wang, and Xiang Bai. Extending large vision-language model for diverse interactive tasks in autonomous driving.arXiv preprint arXiv:2505.08725, 2025. 3

    Pith/arXiv arXiv 2025

  61. [61]

    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. InProc. of European Conference on Com- puter Vision, pages 87–104, 2024. 3, 7

    2024

  62. [62]

    Hermes: A unified self-driving world model for simultaneous 3d scene understanding and generation

    Xin Zhou, Dingkang Liang, Sifan Tu, Xiwu Chen, Yikang Ding, Dingyuan Zhang, Feiyang Tan, Hengshuang Zhao, and Xiang Bai. Hermes: A unified self-driving world model for simultaneous 3d scene understanding and generation. In Porc. of IEEE Intl. Conf. on Computer Vision, 2025. 3

    2025

  63. [63]

    Embodied understanding of driving scenarios

    Yunsong Zhou, Linyan Huang, Qingwen Bu, Jia Zeng, Tianyu Li, Hang Qiu, Hongzi Zhu, Minyi Guo, Yu Qiao, and Hongyang Li. Embodied understanding of driving scenarios. InProc. of European Conference on Computer Vision, pages 129–148, 2024. 3

    2024

  64. [64]

    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. Autovla: A vision- language-action model for end-to-end autonomous driving with adaptive reasoning and reinforcement fine-tuning. In Proc. of Advances in Neural Information Processing Sys- tems, 2025. 2, 3 11 MindDrive: A Vision-Language-Action Model for Autonomous Dri...

    2025