Pith. sign in

REVIEW 5 major objections 6 minor 13 cited by

Autonomous driving can be learned end-to-end from images and rule-based rewards alone, with no human driving demonstrations.

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

2026-08-04 07:48 UTC pith:SSPO5RS5

load-bearing objection Reward-only trajectory scoring is a real idea, but the headline SOTA and 'zero-human' claim need closer scrutiny. the 5 major comments →

arxiv 2510.24108 v2 pith:SSPO5RS5 submitted 2025-10-28 cs.RO cs.CV

Zero-Human Demonstration End-to-end Autonomous Driving with Trajectory Scorer

Zhenxin Li , Nadine Chang , Wenhao Yao , Xinglong Sun , Zi Wang , Maying Shen , Jingde Chen , Jingyu Song
show 4 more authors
Kailin Li Zuxuan Wu Shiyi Lan Jose M. Alvarez
This is my paper
classification cs.RO cs.CV
keywords end-to-end autonomous drivingoffline reinforcement learningtrajectory scoringpolicy gradientimitation learningrule-based rewardsExhaustive Policy Optimizationlong-tail driving
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.

This paper claims that an end-to-end autonomous driving planner can be trained from scratch using only real-world camera images and rule-based rewards, without any human demonstration trajectories as training labels. The proposed framework, ZTRS, treats planning as scoring a fixed vocabulary of candidate trajectories and uses a policy-gradient variant, Exhaustive Policy Optimization, that evaluates every trajectory in the vocabulary at once, giving dense supervision that overcomes the cold-start problem of reinforcement learning. The authors report that ZTRS reaches a new state-of-the-art on the challenging open-loop planning benchmark Navhard and outperforms imitation-learning baselines on the closed-loop HUGSIM benchmark, while never training on simulated data. If correct, this would show that scarce and sometimes suboptimal human demonstration data is not a prerequisite for robust end-to-end driving.

Core claim

The central discovery is that eliminating imitation learning entirely from end-to-end driving training is possible: ZTRS learns a policy purely from rule-based rewards while operating directly on high-dimensional real-world images. This is achieved by discretizing the action space into a fixed set of 16,384 trajectories (obtained by K-means clustering of human driving logs), and then using Exhaustive Policy Optimization (EPO), a policy gradient variant that sums the gradient over every action in the set instead of sampling one, providing dense supervision at every offline state. With the rule-based EPDMS reward and a temporal-consistency correction term, the policy learns driving patterns th

What carries the argument

The central mechanism is the combination of a finite trajectory vocabulary and exhaustive policy gradient. The vocabulary A is a set of 16,384 trajectories spanning 4 seconds at 10Hz, clustered from the nuPlan dataset, which discretizes continuous driving into enumerable actions so rewards can be computed offline for every action. Exhaustive Policy Optimization (EPO) then computes the policy gradient as a sum over all actions of the rule-based reward times the gradient of the action probability, rather than the log-likelihood of a single sampled action, yielding much denser supervision. The reward is the Extended Predictive Driver Model Score (EPDMS), a rule-based metric combining multiplica

Load-bearing premise

The load-bearing premise is that the fixed set of 16,384 trajectories obtained by clustering human driving logs covers 'almost all driving possibilities'; if a necessary evasive maneuver is missing from this vocabulary, the policy cannot express it, no matter how strong the perception or optimization is.

What would settle it

Run ZTRS in a closed-loop scenario that demands a maneuver absent from the nuPlan-derived vocabulary, such as a sharp reverse-direction swerve into a gap or a U-turn in a confined space. If the planner fails because the best available action in its vocabulary is still unsafe, while a planner with a larger or differently derived vocabulary succeeds, the central claim of human-demonstration-free long-tail generalization is falsified. Concretely, compare the highest-reward trajectory in the failure scenario: if it lies outside the 16,384-set, the failure is purely vocabulary-bound.

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

If this is right

  • End-to-end driving policies can be trained from real-world images alone, using rule-based rewards, removing the need for human demonstration labels in training.
  • Dense supervision over an enumerable action space solves the cold-start problem that makes RL impractical for continuous trajectory generation from high-dimensional inputs.
  • The resulting policy transfers zero-shot to closed-loop 3D-Gaussian-Splatting simulated environments (HUGSIM), outperforming imitation-learning trajectory scorers without any simulated-data fine-tuning.
  • On the long-tail open-loop benchmark Navhard, ZTRS sets a new state-of-the-art EPDMS, suggesting rule-based rewards are sufficient for safety-critical planning scenarios.

Where Pith is reading between the lines

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

  • The 'zero-human demonstration' claim is narrower than it appears: the trajectory vocabulary itself is derived from human driving logs (K-means clustering on nuPlan), so the policy cannot output any maneuver that is not represented in human data; a fully human-free pipeline would need a vocabulary generated by non-human means such as kinematic sampling or rule-based generation.
  • EPO's exhaustive-sum gradient is a general recipe for any discrete-action control problem with an offline-computable reward; it could extend beyond driving to robot manipulation or navigation tasks where demonstrations are scarce but rule-based success criteria exist.
  • Because EPDMS is itself a human-designed metric with specific weights and comfort thresholds, the pipeline still embeds human knowledge; the contribution is precisely the removal of trajectory labels, not the removal of all human priors.
  • A testable extension: retrain ZTRS with a vocabulary produced by a vehicle dynamics model or random sampling instead of nuPlan clustering. If performance on Navhard and HUGSIM stays comparable, then the vocabulary is not the bottleneck for long-tail generalization; if it drops, the ceiling of the method is set by the human-data-derived action set.

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

5 major / 6 minor

Summary. The paper proposes ZTRS, an end-to-end trajectory-scoring planner for autonomous driving trained without imitation-learning labels. Instead of regressing a continuous trajectory, the model scores a fixed set of 16,384 candidate trajectories produced by K-means clustering on nuPlan. The policy head is optimized by "Exhaustive Policy Optimization" (EPO), which sums the gradient over all actions weighted by a rule-based EPDMS reward minus a temporal-consistency penalty. Experiments on Navtest, Navhard, and HUGSIM are reported; the headline claims are state-of-the-art Navhard EPDMS and zero-shot HUGSIM performance above IL baselines.

Significance. If the central claims held, ZTRS would be a meaningful step: it would show that a sensor-based end-to-end planner can be trained with rule-based rewards alone, avoiding covariate shift and demonstration scarcity. The EPO formulation is simple and the ablation in Table 4 supports the value of dense likelihood updates and the temporal penalty. The paper also evaluates on three benchmarks and commits to code release. However, the significance is reduced by the human-derived action vocabulary and pretrained backbones, and the headline Navhard result is obtained under a benchmark-specific inference setting not disclosed in Table 1. These issues do not invalidate the method, but they require the claims to be restated and the experiments to be reported more transparently.

major comments (5)
  1. [Sec. 3.2 / Sec. 2.3] The central "zero-human demonstration"/"eliminates imitation learning" claim is overstated. The action space A is obtained by K-means clustering on the nuPlan human driving logs (Sec. 3.2), so the set of candidate trajectories is derived from human demonstrations; the policy can only select among these human-derived trajectories. Additionally, the V2-99 and ViT-L backbones are pretrained (DD3D, Depth-Anything). Thus the statement in the abstract that training is "solely on real-world images and rule-based rewards, entirely without human demonstration" is not accurate unless "demonstration" is restricted to trajectory labels. Please rephrase the claim and discuss the ancestry of the vocabulary.
  2. [Table 1 vs. Table 5, Sec. 3.2] The Navhard SOTA claim is not robust. Table 5 shows that with full 16,384-action inference, V2-99 obtains Navhard EPDMS 43.4; the reported 45.5 in Table 1 is obtained by shrinking the inference action space to 8,192. This configuration is not disclosed in Table 1 and contradicts the stated default of 16,384 actions. The margin over GTRS-Dense ViT-L (45.3) is 0.2 points, with no error bars or seeds. The result therefore appears to be a benchmark-specific selection artifact. Please report both inference configurations in Table 1, provide variance across seeds, and either use a pre-registered/held-out selection rule or temper the SOTA claim.
  3. [Eq. (6)] Equation (6) as written, E_{s∼D,a∼A}[Ψ(s,a)], does not contain the policy and is not the objective being optimized. The correct objective for EPO is E_{s∼D}[Σ_{a∈A} π_θ(a|s) Ψ(s,a)], or equivalently E_{s∼D, a∼π_θ}[Ψ(s,a)]. The gradient in Eq. (7) is correct for that objective, but the derivation leading to it is misstated. Please fix Eq. (6) and explain that EPO maximizes expected reward under the policy, not an expectation over actions drawn uniformly from A.
  4. [Sec. 2.3 / Sec. 3.2] The assumption that the 16,384-trajectory vocabulary "cover[s] almost all driving possibilities" is unsupported. No coverage analysis is provided, and because the vocabulary is a finite clustering of nuPlan logs, any maneuver absent from the clusters is inexpressible by the policy. This directly bounds the claimed long-tail generalization. Please quantify vocabulary coverage (e.g., distance from held-out human trajectories or success rate on a maneuver taxonomy) or soften the claim and add this as a limitation.
  5. [Table 3 / Abstract] The HUGSIM claim is overstated as stated. The table mixes official results on private scenarios (marked *) with runs on public scenarios, so rows are not directly comparable. Against the directly comparable public-scenario GTRS-Dense, ZTRS wins overall HD-Score by 0.3 points but loses on the medium and extreme difficulty categories. Please restrict the comparison to the same scenario subset, report per-category significance, and qualify the claim accordingly.
minor comments (6)
  1. [Sec. 2.1] The final trajectory is chosen by a weighted average of π and the m scoring-head outputs, but the weights are never given. Please provide the aggregation formula and any temperature/scaling.
  2. [Eq. (8)] The notation λ1[EC(a_{t-1}, a_t)] is unclear. Is 1 an indicator and EC a boolean violation? Also a_{t-1} is the argmax of the current policy, so the reward is non-stationary during training; please state how this is computed and whether gradients flow through b.
  3. [Abstract / Introduction] Use consistent terminology — "zero-human demonstration", "zero-imitation", and "eliminates imitation learning entirely" are claimed interchangeably; the actual scope is no imitation training labels.
  4. [Table 3] Add a column or footnote specifying the number of scenarios and the subset (public/private) for each method, and avoid placing official and public runs in the same group without a clear separation.
  5. [Tables 1-5] Report the number of random seeds/independent runs for the main tables; the Navhard margin is small and no variance is given.
  6. [Sec. 3.2] Please provide the K-means clustering details (number of nuPlan trajectories, feature representation, distance metric) for reproducibility.

Circularity Check

0 steps flagged

No significant circularity: training and evaluation share the EPDMS metric family, but benchmark scores are not forced by construction; remaining concerns are benchmark alignment, mild self-citation, and an inference-configuration reporting issue.

full rationale

ZTRS's derivation chain is not circular. The training reward is Ψ(s,a)=E(s,a)−b(s,a,a_{t−1}) (Eq. 8), where E is the EPDMS score computed on offline Navtrain states; the reported benchmark numbers are EPDMS computed on held-out Navtest/Navhard/HUGSIM evaluations. Having the training reward and evaluation metric belong to the same metric family is benchmark alignment, not a reduction of the result to its inputs. The paper's own Table 4 supplies a direct non-circularity check: training on the argmax-EPDMS trajectory as an imitation label yields only 76.7 EPDMS on Navtest, while the proposed EPO reward-trained policy reaches 85.3, so high evaluation scores are not statistically forced by the reward construction. The main self-citations (the EC term from Li et al. 2025b entering Eq. 9 and the temporal-consistency correction b) are incremental prior-work components used inside an externally defined benchmark metric; they do not by themselves define the reported outcome and are available to all baselines. The 'zero-human demonstration' claim is imprecise because the 16,384-action vocabulary is K-means clustered from nuPlan human driving logs, but this is a limitation or assumption about the action prior, not a circular derivation of the planning result. Finally, the Navhard SOTA in Table 1 corresponds to the 8,192-inference-action configuration in Table 5, whereas Sec. 3.2 states the default action space is 16,384; this is an undisclosed inference-time configuration choice (a reporting/correctness concern), not circularity.

Axiom & Free-Parameter Ledger

3 free parameters · 4 axioms · 0 invented entities

The central claim rests on the trajectory vocabulary covering all relevant maneuvers, on the sufficiency of per-state open-loop EPDMS as a reward, and on ignoring sequential credit assignment. These are domain assumptions, not derived results. No new physical entities are introduced.

free parameters (3)
  • lambda (temporal consistency weight) = 0.2
    In Eq. 8, lambda scales the Extended Comfort penalty term b(st, at, at-1); set by hand in Sec. 3.2.
  • |A| training action-space size = 16384
    Number of K-means trajectory clusters on nuPlan; chosen by hand, defines the policy's output space.
  • |A| inference action-space size = 8192
    Final Navhard results use 8192 trajectories at inference after training with 16384 (Table 5), a post-hoc selection.
axioms (4)
  • domain assumption A 'covers almost all driving possibilities'
    Sec. 2.3: 'When the action space A is a set of trajectories covering almost all driving possibilities, policy optimization can be formulated...' This is necessary for the EPO objective to be sound.
  • domain assumption One-step bandit / no temporal credit assignment
    Eqs. 6-8 optimize per-state expected EPDMS without a value function or environment interaction; assumes immediate reward captures all that matters for the trajectory.
  • domain assumption EPDMS open-loop reward transfers to closed-loop performance
    Training reward is EPDMS (open-loop non-reactive metrics), while HUGSIM evaluates closed-loop; assumes optimizing open-loop scores yields safe closed-loop behavior.
  • standard math Policy gradient theorem / expectation expansion
    Eqs. 1-5 use the standard identity for policy gradients; accepted background.

pith-pipeline@v1.3.0-alltime-deepseek · 13690 in / 12223 out tokens · 80608 ms · 2026-08-04T07:48:53.105148+00:00 · methodology

0 comments
read the original abstract

Human demonstrations are widely considered the cornerstone of end-to-end (E2E) autonomous driving despite human demonstration's scarcity for long-tail and safety-critical scenarios. Nonetheless, current E2E autonomous driving (AD) training paradigms continue to rely on human demonstrations. Imitation learning (IL) requires human demonstrations for training, whereas reinforcement learning (RL) has emerged as a promising alternative to reduce this dependency. However, most existing RL methods for E2E AD still rely implicitly on human demonstrations. A pure rewards-based RL method can overcome the need for human demonstrations, but general RL policy gradient methods suffer from the cold-start problem. In this paper, we propose ZTRS (Zero-human demonstration end-to-end autonomous driving with TRajectory Scorer) - a complete RL-based E2E planning paradigm trained solely on real-world images and rule-based rewards, entirely without human demonstration. Through our proposed Exhaustive Policy Optimization (EPO), a policy gradient variant tailored for enumerable trajectory actions and dense supervision, ZTRS enables the model to generalize better to long-tail driving scenarios. We demonstrate this generalization through our SOTA performance against IL approaches on both long-tail Navhard and closed-loop HUGSIM datasets. Project page: https://zhenxinli.net/ZTRS/.

Figures

Figures reproduced from arXiv: 2510.24108 by Jingde Chen, Jingyu Song, Jose M. Alvarez, Kailin Li, Maying Shen, Nadine Chang, Shiyi Lan, Wenhao Yao, Xinglong Sun, Zhenxin Li, Zi Wang, Zuxuan Wu.

Figure 1
Figure 1. Figure 1: Comparisons between three paradigms for end-to-end autonomous driving. ZTRS builds on three pillars: data, rewards, and policy optimization. For the data pillar, since large￾scale sensor data collection is difficult, we rely on offline driving datasets. This naturally transforms our problem into offline reinforcement learning (Levine et al., 2020), where the planner learns to maximize the reward at each da… view at source ↗
Figure 2
Figure 2. Figure 2: The Overall Framework of ZTRS. Given offline sensor data and a fixed set of trajec￾tories, ZTRS first tokenizes these two modalities. In a Transformer Decoder, the trajectory tokens attend to image tokens to acquire the context. Finally, scoring heads and a policy head map the trajectory tokens to rule-based scores and action likelihoods. 2 METHODOLOGY In this section, we elaborate on the framework and the… view at source ↗
Figure 3
Figure 3. Figure 3: Visualizations of planned trajectories (blue curves) and the human trajectory (green curves) on the open-loop planning benchmark Navtest. (c) Overtaking with oncoming traffic (b) Overtaking a slow vehicle (a) Collision avoidance with rogue vehicles [PITH_FULL_IMAGE:figures/full_fig_p008_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Visualizations of planned trajectories (orange dots) on the challenging closed-loop driving benchmark HUGSIM. CARLA (Dosovitskiy et al., 2017). Subsequent works improve the closed-loop driving performance with modern neural architectures (e.g. Transformers (Vaswani et al., 2017)) (Chitta et al., 2022), intermediate representations (Hu et al., 2022; Renz et al., 2022; Shao et al., 2023; Jia et al., 2023b), … 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 13 Pith papers

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

  1. Test-Time Coverage: Test-Conditioned Data Curation for Deployment-Aware Learning

    cs.AI 2026-07 conditional novelty 6.0

    TTCov curates training data for deployment by building an LLM-generated atomic-proposition atlas of the test distribution and greedily selecting clips that match it.

  2. Scaling Self-Play for End-to-End Driving

    cs.RO 2026-06 unverdicted novelty 6.0

    Self-play DAgger training in a batched pixel renderer produces end-to-end driving policies that reach competitive performance on HUGSIM and NAVSIM-v2 after real-world adaptation and improve with more self-play compute.

  3. DriveJudge: Rethinking Autonomous Driving Evaluation with Vision-Language Models

    cs.CV 2026-06 unverdicted novelty 6.0

    DriveJudge combines VLM reasoning with rule functions on a new 33,577-sample human-annotated dataset, outperforming EPDMS by 21.23 AUC on quality classification and DriveCritic by 6.5% on trajectory preference.

  4. Test-Time Trajectory Optimization for Autonomous Driving

    cs.RO 2026-06 unverdicted novelty 6.0

    TOAD applies test-time Cross-Entropy Method optimization to refine trajectories using the planner's scorer as a reward function, improving end-to-end autonomous driving performance without retraining.

  5. DriveFuture: Future-Aware Latent World Models for Autonomous Driving

    cs.CV 2026-05 unverdicted novelty 6.0

    DriveFuture achieves SOTA results on NAVSIM by conditioning latent world model states on future predictions to directly inform trajectory planning.

  6. GSDrive: Reinforcing Driving Policies by Multi-mode Future Trajectory Probing with 3D Gaussian Splatting Environment

    cs.RO 2026-04 unverdicted novelty 6.0

    GSDrive improves end-to-end driving policies through 3D Gaussian Splatting simulation and multi-mode trajectory probing that supplies dense, differentiable rewards for reinforcement learning.

  7. GSDrive: Reinforcing Driving Policies by Multi-mode Future Trajectory Probing with 3D Gaussian Splatting Environment

    cs.RO 2026-04 unverdicted novelty 6.0

    GSDrive combines IL priors with RL feedback by probing multi-mode futures inside a 3D Gaussian Splatting simulator to supply dense rewards for closed-loop driving policy improvement on nuScenes.

  8. BridgeSim: Unveiling the OL-CL Gap in End-to-End Autonomous Driving

    cs.RO 2026-04 unverdicted novelty 6.0

    The primary OL-CL gap in end-to-end autonomous driving arises from objective mismatch creating structural inability to model reactive behaviors, which a test-time adaptation method can mitigate.

  9. SimScale: Learning to Drive via Real-World Simulation at Scale

    cs.CV 2025-11 conditional novelty 6.0

    SimScale synthesizes unseen driving states from real logs via neural rendering and reactive environments, generates pseudo-expert trajectories, and shows that co-training on real plus simulated data improves planning ...

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

    cs.RO 2026-04 conditional novelty 5.5

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

  11. PriorEye: Geospatial Visual Priors for End-to-End Autonomous Driving

    cs.CV 2026-06 unverdicted novelty 5.0

    PriorEye augments end-to-end driving models with a dual-memory architecture that stores and gates geospatial visual priors to improve performance and robustness to sensor corruption on NAVSIM-v2.

  12. Slow Brain, Fast Planner: Latency-Resilient VLM-Augmented Urban Navigation

    cs.RO 2026-06 unverdicted novelty 5.0

    A training-free fusion layer enables stale VLM selections to improve a real-time planner's trajectory scoring for urban sidewalk navigation, yielding 30% ADE reduction in challenging scenarios.

  13. 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

21 extracted references · 16 linked inside Pith · cited by 12 Pith papers

  1. [1]

    Cosmos world foundation model platform for physical ai.arXiv preprint arXiv:2501.03575,

    Niket Agarwal, Arslan Ali, Maciej Bala, Yogesh Balaji, Erik Barker, Tiffany Cai, Prithvijit Chat- topadhyay, Yongxin Chen, Yin Cui, Yifan Ding, et al. Cosmos world foundation model platform for physical ai.arXiv preprint arXiv:2501.03575,

    Pith/arXiv arXiv

  2. [6]

    Rad: Training an end-to-end driving policy via large-scale 3dgs-based reinforcement learning.arXiv preprint arXiv:2502.13144,

    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 driving policy via large-scale 3dgs-based reinforcement learning.arXiv preprint arXiv:2502.13144,

    arXiv

  3. [10]

    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, 2025a. Kailin Li, Zhenxin Li, Shiyi Lan, Yuan Xie, Zhizhong Zhang, Jiayi Liu, Zuxuan Wu, Zhiding Yu, and Jose M Al...

    arXiv

  4. [11]

    Continuous control with deep reinforcement learning.arXiv preprint arXiv:1509.02971,

    Timothy P Lillicrap, Jonathan J Hunt, Alexander Pritzel, Nicolas Heess, Tom Erez, Yuval Tassa, David Silver, and Daan Wierstra. Continuous control with deep reinforcement learning.arXiv preprint arXiv:1509.02971,

    Pith/arXiv arXiv

  5. [12]

    Safe navigation: Training autonomous vehicles using deep rein- forcement learning in carla.arXiv preprint arXiv:2311.10735,

    Ghadi Nehme and Tejas Y Deo. Safe navigation: Training autonomous vehicles using deep rein- forcement learning in carla.arXiv preprint arXiv:2311.10735,

    Pith/arXiv arXiv

  6. [13]

    High- dimensional continuous control using generalized advantage estimation.arXiv preprint arXiv:1506.02438,

    John Schulman, Philipp Moritz, Sergey Levine, Michael Jordan, and Pieter Abbeel. High- dimensional continuous control using generalized advantage estimation.arXiv preprint arXiv:1506.02438,

    Pith/arXiv arXiv

  7. [14]

    Deepseekmath: Pushing the limits of mathemati- cal reasoning in open language models.arXiv preprint arXiv:2402.03300,

    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 mathemati- cal reasoning in open language models.arXiv preprint arXiv:2402.03300,

    Pith/arXiv arXiv

  8. [15]

    Centaur: Robust end-to-end autonomous driving with test-time training

    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,

    Pith/arXiv arXiv

  9. [16]

    He-drive: Human-like end-to-end driving with vision language models.arXiv preprint arXiv:2410.05051,

    Junming Wang, Xingyu Zhang, Zebin Xing, Songen Gu, Xiaoyang Guo, Yang Hu, Ziying Song, Qian Zhang, Xiaoxiao Long, and Wei Yin. He-drive: Human-like end-to-end driving with vision language models.arXiv preprint arXiv:2410.05051,

    arXiv

  10. [17]

    Enhancing autonomous driving safety with collision scenario integration.arXiv preprint arXiv:2503.03957,

    Zi Wang, Shiyi Lan, Xinglong Sun, Nadine Chang, Zhenxin Li, Zhiding Yu, and Jose M Al- varez. Enhancing autonomous driving safety with collision scenario integration.arXiv preprint arXiv:2503.03957,

    Pith/arXiv arXiv

  11. [19]

    Raw2drive: Reinforcement learning with aligned world models for end-to-end autonomous driving (in carla v2).arXiv preprint arXiv:2505.16394,

    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 driving (in carla v2).arXiv preprint arXiv:2505.16394,

    arXiv

  12. [21]

    Hugsim: A real-time, photo-realistic and closed-loop simulator for autonomous driving.arXiv preprint arXiv:2412.01718,

    Hongyu Zhou, Longzhong Lin, Jiabao Wang, Yichong Lu, Dongfeng Bai, Bingbing Liu, Yue Wang, Andreas Geiger, and Yiyi Liao. Hugsim: A real-time, photo-realistic and closed-loop simulator for autonomous driving.arXiv preprint arXiv:2412.01718,

    Pith/arXiv arXiv

  13. [2017]

    An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929,

    Alexey Dosovitskiy, Lucas Beyer, Alexander Kolesnikov, Dirk Weissenborn, Xiaohua Zhai, Thomas Unterthiner, Mostafa Dehghani, Matthias Minderer, Georg Heigold, Sylvain Gelly, et al. An image is worth 16x16 words: Transformers for image recognition at scale.arXiv preprint arXiv:2010.11929,

    Pith/arXiv arXiv 2010

  14. [2018]

    Robust autonomy emerges from self-play.arXiv preprint arXiv:2502.03349,

    Marco Cusumano-Towner, David Hafner, Alex Hertzberg, Brody Huval, Aleksei Petrenko, Eugene Vinitsky, Erik Wijmans, Taylor Killian, Stuart Bowers, Ozan Sener, et al. Robust autonomy emerges from self-play.arXiv preprint arXiv:2502.03349,

    Pith/arXiv arXiv

  15. [2019]

    Offline reinforcement learning: Tuto- rial, review, and perspectives on open problems.arXiv preprint arXiv:2005.01643,

    Sergey Levine, Aviral Kumar, George Tucker, and Justin Fu. Offline reinforcement learning: Tuto- rial, review, and perspectives on open problems.arXiv preprint arXiv:2005.01643,

    Pith/arXiv arXiv 2005

  16. [2020]

    End- to-end autonomous driving: Challenges and frontiers.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024a

    Li Chen, Penghao Wu, Kashyap Chitta, Bernhard Jaeger, Andreas Geiger, and Hongyang Li. End- to-end autonomous driving: Challenges and frontiers.IEEE Transactions on Pattern Analysis and Machine Intelligence, 2024a. Shaoyu Chen, Bo Jiang, Hao Gao, Bencheng Liao, Qing Xu, Qian Zhang, Chang Huang, Wenyu Liu, and Xinggang Wang. Vadv2: End-to-end vectorized au...

    Pith/arXiv arXiv

  17. [2021]

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

    Lihe Yang, Bingyi Kang, Zilong Huang, Xiaogang Xu, Jiashi Feng, and Hengshuang Zhao. Depth anything: Unleashing the power of large-scale unlabeled data.arXiv preprint arXiv:2401.10891,

    Pith/arXiv arXiv

  18. [2022]

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

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

    Pith/arXiv arXiv

  19. [2023]

    Dist-4d: Disentangled spatiotemporal diffusion with metric depth for 4d driving scene generation.arXiv preprint arXiv:2503.15208,

    Jiazhe Guo, Yikang Ding, Xiwu Chen, Shuo Chen, Bohan Li, Yingshuang Zou, Xiaoyang Lyu, Feiyang Tan, Xiaojuan Qi, Zhiheng Li, and Hao Zhao. Dist-4d: Disentangled spatiotemporal diffusion with metric depth for 4d driving scene generation.arXiv preprint arXiv:2503.15208,

    Pith/arXiv arXiv

  20. [2024]

    A comprehensive review of reinforce- ment learning for autonomous driving in the carla simulator.arXiv preprint arXiv:2509.08221,

    Elahe Delavari, Feeza Khan Khanzada, and Jaerock Kwon. A comprehensive review of reinforce- ment learning for autonomous driving in the carla simulator.arXiv preprint arXiv:2509.08221,

    Pith/arXiv arXiv

  21. [2025]

    Drivesuprim: Towards precise trajectory selection for end-to-end planning.arXiv preprint arXiv:2506.06659,

    Wenhao Yao, Zhenxin Li, Shiyi Lan, Zi Wang, Xinglong Sun, Jose M Alvarez, and Zuxuan Wu. Drivesuprim: Towards precise trajectory selection for end-to-end planning.arXiv preprint arXiv:2506.06659,

    arXiv