ReconDreamer-RL: Enhancing Reinforcement Learning via Diffusion-based Scene Reconstruction

Chaojun Ni; Guanghong Jia; Guan Huang; Guosheng Zhao; Wenjun Mei; Wenkang Qin; Xiaofeng Wang; Xinze Chen; Zheng Zhu

REVIEW 3 major objections 5 minor 2 cited by

ReconDreamer-RL: Enhancing Reinforcement Learning via Diffusion-based Scene Reconstruction

T0 review · 3 major / 5 minor · reviewed 2026-08-05 · deepseek-v4-flash

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

Pith's one-line read ReconDreamer-RL claims a 5x collision-ratio reduction for end-to-end driving by training in diffusion-reconstructed scenes.

desk verdict Useful RL-in-3DGS driving framework whose headline collision gains are probably inflated because the evaluation scenes come from the same generator used to create training tasks. read the letter →

arxiv 2508.08170 v2 pith:BMBG3XJW submitted 2025-08-11 cs.CV

Chaojun Ni , Guosheng Zhao , Xiaofeng Wang , Zheng Zhu , Wenkang Qin , Xinze Chen , Guanghong Jia , Guan Huang

show 1 more author

Wenjun Mei

This is my paper

classification cs.CV

keywords autonomousdrivingreinforcementlearning3DGaussianSplattingvideodiffusionpriorscenereconstructionclosed-loopsimulationcornercasesbehaviorcloning

verification ladder T0 review T1 audit T2 compute T3 formal

The pith

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

The reading

ReconDreamer-RL aims to make closed-loop reinforcement learning for end-to-end autonomous driving practical by building the training environment itself from real driving data. The authors argue that pure 3D Gaussian Splatting reconstruction renders poorly on trajectories not seen in the recorded data, so they add a video diffusion prior to restore novel-view videos and a kinematic bicycle model to keep vehicle motion physically plausible. On top of this simulator, a Dynamic Adversary Agent (DAA) generates corner cases such as cut-ins, and a Cousin Trajectory Generator (CTG) diversifies the ego vehicle's action distribution away from straight-line driving. The central claim is that this combination lowers the Collision Ratio in closed-loop evaluation to 0.077, a 3x improvement over the RAD reinforcement-learning baseline and roughly 5x over imitation-learning baselines. If true, it would mean safer driving policies can be trained without expensive real-world trial and error.

What carries the argument

The operating core is the ReconSimulator loop: 3D Gaussian Splatting scene reconstruction for appearance, a video diffusion prior (DriveRestorer) that restores rendered novel-trajectory videos and is used to refine the reconstruction, and a kinematic bicycle model that updates each vehicle pose $W_t=[R_t|P_t]$ via $P_{t+1}=P_t+v_t\Delta t\,\hat{d}_t$ and $R_{t+1}=\mathrm{Rot}_z(\Delta\theta_t)R_t$ with $\Delta\theta_t=\frac{v_t}{L}\tan(\delta_t)\Delta t$, so edited trajectories stay physically plausible. Around this core, DAA is the mechanism that injects corner cases into both training stages, and CTG is the mechanism that rebalances the action distribution.

What would settle it

Run the trained policy on a disjoint set of corner-case scenarios produced by an independent generator, or on a real closed-loop test track, with DAA-edited scenes excluded from evaluation; if the collision-ratio gap over RAD collapses, the reported 3x-5x reductions are largely distribution overfitting.

Watch

Extended reading notes

Core claim

The paper's central discovery is that integrating a video diffusion prior into a 3DGS-based simulator removes the reconstruction-quality bottleneck that stops reinforcement learning from exploiting novel trajectories, and that adversarial trajectory editing plus trajectory interpolation supplies the missing corner-case and action diversity. ReconSimulator separates static background and rigid moving-vehicle Gaussians, renders novel paths, and uses the DriveRestorer from ReconDreamer to clean artifacts and fine-tune the reconstruction. DAA selects target vehicles in BEV, generates behavior-conditioned trajectories, checks them against drivable-region, kinematic, and inter-vehicle constraints,

Load-bearing premise

The load-bearing premise is that the closed-loop evaluation benchmark, which includes corner-case scenes edited by the same DAA machinery used in training, is an unbiased test of collision avoidance rather than a test of familiarity with the scenario generator.

Editorial extensions

If this is right

If ReconSimulator renders faithfully enough on novel trajectories, reinforcement learning can be run in closed loop on scenes reconstructed from real sensor data, without a game-engine simulator.
DAA-style adversarial trajectory generation can populate imitation-learning datasets with cut-in and hard-brake cases that are rare in human driving logs.
CTG's trajectory extension and interpolation counters the straight-line bias in expert data, which should improve behavior-cloning initialization before RL.
The reported collision-ratio numbers imply that the diffusion prior, not just RL itself, is a major driver of safety in corner cases.
The framework extends to Waymo reconstructed scenes, suggesting the recipe is dataset-general rather than nuScenes-specific.

Reading between the lines

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

The evaluation scenes are edited with the same DAA trajectory-edit machinery used in training, so the 3x-5x collision-ratio reductions may partly measure familiarity with the scenario generator; a held-out split generated by an independent simulator would test this.
Because CTG generates kinematically checked trajectories from expert data, the same interpolation-and-extension idea could be applied as a standalone data augmentation for imitation-only driving pipelines, with or without RL.
The kinematic bicycle-model constraint is what makes DAA edits physically plausible; if the constraint is too permissive or too strict, the difficulty distribution of generated corner cases shifts, so tuning it could directly trade off training difficulty against realism.
A natural extension is to use the same diffusion prior to render sensor data for perception fine-tuning on novel routes, such as evaluating sensor failure or weather shifts, which the paper does not test.

Share X Bluesky LinkedIn Reddit HN

Editorial analysis

A structured set of objections, weighed in public.

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

Desk Editor's Note

read the letter

Here's the takeaway: ReconDreamer-RL is a coherent, well-engineered framework for training end-to-end driving policies in reconstructed 3D environments, and the ablations support the value of each module. But the headline safety numbers (3x vs RAD, 5x vs IL) should be read skeptically. The test benchmark includes edited scenes (DynamicCutIn, HardBrake, etc.) that are generated by the DAA mechanism, and DAA is also used during both training stages, with trajectory fine-tuning at some probability. So the test distribution overlaps with the training task distribution. Baselines like VAD or RAD never saw DAA-generated conditions, so part of the gap is exposure rather than learned generalization.

What's genuinely useful: ReconSimulator adds a video diffusion prior to 3DGS rendering, which measurably improves novel-view quality (NTA-IoU, FID) and rendering speed is practical (125 FPS). The kinematic bicycle model for physical plausibility is a sensible addition. CTG is a simple but effective data-augmentation idea, and the Cousin-nuScenes distribution shift is visible. The Waymo cross-dataset experiment is a real plus, even though it uses the same edited-scene protocol.

The soft spots: besides the train/test overlap, there is no variance reporting (single run per metric), no disjoint held-out scenario set, and no code or trained-model release. Distance thresholds and trajectory-generation details are underspecified. The Waymo consistency check helps but doesn't resolve the overlap because it uses the same scenario-generation machinery.

The math is standard and internally consistent; the paper is not sloppy. The central claim is plausible but not established by the current evaluation.

This is a paper for builders of closed-loop AD simulators. It deserves a serious referee: the core idea is solid, the flaws are correctable, and a peer-review round could force the authors to add a held-out scenario split and report variance. I'd send it to review, but with a clear request to make the evaluation adversarial to their own generator.

Referee Report

3 major / 5 minor

Summary. The paper proposes ReconDreamer-RL, a framework for training end-to-end autonomous driving policies in closed-loop, photorealistic 3DGS-based simulators. The method has three components: (i) ReconSimulator, which augments 3D Gaussian Splatting scene reconstruction with a video diffusion prior (DriveRestorer) and enforces physical plausibility through a kinematic bicycle model; (ii) the Dynamic Adversary Agent (DAA), which generates corner-case trajectories such as cut-ins by modifying surrounding vehicles' paths in both the imitation-learning and reinforcement-learning stages; and (iii) the Cousin Trajectory Generator (CTG), which extends and interpolates expert trajectories to create a more action-diverse training set (Cousin-nuScenes). The authors train a RAD-style driving policy with a two-stage IL+RL procedure and evaluate it in closed-loop 3DGS environments reconstructed from nuScenes and Waymo, including edited corner-case scenes. The headline results are a Collision Ratio of 0.077, a 3x reduction relative to RAD (0.238) and roughly a 5x reduction relative to imitation-learning baselines. Ablations show each module reducing CR in the expected direction, and a Waymo experiment reports a similar improvement over RAD (0.093 vs 0.140).

Significance. If the results are taken at face value, the paper makes a useful empirical contribution: it demonstrates that combining diffusion-based appearance restoration with kinematic constraints in a reconstructed 3DGS simulator can improve closed-loop RL training for driving, and that adversarial trajectory augmentation plus trajectory diversity is helpful. The mathematical components are standard and clearly described, and the ablations in Tab. 2 are internally consistent. The cross-dataset Waymo check, though using the same evaluation protocol, is a positive sign. However, the central safety claim rests on an evaluation benchmark whose edited scenarios are generated by the same DAA module used during training, so the reported improvements may reflect train/test overlap rather than safer generalizable driving. Because no disjoint held-out scenario split or independent closed-loop evaluation is reported, the main quantitative conclusion is currently not adequately supported. The paper is nevertheless a reasonable candidate after a major revision that adds unbiased evaluation.

major comments (3)

[Method, 'Dynamic Adversary Agent'; Experiments, 'Main Results'; Supplement, 'Details of Edited Scenes' and 'Visualizatio] The evaluation benchmark and the training data are not independent. The test benchmark includes edited scenarios such as DynamicCutIn and HardBrake (Supplement, 'Details of Edited Scenes'), while the same DAA module is used in both training stages: the method states DAA 'can be used in both imitation learning and reinforcement learning stages' and during RL it 'continues to dynamically alter surrounding vehicle trajectories to create corner cases.' The supplement further states that 'DAA has a certain probability of fine-tuning the trajectories in the edited scenes' during RL. Thus the test scenarios are sampled from the same generator that produced the training tasks. The IL/RL baselines (VAD, GenAD, VADv2, RAD) were not trained on DAA-generated corner cases, so their higher CR in Tabs. 1 and 4 may simply reflect lack of exposure to the test generator's output, not poorer driving abilit
[Experiments, 'Evaluation Metrics'; Tabs. 1 and 4] The paper reports CR, DCR, SCR, DR, PDR, and HDR as ratios over clips but does not state the total number of evaluation clips, the number of repeated runs or seeds, or any confidence intervals. Given that the main claim is a large reduction in collision ratio from 0.238 to 0.077, the lack of uncertainty quantification makes it impossible to assess whether the difference is statistically significant or stable. In addition, the PDR and HDR thresholds in Eqs. (15)-(16) are not specified, so the DR numbers in Tab. 1 cannot be reproduced. Please provide the evaluation corpus size, per-scenario breakdown, variance over seeds, and the exact deviation thresholds.
[Main Results, 'Corner Case Results', Tab. 4] The statement that ReconDreamer-RL achieves a '404.5% improvement in Collision Ratio over imitation methods' is not a standard usage. The CR improves from, e.g., 0.449 to 0.089, which is an 80.2% relative reduction, not a 404.5% improvement. If the authors intend the reciprocal ratio, that should be stated explicitly and consistently with the abstract's '5x reduction' phrasing. As written, the percentage claim is misleading and should be corrected.

minor comments (5)

[Experiments, 'Ablation Studies', Tab. 3] The text says 'without ReconSimulator,' but the ablation actually removes the video diffusion prior, not the entire ReconSimulator. Rephrase to 'w/o video diffusion prior' or clarify the component under study.
[Method, 'Dynamic Adversary Agent', Eq. (9)] The sentence 'avoid collisions with other vehicles (collisions with the ego-vehicle are allowed)' is ambiguous. Does the set {o_j} include the ego vehicle? If collisions with the ego are allowed, the condition should clearly exclude the ego from the set; otherwise the text contradicts the constraint.
[Abstract and Introduction] Minor typographical issues: 'InReconDreamer-RL' missing a space; the abstract states 'a 5x reduction in the Collision Ratio' while the introduction says 'a 5× reduction in collision rate.' Use consistent terminology and formatting.
[Supplement, 'Visualization of DAA'] The probability of DAA fine-tuning trajectories in the RL stage is mentioned but never quantified, and the effect of the fine-tuning probability is not ablated. Please state the value used and, ideally, include a sensitivity analysis.
[Figure 6] The claim that Cousin-nuScenes has '4x more non-straight-line driving maneuvers' is based on the bar chart, but the definition of 'Others' and the exact counting procedure are not given. Please define the categories and report the composition of the final training set.

Circularity Check

1 steps flagged · score 4.0 of 10

Headline collision-ratio reductions are measured on edited scenes whose trajectories are produced/fine-tuned by the same DAA mechanism used during training, so the benchmark partially overlaps the training distribution; otherwise the paper's components are validated by ablations and not derivationally circular.

fitted input called prediction [Experiments (Experimental Setup); Method (Dynamic Adversary Agent); Supplement (Visualization of DAA)]
"Finally, we build an evaluation benchmark including edited 3DGS environments... DAA can be used in both imitation learning and reinforcement learning stages. In the first stage, DAA generates corner case scenarios... ReconSimulator then renders offline autonomous driving data... As mentioned in the main text, in the reinforcement learning stage, the Dynamic Adversary Agent (DAA) has a certain probability of fine-tuning the trajectories in the edited scenes... instead of directly reusing those generated during the imitation stage."

The headline 3x/5x collision-ratio reductions are computed on an evaluation benchmark that includes edited 3DGS scenes. Those edited scenes are the same ones in which DAA generates/fine-tunes trajectories during training: DAA generates corner cases for the imitation stage and continues to alter surrounding-vehicle trajectories in the RL stage, and the Supplement states that DAA fine-tunes trajectories in the edited scenes rather than reusing imitation-stage ones. Thus the test scenarios are drawn from the same generator that produced the training tasks. Reporting CR on this benchmark as evidence of safer driving is therefore partly a re-measurement of the policy's exposure to DAA-generated data, not an independent prediction of generalization. Baselines were not trained on DAA-generated co

full rationale

The paper is an empirical systems paper; there is no derivation whose equations reduce to its inputs. ReconSimulator's diffusion prior is supported by an internal ablation (Tab. 3) using detector-based and FID metrics, so citing ReconDreamer/DriveDreamer-2 is not load-bearing. CTG is validated by ablations on original scenes, and the Waymo experiment is on a different dataset (though it uses the same edited-scene protocol). The main circularity concern is the DAA train/test overlap: the evaluation benchmark's edited scenes are the same ones DAA operates in during training, so the headline collision-ratio numbers are not a clean held-out test. This is a validity threat, not a self-definitional derivation, so the score is 4.

Assumptions & free parameters 6 free parameters · 5 assumptions · 0 invented entities

No new physical entities are introduced. The central claim rests on the domain assumptions above and on a set of hand-chosen thresholds and constraints that are not fully specified. The software modules DAA, CTG, and ReconSimulator are engineering constructs, not invented entities in the sense of new physical postulates.

free parameters (6)

DAA distance thresholds for target vehicle selection = not specified (referenced to Kesting and Treiber 2013)
Distance settings that choose which vehicles become adversarial targets are behavior-dependent and not reported in the paper.
dmin, minimum distance in DAA trajectory feasibility check (Eq. 9) = not specified
Threshold that accepts or rejects generated trajectories; its value is not given, changing it alters which corner cases are produced.
Per-vehicle-category kinematic constraints (wheelbase L, max steering angle, max velocity) = not specified
Used in Eqs. 4-7 and in trajectory checks; the paper says parameters differ per category but lists no values.
CTG interpolation density m (Eq. 11) = not specified
Number of interpolated points between expert trajectory timesteps; directly controls the size and variety of Cousin-nuScenes.
PDR and HDR deviation thresholds (Eqs. 15-16) = not specified
Predefined thresholds decide which clips count as deviations; they define the headline CR and DR metrics, so they are load-bearing.
DAA trajectory fine-tuning probability in the RL stage = not specified ('a certain probability')
Controls how often DAA perturbs trajectories during RL training instead of reusing imitation-stage ones; value is not reported.

assumptions (5)

domain assumption 3DGS reconstruction of nuScenes/Waymo scenes is a faithful enough simulator that policies trained in it transfer to the evaluation benchmark and to real driving.
The whole paper measures training improvements only inside the same reconstructed environments (Experiments). Real-world transfer is never tested.
domain assumption DriveRestorer's video-diffusion restoration improves novel-trajectory rendering without introducing plan-misleading artifacts.
Assumed from prior work (Ni et al. 2024); Tab. 3 shows rendering-quality gains but no direct causal link to policy safety.
standard math The kinematic bicycle model (Eqs. 4-7) is sufficient to model physical feasibility of all vehicle trajectories in the simulator.
Standard low-speed vehicle kinematics; the paper does not validate it against recorded vehicle dynamics.
domain assumption DAA-generated corner-case trajectories are representative of real dangerous traffic events, so training on them generalizes to reality.
Evaluation uses the same editing pipeline that DAA uses for training (selection-bias risk); no real-world validation is provided.
domain assumption The evaluation benchmark's edited scenes (DynamicCutIn, HardBrake, etc.) are a fair and unbiased test of collision avoidance.
The edited scene list overlaps with DAA training behaviors; fairness is asserted, not established through a disjoint held-out split.

how reviews work

0 comments

Cite this review

Pith. "Pith review of ReconDreamer-RL: Enhancing Reinforcement Learning via Diffusion-based Scene Reconstruction." pith.science (2026). https://pith.science/paper/BMBG3XJW

@misc{pith2026250808170,
  author       = {Pith},
  title        = {Pith review of: ReconDreamer-RL: Enhancing Reinforcement Learning via Diffusion-based Scene Reconstruction},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/BMBG3XJW}},
  note         = {Machine review of arXiv:2508.08170}
}

read the original abstract

Reinforcement learning for training end-to-end autonomous driving models in closed-loop simulations is gaining growing attention. However, most simulation environments differ significantly from real-world conditions, creating a substantial simulation-to-reality (sim2real) gap. To bridge this gap, some approaches utilize scene reconstruction techniques to create photorealistic environments as a simulator. While this improves realistic sensor simulation, these methods are inherently constrained by the distribution of the training data, making it difficult to render high-quality sensor data for novel trajectories or corner case scenarios. Therefore, we propose ReconDreamer-RL, a framework designed to integrate video diffusion priors into scene reconstruction to aid reinforcement learning, thereby enhancing end-to-end autonomous driving training. Specifically, in ReconDreamer-RL, we introduce ReconSimulator, which combines the video diffusion prior for appearance modeling and incorporates a kinematic model for physical modeling, thereby reconstructing driving scenarios from real-world data. This narrows the sim2real gap for closed-loop evaluation and reinforcement learning. To cover more corner-case scenarios, we introduce the Dynamic Adversary Agent (DAA), which adjusts the trajectories of surrounding vehicles relative to the ego vehicle, autonomously generating corner-case traffic scenarios (e.g., cut-in). Finally, the Cousin Trajectory Generator (CTG) is proposed to address the issue of training data distribution, which is often biased toward simple straight-line movements. Experiments show that ReconDreamer-RL improves end-to-end autonomous driving training, outperforming imitation learning methods with a 5x reduction in the Collision Ratio.

Figures

Figures reproduced from arXiv: 2508.08170 by the authors.

**Figure 1.** In ReconDreamer-RL, ReconSimulator improves appearance modeling by ReconDreamer and incorporates physical modeling to reconstruct driving scenes. In the imitation learning stage, DAA generates corner-case scenario trajectories, while CTG diversifies the ego vehicle’s actions and uses ReconSimulator to render sensor data for training the policy. In the reinforcement learning stage, the policy is trained in a closed-… view at source ↗

**Figure 2.** The process of integrating the diffusion prior for [PITH_FULL_IMAGE:figures/full_fig_p003_2.png] view at source ↗

**Figure 3.** Examples of Dynamic Adversary Agent (DAA) controlling surrounding vehicles to simulate cut-in scenarios. [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗

Figures from the paper (12 more)

**Figure 4.** Figure 4: The pipeline of the DAA. DAA identifies the tar [PITH_FULL_IMAGE:figures/full_fig_p004_4.png]

**Figure 5.** Figure 5: Cousin Trajectory Generator (CTG) generates cousin trajectories and performs trajectory checks to eliminate unrea [PITH_FULL_IMAGE:figures/full_fig_p005_5.png]

**Figure 6.** Figure 6: The Cousin-nuScenes dataset created by CTG has [PITH_FULL_IMAGE:figures/full_fig_p005_6.png]

**Figure 7.** Figure 7: Comparison of different methods in challenging corner cases, with collisions highlighted by orange boxes. [PITH_FULL_IMAGE:figures/full_fig_p006_7.png]

**Figure 8.** Figure 8: Various trajectories generated by the Dynamic Adversary Agent (DAA) for the vehicle in the red box at the same [PITH_FULL_IMAGE:figures/full_fig_p013_8.png]

**Figure 9.** Figure 9: Comparison of different methods in challenging corner cases, with collisions highlighted by red boxes. [PITH_FULL_IMAGE:figures/full_fig_p014_9.png]

**Figure 10.** Figure 10: Example scenario of BlockedIntersection. The vehicle in the red box is stationary and blocks the intersection when [PITH_FULL_IMAGE:figures/full_fig_p015_10.png]

**Figure 11.** Figure 11: Example scenario of DynamicCutIn. The vehicle in the red box is attempting to cut in from an adjacent lane while [PITH_FULL_IMAGE:figures/full_fig_p015_11.png]

**Figure 12.** Figure 12: Example scenario of OppositeLaneIntrusion. The vehicle in the red box is attempting to encroach into the ego [PITH_FULL_IMAGE:figures/full_fig_p016_12.png]

**Figure 13.** Figure 13: Example scenario of ParkingCutIn. The vehicle in the red box is attempting to exit a parallel parking space and [PITH_FULL_IMAGE:figures/full_fig_p016_13.png]

**Figure 14.** Figure 14: Example scenario of HardBrake. The vehicle in the red box is decelerating abruptly. The ego vehicle needs to change [PITH_FULL_IMAGE:figures/full_fig_p017_14.png]

**Figure 15.** Figure 15: Example scenario of HazardAtSideLane. The vehicle in the red box is slowly moving and partially occupying the [PITH_FULL_IMAGE:figures/full_fig_p017_15.png]

Discussion (0). Sign in to comment.

Forward citations

Cited by 2 Pith papers

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

GigaWorld-1: A Roadmap to Build World Models for Robot Policy Evaluation
cs.RO 2026-07 conditional novelty 6.0 of 10

Long-horizon action-faithful consistency, not short-term visual realism, dominates world-model reliability for robot policy evaluation; GigaWorld-1 implements that roadmap and gains 14.9% on evaluator-alignment metrics.
GigaWorld-Policy-0.5: A Faster and Stronger WAM Empowered by AutoResearch
cs.RO 2026-07 conditional novelty 4.0 of 10

GigaWorld-Policy-0.5 uses a Mixture-of-Transformers action-expert split and mixed world-model pretraining to reach 85 ms action-only inference with claimed success-rate gains.

Reference graph

Works this paper leans on

10 extracted references · 1 canonical work pages · cited by 2 Pith papers

[3]

arXiv preprint arXiv:2311.18561

Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561. Chen, Z.; Yang, J.; Huang, J.; Lutio, R. d.; Esturo, J. M.; Ivanovic, B.; Litany, O.; Gojcic, Z.; Fidler, S.; Pavone, M.; Song, L.; and Wang, Y . 2024b. OmniRe: Omni Urban Scene Reconstruction. arXiv preprint arXiv:2408.16760. Cheng, J...

arXiv 2024
[6]

ACM Trans

3D Gaussian splatting for real-time radiance field ren- dering. ACM Trans. Graph., 42(4): 139–1. Kesting, A.; and Treiber, M. 2013. Traffic flow dynamics: data, models and simulation. No. Book, Whole)(Springer Berlin Heidelberg, Berlin, Heidelberg. Khanam, R.; and Hussain, M. 2024. Yolov11: An overview of the key architectural enhancements. arXiv preprint...

arXiv 2013
[9]

In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , 15449– 15458

Para-drive: Parallelized architecture for real-time au- tonomous driving. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , 15449– 15458. Wu, Z.; Liu, T.; Luo, L.; Zhong, Z.; Chen, J.; Xiao, H.; Hou, C.; Lou, H.; Chen, Y .; Yang, R.; et al. 2023. Mars: An instance-aware, modular and realistic simulator for au- tonomo...

arXiv 2023
[10]

IEEE Transactions on Cy- bernetics

A survey of imitation learning: Algorithms, recent developments, and challenges. IEEE Transactions on Cy- bernetics. Zhang, D.; Liang, J.; Guo, K.; Lu, S.; Wang, Q.; Xiong, R.; Miao, Z.; and Wang, Y . 2025. Carplanner: Consistent auto- regressive trajectory planning for large-scale reinforcement learning in autonomous driving. In Proceedings of the Com- p...

arXiv 2025
[158]

Li, Z.; Li, K.; Wang, S.; Lan, S.; Yu, Z.; Ji, Y .; Li, Z.; Zhu, Z.; Kautz, J.; Wu, Z.; et al

Springer. Li, Z.; Li, K.; Wang, S.; Lan, S.; Yu, Z.; Ji, Y .; Li, Z.; Zhu, Z.; Kautz, J.; Wu, Z.; et al. 2024b. Hydra-mdp: End-to-end mul- timodal planning with multi-target hydra-distillation. arXiv preprint arXiv:2406.06978. Li, Z.; Wang, W.; Li, H.; Xie, E.; Sima, C.; Lu, T.; Yu, Q.; and Dai, J. 2024c. Bevformer: learning bird’s-eye-view rep- resentati...

arXiv 2024
[2019]

CARIL: Confidence-Aware Regression in Imitation Learning for Autonomous Driving

Exploring the limitations of behavior cloning for au- tonomous driving. In Proceedings of the IEEE/CVF inter- national conference on computer vision, 9329–9338. Delavari, E.; Khalil, A.; and Kwon, J. 2025. CARIL: Confidence-Aware Regression in Imitation Learning for Au- tonomous Driving. arXiv preprint arXiv:2503.00783. Di Palo, N.; and Johns, E. 2024. Ke...

work page Pith review arXiv 2025
[2020]

nuscenes: A multimodal dataset for autonomous driv- ing. In CVPR. Caesar, H.; Kabzan, J.; Tan, K. S.; Fong, W. K.; Wolff, E.; Lang, A.; Fletcher, L.; Beijbom, O.; and Omari, S. 2021. nuplan: A closed-loop ml-based planning benchmark for au- tonomous vehicles. arXiv preprint arXiv:2106.11810. Che, Q.-H.; Nguyen, D.-P.; Pham, M.-Q.; and Lam, D.-K

arXiv 2021
[2021]

Advances in Neural Informa- tion Processing Systems, 34: 11702–11716

Bridging offline reinforcement learning and imitation learning: A tale of pessimism. Advances in Neural Informa- tion Processing Systems, 34: 11702–11716. Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; and Om- mer, B. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF confer- ence on computer vision and ...

arXiv 2022

Show all 10 references

[2023]

TwinLiteNet: An efficient and lightweight model for driveable area and lane segmentation in self-driving cars. In MAPR. Chen, S.; Jiang, B.; Gao, H.; Liao, B.; Xu, Q.; Zhang, Q.; Huang, C.; Liu, W.; and Wang, X. 2024a. Vadv2: End-to- end vectorized autonomous driving via proba...

arXiv
[2024]

Advances in Neural Information Processing Systems, 37: 819–844

Bench2drive: Towards multi-ability benchmarking of closed-loop end-to-end autonomous driving. Advances in Neural Information Processing Systems, 37: 819–844. Jiang, B.; Chen, S.; Xu, Q.; Liao, B.; Chen, J.; Zhou, H.; Zhang, Q.; Liu, W.; Huang, C.; and Wang, X. 2023. Vad: Vecto...

2023 arXiv

Pith tools

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

[1] [3]

arXiv preprint arXiv:2311.18561

Periodic vibration gaussian: Dynamic urban scene reconstruction and real-time rendering. arXiv preprint arXiv:2311.18561. Chen, Z.; Yang, J.; Huang, J.; Lutio, R. d.; Esturo, J. M.; Ivanovic, B.; Litany, O.; Gojcic, Z.; Fidler, S.; Pavone, M.; Song, L.; and Wang, Y . 2024b. OmniRe: Omni Urban Scene Reconstruction. arXiv preprint arXiv:2408.16760. Cheng, J...

arXiv 2024

[2] [6]

ACM Trans

3D Gaussian splatting for real-time radiance field ren- dering. ACM Trans. Graph., 42(4): 139–1. Kesting, A.; and Treiber, M. 2013. Traffic flow dynamics: data, models and simulation. No. Book, Whole)(Springer Berlin Heidelberg, Berlin, Heidelberg. Khanam, R.; and Hussain, M. 2024. Yolov11: An overview of the key architectural enhancements. arXiv preprint...

arXiv 2013

[3] [9]

In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , 15449– 15458

Para-drive: Parallelized architecture for real-time au- tonomous driving. In Proceedings of the IEEE/CVF Confer- ence on Computer Vision and Pattern Recognition , 15449– 15458. Wu, Z.; Liu, T.; Luo, L.; Zhong, Z.; Chen, J.; Xiao, H.; Hou, C.; Lou, H.; Chen, Y .; Yang, R.; et al. 2023. Mars: An instance-aware, modular and realistic simulator for au- tonomo...

arXiv 2023

[4] [10]

IEEE Transactions on Cy- bernetics

A survey of imitation learning: Algorithms, recent developments, and challenges. IEEE Transactions on Cy- bernetics. Zhang, D.; Liang, J.; Guo, K.; Lu, S.; Wang, Q.; Xiong, R.; Miao, Z.; and Wang, Y . 2025. Carplanner: Consistent auto- regressive trajectory planning for large-scale reinforcement learning in autonomous driving. In Proceedings of the Com- p...

arXiv 2025

[5] [158]

Li, Z.; Li, K.; Wang, S.; Lan, S.; Yu, Z.; Ji, Y .; Li, Z.; Zhu, Z.; Kautz, J.; Wu, Z.; et al

Springer. Li, Z.; Li, K.; Wang, S.; Lan, S.; Yu, Z.; Ji, Y .; Li, Z.; Zhu, Z.; Kautz, J.; Wu, Z.; et al. 2024b. Hydra-mdp: End-to-end mul- timodal planning with multi-target hydra-distillation. arXiv preprint arXiv:2406.06978. Li, Z.; Wang, W.; Li, H.; Xie, E.; Sima, C.; Lu, T.; Yu, Q.; and Dai, J. 2024c. Bevformer: learning bird’s-eye-view rep- resentati...

arXiv 2024

[6] [2019]

CARIL: Confidence-Aware Regression in Imitation Learning for Autonomous Driving

Exploring the limitations of behavior cloning for au- tonomous driving. In Proceedings of the IEEE/CVF inter- national conference on computer vision, 9329–9338. Delavari, E.; Khalil, A.; and Kwon, J. 2025. CARIL: Confidence-Aware Regression in Imitation Learning for Au- tonomous Driving. arXiv preprint arXiv:2503.00783. Di Palo, N.; and Johns, E. 2024. Ke...

work page Pith review arXiv 2025

[7] [2020]

nuscenes: A multimodal dataset for autonomous driv- ing. In CVPR. Caesar, H.; Kabzan, J.; Tan, K. S.; Fong, W. K.; Wolff, E.; Lang, A.; Fletcher, L.; Beijbom, O.; and Omari, S. 2021. nuplan: A closed-loop ml-based planning benchmark for au- tonomous vehicles. arXiv preprint arXiv:2106.11810. Che, Q.-H.; Nguyen, D.-P.; Pham, M.-Q.; and Lam, D.-K

arXiv 2021

[8] [2021]

Advances in Neural Informa- tion Processing Systems, 34: 11702–11716

Bridging offline reinforcement learning and imitation learning: A tale of pessimism. Advances in Neural Informa- tion Processing Systems, 34: 11702–11716. Rombach, R.; Blattmann, A.; Lorenz, D.; Esser, P.; and Om- mer, B. 2022. High-resolution image synthesis with latent diffusion models. In Proceedings of the IEEE/CVF confer- ence on computer vision and ...

arXiv 2022