Pith. sign in

REVIEW 4 major objections 4 minor 49 references

DecoupleGS: a decoupled 3DGS simulator for interactive E2E driving testing.

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 ↗

DecoupleGS splits 3D Gaussian Splatting scenes into a persistent background and compressible, relightable vehicle assets to run interactive closed-loop tests of end-to-end driving policies.

T0 review reviewed 2026-08-04 challenge →

load-bearing objection Useful decoupled 3DGS system for E2E simulation, but the headline closed-loop comparison is not controlled and the ADE metric is entangled with the registration objective. the 4 major comments →

arxiv 2608.01761 v1 pith:SGXSS3H5 submitted 2026-08-03 cs.CV

DecoupleGS: Interactive 3D Gaussian Splatting for End-to-End Autonomous Driving Testing

Siying Li , Ying Ni , Jie Sun , Jian Sun , Haotian Shi This is my paper
classification cs.CV
keywords 3D Gaussian Splattingend-to-end autonomous drivingclosed-loop simulationsensor simulationdynamic scene compositionasset compressionmap-guided registrationrelighting
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.

The reading

DecoupleGS is a simulation framework for testing end-to-end autonomous driving policies. It splits each recorded scene into a fixed, high-fidelity 3D Gaussian Splatting background and a library of compact, moveable vehicle models, then recombines them in real time. The paper argues this decomposition resolves three conflicts—memory, geometric alignment, and lighting—that block previous neural simulators from closing the loop around a planner. If the framework works as reported, it gives E2E algorithms an interactive, photorealistic testbed that runs at 45 FPS, with vehicles that respond to the ego car rather than replaying a fixed log.

Core claim

The paper's central claim is that a scene can be decomposed into a persistent static background and object-centric canonical vehicle volumes without losing photorealism or interactivity. Three targeted modules carry the argument: semantic-aware compression prunes perceptually unimportant Gaussians and vector-quantizes the rest so many vehicles fit in memory; map-guided registration aligns vehicle trajectories to lane topology via DTW, SE(2) Procrustes correction, and opacity-accumulated vertical grounding; proxy-based relighting transfers local ambient SH illumination to inserted assets and adds dynamic contact shadows. With these modules, the framework reports the best balance of fidelity a

What carries the argument

The load-bearing mechanism is the decoupled scene representation: a time-invariant background field plus K canonical vehicle volumes, fused by SE(3) transforms before rasterization. View-dependent color is kept correct under rotation by applying the Wigner D-matrix to spherical-harmonic coefficients. Around that core sit the three conflict-solving modules—importance scoring (visibility, contrast, entropy) plus vector quantization for compression; constrained DTW lane matching, Orthogonal Procrustes, and opacity-weighted vertical grounding for geometry; and a linear SH transfer operator plus super-ellipse contact shadow mask for relighting.

Load-bearing premise

The load-bearing premise is that every scene where agents are inserted has reliable lane centerlines (from an HD map or a vectorized map extractor) and a clean segmenter-inpainted background with no residual floaters or shadows.

What would settle it

Run DecoupleGS on an overpass or under-bridge scene whose lane polylines are missing or poorly detected, insert a vehicle, and measure Trajectory ADE and Ground Penetration Rate; if ADE departs substantially from the reported 0.05 m or GPR rises, the geometric-registration claim fails wherever map topology is unreliable. A second decisive check: relight a glossy vehicle under a high-frequency specular environment and measure PAE; the linear SH transfer predicts low PAE, and a large jump would falsify the photometric claim.

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

If this is right

  • Multi-agent interactive closed-loop testing becomes feasible on one GPU: 45 FPS with up to 50 vehicles and roughly linear VRAM growth.
  • E2E planners can be stress-tested across controlled difficulty levels; reported Driving Score and Success Rate degrade monotonically from Easy to Extreme, exposing planner-specific weaknesses.
  • Inserted vehicles inherit scene illumination and ground contact, so open-loop planner behavior on rendered frames approaches behavior on real frames (mADE 0.82 vs 0.76 real).
  • The compression, registration, and relighting modules are solver-agnostic and can be grafted onto other neural simulators to improve their speed and photometric accuracy.
  • Scene editing—moving, swapping, or re-lighting vehicles—needs no per-scene retraining.

Where Pith is reading between the lines

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

  • If reliable lane topology can be obtained for multi-level roads and overpasses, the same registration pipeline could extend to arbitrary infrastructure like parking garages and intersections without stop lines.
  • The linear SH relighting operator is a proxy; scenes with strong multi-bounce or high-frequency specular lighting would likely need additional terms, so a testable extension is measuring PAE on such scenes.
  • Because the framework is rigid-agent and camera-only, extending to pedestrians or LiDAR would require canonical deformation models and a range-sensing rasterizer—an avenue the paper lists as future work.
  • The difficulty-graded scenario protocol could serve as a reusable benchmark for comparing E2E planners' safety margins across simulation platforms.
Share X Bluesky LinkedIn Reddit HN

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 DecoupleGS, a 3D Gaussian Splatting-based closed-loop sensor simulator for end-to-end autonomous driving testing. The scene is decomposed into a persistent static background and object-centric canonical vehicle assets, with three targeted modules: semantic-aware asset compression (importance pruning + VQ), map-guided geometric registration (DTW + Procrustes + opacity-weighted vertical grounding), and proxy-based relighting (affine SH transfer + parametric contact shadows). The authors evaluate on nuScenes and PandaSet backgrounds with 3DRealCar assets, conduct open-loop and closed-loop experiments with UniAD/VAD, and claim a balanced fidelity-efficiency trade-off, improved metric and photometric consistency, and a practical interactive closed-loop E2E testing platform.

Significance. If the empirical claims hold under controlled evaluation, DecoupleGS would be a useful contribution to neural closed-loop simulation for end-to-end driving. The decoupled background/asset representation is natural and the three modules address genuinely important bottlenecks. The paper is also reproducible in spirit: the algorithms and hyperparameters are described in detail, and the supplement includes pseudocode and metric definitions. The ablation in Table 4 gives each module a clear role. However, the current evidence has two load-bearing gaps: the closed-loop comparison in Table 2 lacks a controlled protocol, and the ADE metric used for geometric consistency is essentially the same objective the registration module minimizes. These issues must be resolved before the central claims can be accepted.

major comments (4)
  1. [§4.5, Table 2] The central claim of a practical closed-loop E2E testing engine rests on Table 2, where DecoupleGS reports the highest Driving Score and Route Completion. The paper does not state whether HUGSIM, RealEngine, and OASim were evaluated under the same protocol: same E2E policy, same route/scenario set, same traffic-behavior engine, same vehicle dynamics, same infraction penalties, same camera configuration, and same rendering resolution. Driving Score and Route Completion are highly sensitive to these choices, and 'normalized to [0,1]' does not make different protocols comparable. The FPS comparison (45 vs 12/32/18) is likewise only meaningful if resolution, scene complexity, and hardware were fixed. Please provide a common protocol, or substantially narrow the claim to what can be supported.
  2. [Supp. Eq. (21), §3.3, Table 4] ADE is defined as the trajectory's distance to lane centerlines, and the map-guided registration module explicitly minimizes trajectory-to-lane distance via constrained DTW and SE(2) Procrustes. The reported improvement from 0.48 m to 0.05 m in Table 4 is therefore in large part a measure of how well the optimization objective was solved, not an independent evaluation of geometric consistency. This makes the geometric-fidelity claim circular. Please include an external metric (e.g., ground-truth pose error or multi-view reprojection error) or explicitly state that ADE is an objective-tracking measure rather than an external evaluation.
  3. [Tables 1, 2, 4] Fidelity, efficiency, geometric, and closed-loop results are reported as single values without variance, number of repetitions, or statistical significance. Given the small curated set (25 clips, 20 assets), statements such as 'consistently top-2' and the closed-loop ordering may not be robust. Table 3 already reports mean±std over 50 episodes; the same practice should be extended to Tables 1, 2, and 4, or the claims should be attenuated.
  4. [§3.3, §4.7] The map-guided registration and all geometry metrics inherit the accuracy of the input lane topology, whether from HD maps or MapTRv2. The authors acknowledge this limitation for overpasses, under-bridges, and poorly detected lanes, but the abstract and Section 4 use stronger language such as 'strictly align trajectories.' Since topology is an input assumption, the method does not guarantee strict alignment in general. Please either soften the wording or quantify the failure rate on the curated clips, especially where MapTRv2 is the topology source.
minor comments (4)
  1. [§3.2] The importance-score weights w_vis, w_col, w_ent are obtained by grid search and claimed to 'demonstrate robust generalization,' but no sensitivity analysis is shown. A small grid or a plot showing the effect of varying these weights on Table 1 would make the claim more credible.
  2. [§4.6, Fig. 9] The plug-and-play experiment in Fig. 9 is described only qualitatively. Please report the underlying numerical values for rendering speed and photometric accuracy when the modules are inserted into HUGSIM/OASim, so the reader can assess the magnitude of the claimed improvement.
  3. [§4.5, Table 3] Table 3 reports mean±std over 50 independent episodes, but it is not defined how the episodes are drawn for each difficulty level (e.g., initial states, route variations, random seeds). This is needed to interpret the variance.
  4. [References] Reference formatting is inconsistent in places (e.g., the DOI/URL style for [11] and [47]); please unify according to the venue style.

Circularity Check

1 steps flagged

ADE improvement is the registration module's own objective, making the geometric-fidelity claim partly circular; the central simulator claim is not reduced by construction.

specific steps
  1. fitted input called prediction [Supp. C.2, Eq. (21); Sec. 3.3 (Map-Guided Geometric Registration); Sec. 4.6, Table 4]
    "ADE= 1/T Σ_{t=1}^{T} min_{l∈L} ||p_t − l||_2 (Supp. Eq. 21); we build a cost matrix that penalizes both Euclidean distance and heading mismatch. We utilize constrained Dynamic Time Warping (DTW) to find the optimal warping path establishing trajectory-to-lane correspondences. Subsequently, an Orthogonal Procrustes analysis computes the globally optimal 2D rigid transform T_align ∈ SE(2) to eliminate systemic lateral drift. (Sec. 3.3); Without map-guided registration, ADE increases from 0.05 m to 0.48 m (Sec. 4.6)."

    The ADE metric is defined as the average distance from the agent trajectory to the lane centerlines L. The map-guided registration module is explicitly built to make trajectories coincide with those same centerlines: DTW establishes trajectory-to-lane correspondences and the SE(2) Procrustes transform eliminates lateral drift, i.e., minimizes the very quantity ADE measures. Thus Table 4's ablation ('w/o map-guided reg.' ADE 0.48 m vs. full framework ADE 0.05 m) is not independent evidence that geometric fidelity improved; it reports fit quality to the lane geometry that was the module's optimization target. The claim in Sec. 4.3 that registration 'significantly reduces trajectory drift' is therefore partly circular for the ADE component, although GPR, photometric, and closed-loop results a

full rationale

The only concrete circularity I can exhibit is the ADE-based geometric-fidelity claim. The map-guided registration module's DTW+Procrustes procedure minimizes trajectory-to-lane-centerline distance, and ADE (Supp. Eq. 21) is literally that distance; Table 4's 'w/o map-guided reg. ADE 0.48 m → 0.05 m' therefore reports the module's own objective as an independent metric. This is a genuine but localized circularity: it affects the 'improves metric consistency' contribution, while the photometric, efficiency, and closed-loop claims are not reduced by the same construction. I do not find load-bearing self-citation: the paper's reference list contains no self-citations by the authors, and the component citations (MapTRv2, 3DRealCar, IDM/MOBIL) are external and are used in a parameter-free or standard way. The closed-loop comparison in Table 2 is a serious protocol-comparability concern (different simulators, normalized scores, no stated common policy/scenario/dynamics control), but that is an experimental-validity issue, not circularity, and I do not score it as such. The paper's own Sec. 4.7 limitation—'map-guided registration relies on reliable semantic topology'—is acknowledged and scopes the claim, but it does not cure the ADE tautology. Overall score 5: one metric-level circularity; the central simulator claim still has independent, if protocol-confounded, evidence.

Axiom & Free-Parameter Ledger

8 free parameters · 6 axioms · 0 invented entities

The framework introduces no new physical or mathematical entity; the 'canonical asset library' and 'local ambient descriptor' are data structures, not postulated units. The claims rest mainly on standard 3DGS machinery plus three engineering assumptions: clean SAM-based background separation, reliable lane topology, and an affine SH lighting-transfer model. The affine relighting model and the lane-topology assumption are the two places where the method could fail outside the curated test set. Free parameters are mostly standard hyperparameters, but the relighting operator is a genuinely fitted global map, and the pruning/weighting hyperparameters are tuned on the evaluation distribution, which slightly inflates the fidelity numbers.

free parameters (8)
  • Importance weight w_vis = 0.5
    Grid-searched over {0.1,0.3,0.5,0.7,0.9} (Supp. A.1); tuned to balance fidelity, and the reported PSNR/LPIPS values are measured after this search.
  • Importance weight w_col = 0.3
    Grid-searched alongside w_vis and w_ent (Supp. A.1); part of the pruning objective that sets Table 1's fidelity numbers.
  • Importance weight w_ent = 0.2
    Grid-searched alongside w_vis and w_col (Supp. A.1).
  • Pruning threshold tau_s = 0.005
    Set empirically in Sec. 3.2; directly controls asset compactness and therefore the FPS/VRAM versus PSNR trade-off in Table 1.
  • DTW heading penalty lambda_heading = 2.5
    Empirically set in Supp. A.2; affects trajectory-to-lane alignment quality and hence the ADE and GPR metrics.
  • Relighting operator Theta = {W_T, w_0, W_B, b_0} = Fit by OLS over synthetic HDRI dataset
    Global affine SH-transfer map learned offline via least squares (Supp. B.2, Eq. 16); the photometric consistency claims depend on this calibration.
  • Codebook sizes K_color, K_shape = 1024 / 512
    Chosen hyperparameters that trade memory footprint against fidelity in Table 1.
  • Shadow and probe parameters (lambda, tau_decay, sigma_p, p) = not reported
    Contact shadow intensity, shadow softness, and probe spatial bandwidth in Eqs. 7, 9, and Supp. Eq. 17 are not given numeric values, limiting full reproduction.
axioms (6)
  • domain assumption Static background can be cleanly reconstructed from SAM-masked images with 5-pixel morphological dilation, leaving no floaters or shadow entanglements.
    Invoked in Sec. 4.1 dataset curation and visualized in Supp. Fig. 10; if masks leak or inpainting fails, background fidelity and relighting results degrade.
  • domain assumption Opacity-weighted depth accumulation (Eq. 6) yields an accurate local ground plane for vertical grounding.
    Sec. 3.3; replaces ray-casting for 3DGS and determines wheel contact, which feeds the GPR metric.
  • ad hoc to paper Ambient illumination transfer between background and assets is affine-linear in the local SH descriptor L(x) (Eq. 8).
    Sec. 3.4; the linear operator is fit offline, so any lighting interaction outside this affine form (multi-bounce, specular effects, cast shadows from the asset) is outside the model, as the authors concede in Sec. 4.7.
  • domain assumption Accurate lane topology is available (HD map or MapTRv2) for every scene where agents are inserted.
    Sec. 3.3; load-bearing for the ADE/GPR improvements and acknowledged as fragile in Sec. 4.7 for overpasses, under-bridge regions, and poor lane detection.
  • standard math Wigner D-matrix rotation of SH coefficients (Eq. 3) preserves correct view-dependent appearance under rigid motion.
    Sec. 3.1; standard group-theoretic result for spherical harmonics, not a new claim.
  • domain assumption IDM and MOBIL models produce realistic reactive background traffic for closed-loop evaluation.
    Supp. C.3; the interactivity and closed-loop driving-score results rest on these behavioral models reacting to the ego state.

reviewed 2026-08-04 · how reviews work

0 comments
Cite this review

Pith. "Pith review of DecoupleGS: Interactive 3D Gaussian Splatting for End-to-End Autonomous Driving Testing." pith.science (2026). https://pith.science/paper/SGXSS3H5

@misc{pith2026260801761,
  author       = {Pith},
  title        = {Pith review of: DecoupleGS: Interactive 3D Gaussian Splatting for End-to-End Autonomous Driving Testing},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/SGXSS3H5}},
  note         = {Machine review of arXiv:2608.01761}
}
Share X Bluesky LinkedIn Reddit HN
read the original abstract

End-to-end (E2E) autonomous driving algorithms require rigorous closed-loop validation in simulation environments offering high visual fidelity, strong interactivity, and real-time performance. Existing approaches, from game engines to static neural rendering, inherently trade off these requirements and struggle with the dynamic scene composition essential for E2E testing. To bridge this gap, we propose a novel decoupled 3D Gaussian Splatting (3DGS) framework tailored for large-scale E2E evaluation. We fundamentally decompose scenes into a high-fidelity static background and manipulable dynamic agents using an object-centric canonical representation. To resolve resulting representational conflicts, we introduce three targeted modules: (1) asset compression via perceptual pruning and vector quantization for real-time traffic rendering; (2) map-guided geometric registration leveraging semantic topology to strictly align trajectories; and (3) proxy-based relighting transferring ambient illumination for seamless photometric integration. Extensive experiments demonstrate that DecoupleGS achieves a balanced fidelity-efficiency trade-off, improves metric and photometric consistency, and provides a practical closed-loop sensor simulation platform for E2E autonomous driving evaluation.

Figures

Figures reproduced from arXiv: 2608.01761 by Haotian Shi, Jian Sun, Jie Sun, Siying Li, Ying Ni.

Figure 1
Figure 1. Figure 1: Comparison of autonomous driving simulation paradigms. [PITH_FULL_IMAGE:figures/full_fig_p002_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Overview of the DecoupleGS framework. DecoupleGS decouples persistent static backgrounds from canonical dynamic assets, and integrates asset compression, map-guided registration, proxy-based relighting, and unified rasterization to support interactive closed-loop E2E evaluation. 3 Method To enable large-scale closed-loop E2E evaluation, we aim to compose a photo￾realistic yet interactive simulation where (… view at source ↗
Figure 3
Figure 3. Figure 3: Illustration of the semantic-aware asset compression module. Canonical ve￾hicle Gaussians undergo explicit importance scoring (Vi, Di, Hi), adaptive pruning via threshold τs, and high-dimensional attribute vector quantization to form compact canonical volumes. Explicit Importance Scoring and Pruning. Instead of relying on black￾box neural networks, we compute an explicit scalar importance score si for each… view at source ↗
Figure 4
Figure 4. Figure 4: The map-guided geometric registration pipeline. It enforces strict metric align￾ment via trajectory-to-lane DTW sequence alignment, SE(2) Procrustes 2D transfor￾mation, and robust opacity-accumulated 3D vertical grounding. is driven entirely by the unconstrained relative pose updates from the planner to preserve maneuvers like lane changes. The sequence-level registration described below, and illustrated i… view at source ↗
Figure 5
Figure 5. Figure 5: Scalability evaluation. DecoupleGS avoids the OOM bottlenecks of Vanilla 3DGS, demonstrating stable VRAM growth and usable frame rates up to 50 agents. agents react online via coupled IDM and MOBIL behavioral models. The global scene state is then synchronized and re-rendered for the next step. 4.2 Rendering Efficiency and Scalability Asset compression [PITH_FULL_IMAGE:figures/full_fig_p011_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Quantitative comparisons of geometric registration and photometric realism against baseline methods. DecoupleGS achieves lower trajectory drift, fewer ground￾penetration artifacts, and more consistent relighting [PITH_FULL_IMAGE:figures/full_fig_p012_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: Qualitative comparison of the relighting module, demonstrating seamless inte￾gration of dynamic assets under various illuminations. combining trajectory-to-lane sequence alignment, SE(2) planar correction, and opacity-accumulated vertical grounding, DecoupleGS achieves accurate lane￾level placement while maintaining physically plausible contact with the road surface. Photometric realism. The same figure al… view at source ↗
Figure 8
Figure 8. Figure 8: Representative camera views of UniAD evaluated across generated scenarios of increasing difficulty, from Easy to Extreme [PITH_FULL_IMAGE:figures/full_fig_p014_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Plug-and-play capability. Integrating our modules into existing simulators im￾proves rendering speed and photometric accuracy. creases from 6.8 ◦ to 48.5 ◦ , exposing severe foreground-background illumination mismatch. The full framework therefore offers the best balanced trade-off across efficiency, geometric alignment, and photometric consistency. Plug-and-play capability. We further validate that the pr… view at source ↗
Figure 10
Figure 10. Figure 10: Surround-View Scene Decoupling. Top: Original reconstructed scene. Middle: Isolated static background with all dynamic agents perfectly removed. Bot￾tom: Extracted dynamic foreground agents rendered independently. The decoupling is strictly consistent across the entire six-camera suite. Vk rendered independently. Crucially, this decoupling maintains strict multi￾view consistency and preserves the high-fid… view at source ↗
Figure 11
Figure 11. Figure 11: Interactive Scene Editing. Top: Original front-camera view containing native traffic. Bottom: Edited scene where vehicle appearances and spatial positions have been explicitly manipulated. The modified assets seamlessly integrate with the background’s illumination and ground geometry. proxy-based relighting modules, the manipulated and newly inserted vehicles naturally inherit the geometric grounding and … view at source ↗

discussion (0)

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

Reference graph

Works this paper leans on

49 extracted references · 43 canonical work pages

  1. [1]

    In: IEEE Conf

    Barron, J.T., Mildenhall, B., Verbin, D., Srinivasan, P.P., Hedman, P.: Mip-NeRF 360: Unbounded anti-aliased neural radiance fields. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)

    work page 2022

  2. [2]

    Behrisch, M., Bieker, L., Erdmann, J., Krajzewicz, D.: SUMO: Simulation of urban mobility. In: Int. Conf. Advances Syst. Simul. (2011)

    work page 2011

  3. [3]

    In: IEEE Conf

    Caesar, H., Bankiti, V., Lang, A.H., Vora, S., Liong, V.E., Xu, Q., Krishnan, A., Pan, Y., Baldan, G., Beijbom, O.: nuScenes: A multimodal dataset for autonomous driving. In: IEEE Conf. Comput. Vis. Pattern Recog. (2020)

    work page 2020

  4. [4]

    In: Conf

    Cao, W., Hallgarten, M., Li, T., Dauner, D., Gu, X., Wang, C., Miron, Y., Aiello, M., Li, H., Gilitschenski, I., Ivanovic, B., Pavone, M., Geiger, A., Chitta, K.: Pseudo-simulation for autonomous driving. In: Conf. Robot Learn. Proceedings of Machine Learning Research, vol. 305, pp. 4709–4722. PMLR (2025)

    work page 2025

  5. [5]

    Chen, A., Xu, Z., Geiger, A., Yu, J., Su, H.: TensoRF: Tensorial radiance fields. In: Eur. Conf. Comput. Vis. (2022)

    work page 2022

  6. [6]

    IEEE Trans

    Chen, L., Wu, P., Chitta, K., Jaeger, B., Geiger, A., Li, H.: End-to-end autonomous driving: Challenges and frontiers. IEEE Trans. Pattern Anal. Mach. Intell.46(12), 10164–10183 (2024)

    work page 2024

  7. [7]

    IEEE Trans

    Chitta, K., Prakash, A., Jaeger, B., Yu, Z., Renz, K., Geiger, A.: Transfuser: Imi- tation with transformer-based sensor fusion for autonomous driving. IEEE Trans. Pattern Anal. Mach. Intell.45, 12878–12895 (2023)

    work page 2023

  8. [8]

    In: IEEE Int

    Codevilla, F., Müller, M., López, A., Koltun, V., Dosovitskiy, A.: End-to-end driv- ing via conditional imitation learning. In: IEEE Int. Conf. Robot. Autom. (2018)

    work page 2018

  9. [9]

    In: Conf

    Dosovitskiy, A., Ros, G., Codevilla, F., Lopez, A., Koltun, V.: CARLA: An open urban driving simulator. In: Conf. Robot Learn. (2017)

    work page 2017

  10. [10]

    Du, X., Wang, Y., Sun, H., Wu, Z., Sheng, H., Wang, S., Ying, J., Lu, M., Zhu, T., Zhan, K., Yu, X.: 3DRealCar: An in-the-wild RGB-D car dataset with 360-degree views. In: Int. Conf. Comput. Vis. pp. 26488–26498 (2025)

    work page 2025

  11. [11]

    Fan, Z., Wang, K., Wen, K., Zhu, Z., Xu, D., Wang, Z.: LightGaussian: Unbounded 3D gaussian compression with 15x reduction and 200+ FPS. In: Adv. Neural In- form. Process. Syst. (2024).https://doi.org/10.52202/079017-4447

    doi:10.52202/079017-4447 2024

  12. [12]

    In: IEEE Conf

    Fridovich-Keil, S., Yu, A., Tancik, M., Chen, Q., Recht, B., Kanazawa, A.: Plenox- els: Radiance fields without neural networks. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)

    work page 2022

  13. [13]

    In: IEEE Conf

    Fridovich-Keil, S., et al.: K-planes: Explicit radiance fields in space, time, and appearance. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)

    work page 2023

  14. [14]

    Ge, J., Liu, Z., Fan, L., Jiang, Y., Su, J., Li, Y., Zhang, Z., Chen, S.: Unraveling the effects of synthetic data on end-to-end autonomous driving. In: Int. Conf. Comput. Vis. pp. 28859–28869 (2025)

    work page 2025

  15. [15]

    arXiv preprint arXiv:2309.17080 (2023) DecoupleGS 17

    Hu, A., Russell, L., Yeo, H., et al.: GAIA-1: A generative world model for au- tonomous driving. arXiv preprint arXiv:2309.17080 (2023) DecoupleGS 17

    Pith/arXiv arXiv 2023

  16. [16]

    In: IEEE Conf

    Hu, Y., Yang, J., Chen, L., Li, K., Sima, C., Zhu, X., Chai, S., Du, S., Lin, T., Wang, W., et al.: Planning-oriented autonomous driving. In: IEEE Conf. Comput. Vis. Pattern Recog. (2023)

    work page 2023

  17. [17]

    Jiang,B.,Chen,S.,Xu,Q.,Liao,B.,Chen,J.,Zhou,H.,Zhang,Q.,Liu,W.,Huang, C., Wang, X.: VAD: Vectorized scene representation for efficient autonomous driv- ing. In: Int. Conf. Comput. Vis. (2023)

    work page 2023

  18. [18]

    arXiv preprint arXiv:2505.16902 (2025)

    Jiang, J., Song, N., Li, J., Zhu, X., Zhang, L.: RealEngine: Simulating autonomous driving in realistic context. arXiv preprint arXiv:2505.16902 (2025)

    arXiv 2025

  19. [19]

    ACM Trans

    Kerbl, B., Kopanas, G., Leimkühler, T., Drettakis, G.: 3D gaussian splatting for real-time radiance field rendering. ACM Trans. Graph.42(4) (2023)

    work page 2023

  20. [20]

    Transportation Research Record1999(1), 86–94 (2007)

    Kesting, A., Treiber, M., Helbing, D.: General lane-changing model MOBIL for car-following models. Transportation Research Record1999(1), 86–94 (2007)

    work page 2007

  21. [21]

    Kirillov, A., Mintun, E., Ravi, N., Mao, H., Rolland, C., Gustafson, L., Xiao, T., Whitehead, S., Berg, A.C., Lo, W.Y., Dollar, P., Girshick, R.: Segment anything. In: Int. Conf. Comput. Vis. pp. 4015–4026 (2023)

    work page 2023

  22. [22]

    In: IEEE Conf

    Kundu, A., et al.: Panoptic neural fields: A semantic object-aware neural scene representation. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)

    work page 2022

  23. [23]

    IEEE Trans

    Li, Q., Peng, Z., Feng, L., Zhang, Q., Xue, Z., Zhou, B.: Metadrive: Compos- ing diverse driving scenarios for generalizable reinforcement learning. IEEE Trans. Pattern Anal. Mach. Intell.45(3), 3461–3475 (2023)

    work page 2023

  24. [24]

    Liao, B., Chen, S., Zhang, Y., et al.: MapTRv2: An end-to-end framework for online vectorized HD map construction. Int. J. Comput. Vis.133(3), 1352–1374 (2025)

    work page 2025

  25. [25]

    arXiv preprint arXiv:2512.21618 (2025)

    Liu, Z., Fu, D., Cai, P., Wang, L., Liu, Y., Ren, Y., Shi, B., Wang, J.: SymDrive: Realistic and controllable driving simulator via symmetric auto-regressive online restoration. arXiv preprint arXiv:2512.21618 (2025)

    work page arXiv 2025

  26. [26]

    Ljungbergh, W., Tonderski, A., Johnander, J., Caesar, H., Åström, K., Felsberg, M., Petersson, C.: NeuroNCAP: Photorealistic closed-loop safety testing for au- tonomous driving. In: Eur. Conf. Comput. Vis. pp. 161–177 (2024)

    work page 2024

  27. [27]

    Mildenhall, B., Srinivasan, P.P., Tancik, M., Barron, J.T., Ramamoorthi, R., Ng, R.: NeRF: Representing scenes as neural radiance fields for view synthesis. In: Eur. Conf. Comput. Vis. (2020)

    work page 2020

  28. [28]

    ACM Trans

    Müller, T., Evans, A., Schied, C., Keller, A.: Instant neural graphics primitives with a multiresolution hash encoding. ACM Trans. Graph.41(4) (2022)

    work page 2022

  29. [29]

    In: IEEE Conf

    Ost, J., et al.: Neural scene graphs for dynamic scenes. In: IEEE Conf. Comput. Vis. Pattern Recog. (2021)

    work page 2021

  30. [30]

    In: IEEE Conf

    Rombach, R., Blattmann, A., Lorenz, D., Esser, P., Ommer, B.: High-resolution image synthesis with latent diffusion models. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)

    work page 2022

  31. [31]

    In: IEEE Intell

    Rong, G., Shin, B.H., Tabatabaee, H., Lu, Q., Lemke, S., Možeiko, M., Boise, E., Uhm, G., Gerow, M., Mehta, S., Agafonov, E., Kim, T.H., Sterner, E., Ushiroda, K., Reyes, M., Zelenkovsky, D., Kim, S.: LGSVL simulator: A high fidelity simu- lator for autonomous driving. In: IEEE Intell. Transp. Syst. Conf. pp. 1–6 (2020). https://doi.org/10.1109/ITSC45102....

    arXiv 2020

  32. [32]

    In: Field and Service Robotics (2018)

    Shah, S., Dey, D., Lovett, C., Kapoor, A.: AirSim: High-fidelity visual and physical simulation for autonomous vehicles. In: Field and Service Robotics (2018)

    work page 2018

  33. [33]

    IEEE Trans

    Tampuu, A., Matiisen, T., Semikin, M., Fishman, D., Muhammad, N.: A survey of end-to-end driving: Architectures and training methods. IEEE Trans. Neural Netw. Learn. Syst.33(4), 1364–1384 (2022)

    work page 2022

  34. [34]

    In: IEEE Conf

    Tancik, M., et al.: Block-nerf: Scalable large scene neural view synthesis. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022) 18 S. Li et al

    work page 2022

  35. [35]

    Physical Review E62(2), 1805–1824 (2000)

    Treiber, M., Hennecke, A., Helbing, D.: Congested traffic states in empirical obser- vations and microscopic simulations. Physical Review E62(2), 1805–1824 (2000)

    work page 2000

  36. [36]

    In: IEEE Conf

    Turki, H., et al.: Mega-nerf: Scalable construction of large-scale nerfs for virtual fly-throughs. In: IEEE Conf. Comput. Vis. Pattern Recog. (2022)

    work page 2022

  37. [37]

    Wang, X., Zhu, Z., Huang, G., Chen, X., Zhu, J., Lu, J.: DriveDreamer: Towards real-world-drive world models for autonomous driving. In: Eur. Conf. Comput. Vis. pp. 55–72 (2024).https://doi.org/10.1007/978-3-031-73195-2_4

    doi:10.1007/978-3-031-73195-2_4 2024

  38. [38]

    In: Artificial Intelligence

    Wu, Z., Liu, T., Luo, L., Zhong, Z., Chen, J., Xiao, H., Hou, C., Lou, H., Chen, Y., Yang, R., et al.: MARS: An instance-aware, modular and realistic simulator for autonomous driving. In: Artificial Intelligence. Lecture Notes in Computer Science, vol. 14473, pp. 3–15 (2023).https://doi.org/10.1007/978-981-99-8850-1_1

    work page doi:10.1007/978-981-99-8850-1_1 2023

  39. [39]

    In: IEEE Intell

    Xiao, P., Shao, Z., Hao, S., Zhang, Z., Chai, X., Jiao, J., Li, Z., Wu, J., Sun, K., Jiang, K., Wang, Y., Yang, D.: PandaSet: Advanced sensor suite dataset for autonomous driving. In: IEEE Intell. Transp. Syst. Conf. (2021)

    work page 2021

  40. [40]

    arXiv preprint arXiv:2402.03830 (2024)

    Yan, G., Pi, J., Guo, J., Luo, Z., Dou, M., Deng, N., Huang, Q., Fu, D., Wen, L., Cai, P., Gao, X., Cai, X., Zhang, B., Yang, X., Bai, Y., Zhou, H., Shi, B.: OASim: An open and adaptive simulator based on neural rendering for autonomous driving. arXiv preprint arXiv:2402.03830 (2024)

    Pith/arXiv arXiv 2024

  41. [41]

    In: IEEE Conf

    Yan, T., Wu, D., Han, W., Jiang, J., Zhou, X., Zhan, K., Xu, C.Z., Shen, J.: DrivingSphere: Building a high-fidelity 4D world for closed-loop simulation. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 27531–27541 (2025)

    work page 2025

  42. [42]

    Yan, Y., Lin, H., Zhou, C., Wang, W., Sun, H., Zhan, K., Lang, X., Zhou, X., Peng, S.: Street Gaussians: Modeling dynamic urban scenes with gaussian splatting. In: Eur. Conf. Comput. Vis. (2024)

    work page 2024

  43. [43]

    In: IEEE Conf

    Yan, Z., Low, W.F., Chen, Y., Lee, G.H.: Multi-scale 3D gaussian splatting for anti- aliased rendering. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 20923–20931 (2024)

    work page 2024

  44. [44]

    Yang,X.,Wen,L.,Wei,T.,Ma,Y.,Mei,J.,Li,X.,Lei,W.,Fu,D.,Cai,P.,Dou,M., He, L., Liu, Y., Shi, B., Qiao, Y.: DriveArena: A closed-loop generative simulation platform for autonomous driving. In: Int. Conf. Comput. Vis. pp. 26933–26943 (2025)

    work page 2025

  45. [45]

    In: IEEE Conf

    Yang, Z., Chen, Y., Wang, J., Manivasagam, S., Ma, W.C., Yang, A.J., Urtasun, R.: UniSim: A neural closed-loop sensor simulator. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 1389–1399 (2023)

    work page 2023

  46. [46]

    Yu, H.X., et al.: Unsupervised discovery of object radiance fields. In: Int. Conf. Learn. Represent. (2022)

    work page 2022

  47. [47]

    IEEE Trans

    Zhou, H., Lin, L., Wang, J., Lu, Y., Bai, D., Liu, B., Wang, Y., Geiger, A., Liao, Y.: HUGSIM: A real-time, photo-realistic and closed-loop simulator for autonomous driving. IEEE Trans. Pattern Anal. Mach. Intell.48(4), 4673–4691 (2026).https: //doi.org/10.1109/TPAMI.2025.3647952

    work page arXiv 2026

  48. [48]

    In: IEEE Conf

    Zhou, X., Lin, Z., Shan, X., Wang, Y., Sun, D., Yang, M.H.: DrivingGaussian: Composite gaussian splatting for surrounding dynamic autonomous driving scenes. In: IEEE Conf. Comput. Vis. Pattern Recog. pp. 21634–21643 (2024)

    work page 2024

  49. [49]

    floaters

    Zhu, X., Yi, R., Wen, X., Zhu, C., Xu, K.: Relighting scenes with object insertions in neural radiance fields. IEEE Trans. Circuit Syst. Video Technol.35(7), 6787– 6802 (2025) Supplementary Material 19 DecoupleGS: Interactive 3D Gaussian Splatting for End-to-End Autonomous Driving Testing Supplementary Material Overview This supplementary document provide...

    work page 2025

This paper was first reviewed by deepseek-v4-flash on August 4, 2026.