{"id":"b94a80e7-d09b-4be5-a3e3-e6cfcecdc35b","arxiv_id":"2608.01761","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"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.","lead":"The paper builds a simulation system that separates a fixed 3D Gaussian Splatting road scene from movable car models, so test drivers can insert, reposition, and re-light vehicles at interactive frame rates. It is aimed at stress-testing end-to-end autonomous driving policies in photorealistic closed-loop scenarios.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Closed-loop simulator comparison in Table 2 lacks a controlled protocol; the headline DS/RC advantage may be a protocol artifact.","rationale":"The reader's weakest_assumption centers on lane-topology reliance and SAM inpainting artifacts—valid scoped limitations. However, the most load-bearing gap is the uncontrolled closed-loop comparison in Table 2. The central claim is that DecoupleGS is a practical closed-loop E2E testing platform, and the quantitative evidence for that is the DS/RC advantage over other simulators. If those simulators were run in their native environments with different policies, routes, traffic behaviors, or scoring, the comparison is apples-to-oranges. The paper provides no protocol description for baselines in the main text or supplement; the supplementary details only DecoupleGS's own pipeline. This is a correctness risk that does not require questioning any internal consistency—it is an external validity gap. The ADE circularity and lack of error bars are secondary. The reader's conditional verdict is appropriate, but for a different, more fundamental reason: the closed-loop results need a controlled protocol before the central claim can be accepted. Hence I maintain CONDITIONAL/UNCHANGED rather than moving to ACCEPT or REJECT; the contribution may still be valid, but the key comparative evidence is currently unverified.","tokens_in":18896,"tokens_out":4315,"duration_ms":50722,"concrete_test":"Run a single controlled closed-loop benchmark: fix the E2E policy (e.g., UniAD), fix the scenario suite (e.g., the 25 curated clips with the same routes and traffic densities), fix the background-agent behavior engine (IDM/MOBIL), and fix the ego-vehicle dynamics and the scoring function. Then vary only the rendering/sensor backend: DecoupleGS vs. HUGSIM vs. RealEngine vs. OASim (or their output images fed to the same policy). If DecoupleGS no longer achieves the highest DS/RC and minTTC, the current comparative claim is a protocol artifact. If the ordering persists, the concern is resolved.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central empirical pillar for the 'practical closed-loop sensor simulation platform' claim is Table 2's closed-loop comparison, where DecoupleGS reports the highest Driving Score and Route Completion among HUGSIM, RealEngine, and OASim. The paper does not state whether all simulators were evaluated under a common protocol: same E2E policy, same scenario/route set, same traffic-behavior engine, same vehicle dynamics, same infraction definitions, and same rendering resolution/camera configuration. Each simulator has its own closed-loop stack, and Driving Score depends heavily on route length, traffic density, infraction penalties, and episode termination conditions. The note that scores are 'normalized to [0,1]' cannot make fundamentally different protocols comparable. Without a controlled benchmark, the reported ordering (Ours 0.884 vs. HUGSIM 0.765 vs. OASim 0.748) may reflect protocol differences rather than simulator quality. The FPS comparison (45 vs. 12/32/18) is similarly confounded unless resolution, scene complexity, and hardware are fixed. This is more load-bearing than the lane-topology limitation, which the authors explicitly scope, and more direct than the ADE circularity, because the headline claim of an interactive E2E testing engine rests on these numbers.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","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.","tokens_in":19141,"tokens_out":6834,"duration_ms":85119,"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":[{"comment":"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.","section":"§4.5, Table 2"},{"comment":"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.","section":"Supp. Eq. (21), §3.3, Table 4"},{"comment":"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.","section":"Tables 1, 2, 4"},{"comment":"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.","section":"§3.3, §4.7"}],"minor_comments":[{"comment":"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.","section":"§3.2"},{"comment":"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.","section":"§4.6, Fig. 9"},{"comment":"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.","section":"§4.5, Table 3"},{"comment":"Reference formatting is inconsistent in places (e.g., the DOI/URL style for [11] and [47]); please unify according to the venue style.","section":"References"}],"recommendation":"major_revision","confidential_remarks":"This is a promising systems paper, but the empirical core needs strengthening before the claims can be taken at face value. The uncontrolled Table 2 comparison and the circular ADE metric are the two issues I would insist on. I do not see the problems as irreparable within the manuscript's scope, so I recommend major revision rather than rejection. The editor may also want to ask the authors whether code will be released, since the paper emphasizes practical deployability."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short take: DecoupleGS is a solid systems paper that assembles known components — background/agent decoupling, VQ compression, DTW+Procrustes alignment, linear SH relighting — into a working interactive E2E driving simulator, and the authors are honest about its limits. The ablation is informative: removing compression drops FPS from 35 to 8.5, removing registration raises ADE from 0.05 to 0.48, removing relighting raises PAE from 6.8 to 48.5. The plug-and-play claim is plausible.\n\nThe soft spots are where the empirical claims overreach. The ADE metric (Supp. Eq. 21) is distance from trajectory to lane centerlines, and the registration module explicitly minimizes that distance with DTW+Procrustes. So Table 4's ADE improvement is partly by construction; it doesn't independently validate metric consistency. The same metric is used in Fig. 6. That doesn't break the photometric or efficiency results, but it means the geometric claim is weaker than it looks.\n\nBigger issue: Table 2's closed-loop comparison is not a controlled experiment. We have no evidence the baselines ran the same E2E policy, same scenario set, same behavioral engine (IDM/MOBIL is only described for DecoupleGS), same vehicle dynamics, same infraction definitions, or same rendering resolution/hardware. Scores are normalized to [0,1] but normalization can't fix protocol mismatch. The DS advantage (0.884 vs 0.765/0.748/0.682) and the FPS comparison may be protocol artifacts. This is load-bearing because the headline claim is being a practical closed-loop engine. The authors should provide a controlled benchmark, at minimum fixing the policy, route set, and hardware, or running baselines inside their own pipeline.\n\nReproducibility: no code release, single runs in Tables 1 and 2, and the importance weights were chosen by grid search; the paper doesn't say the grid was on a held-out set. That's a fitting risk. Table 3's mean±std over 50 episodes is better.\n\nThe authors deserve credit for the limitation section (Sec. 4.7): relighting is approximate, registration needs reliable lane topology, rigid agents only. Those are real scoping statements, not boilerplate.\n\nWho is this for? Researchers building neural simulators for E2E AD testing. It's a useful reference for module design and for seeing what a full stack looks like. I'd send it to reviewers, but the reviewers should demand a controlled closed-loop evaluation and a disentangled geometric metric before the headline claims are accepted.","headline":"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.","tokens_in":19777,"tokens_out":3118,"would_cite":true,"duration_ms":35664,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"DecoupleGS: a decoupled 3DGS simulator for interactive E2E driving testing.","keywords":["3D Gaussian Splatting","end-to-end autonomous driving","closed-loop simulation","sensor simulation","dynamic scene composition","asset compression","map-guided registration","relighting"],"falsifier":"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.","tokens_in":18638,"feed_emoji":"🚗","tokens_out":5587,"duration_ms":62897,"temperature":0.7,"pith_summary":"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.","feed_headline":"Decoupled 3DGS renders interactive driving sims at 45 FPS","feed_subtitle":"Splitting fixed scenes from re-litable vehicle assets closes the loop for testing end-to-end planners.","key_machinery":"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.","core_discovery":"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","pith_inferences":["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."],"forward_implications":["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."],"supporting_citations":[{"why":"Supplies the base 3D Gaussian Splatting representation and rasterizer that both background and canonical assets are built on.","marker":"[19]"},{"why":"Provides vectorized lane topology when HD maps are absent, feeding the map-guided registration module.","marker":"[24]"},{"why":"Masks dynamic objects in the source clips so the static background can be cleaned by inpainting.","marker":"[21]"},{"why":"Provides the background driving scenes and HD maps used for reconstruction and closed-loop evaluation.","marker":"[3]"},{"why":"Supplies the canonical vehicle assets that are compressed, relit, and inserted into the scenes.","marker":"[10]"},{"why":"One of the two end-to-end planners used in open-loop sim-to-real and closed-loop stress tests.","marker":"[16]"},{"why":"Second end-to-end planner used to measure behavioral consistency and closed-loop performance.","marker":"[17]"},{"why":"IDM longitudinal control model animates background agents reactively in the closed-loop protocol.","marker":"[35]"},{"why":"MOBIL lane-change model governs lateral maneuvers of background agents, enabling interactive traffic.","marker":"[20]"},{"why":"A closed-loop neural simulator baseline that DecoupleGS compares against for fidelity and interactive performance.","marker":"[47]"}],"fun_headline_variants":["DecoupleGS: interactive 3DGS for end-to-end driving tests","Decoupled 3DGS: real-time interactive sims for E2E driving","DecoupleGS splits static and dynamic for real-time E2E testing","Interactive 3DGS: decouple scenes for high-fidelity E2E sims"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"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.","fun_headline_variants_meta":{"raw":{"variants":["DecoupleGS: interactive 3DGS for end-to-end driving tests","Decoupled 3DGS: real-time interactive sims for E2E driving","DecoupleGS splits static and dynamic for real-time E2E testing","Interactive 3DGS: decouple scenes for high-fidelity E2E sims"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000869,"raw_usage":{"total_tokens":3586,"prompt_tokens":712,"completion_tokens":2874,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":456,"completion_tokens_details":{"reasoning_tokens":2787}},"tokens_in":456,"tokens_out":2874,"duration_ms":23525,"temperature":1.0,"reasoning_tokens":2787,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T21:29:38.646525+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"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.","supporting_citations":[{"cited_title":"ACM Trans","cited_arxiv_id":null,"evidence_quote":"Supplies the base 3D Gaussian Splatting representation and rasterizer that both background and canonical assets are built on."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides vectorized lane topology when HD maps are absent, feeding the map-guided registration module."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Masks dynamic objects in the source clips so the static background can be cleaned by inpainting."},{"cited_title":"In: IEEE Conf","cited_arxiv_id":null,"evidence_quote":"Provides the background driving scenes and HD maps used for reconstruction and closed-loop evaluation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the canonical vehicle assets that are compressed, relit, and inserted into the scenes."},{"cited_title":"In: IEEE Conf","cited_arxiv_id":null,"evidence_quote":"One of the two end-to-end planners used in open-loop sim-to-real and closed-loop stress tests."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Second end-to-end planner used to measure behavioral consistency and closed-loop performance."},{"cited_title":"Physical Review E62(2), 1805–1824 (2000)","cited_arxiv_id":null,"evidence_quote":"IDM longitudinal control model animates background agents reactively in the closed-loop protocol."},{"cited_title":"Transportation Research Record1999(1), 86–94 (2007)","cited_arxiv_id":null,"evidence_quote":"MOBIL lane-change model governs lateral maneuvers of background agents, enabling interactive traffic."},{"cited_title":"IEEE Trans","cited_arxiv_id":null,"evidence_quote":"A closed-loop neural simulator baseline that DecoupleGS compares against for fidelity and interactive performance."}],"review_version":1}