{"id":"d8e37a9c-4438-4f83-ab14-657a8caa1e83","arxiv_id":"2506.18885","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":12,"one_line_summary":"GRAND-SLAM combines local submap optimization with inter- and intra-robot loop closure to produce globally consistent multi-agent Gaussian splatting maps in large-scale outdoor environments.","lead":"This paper presents a multi-robot mapping system that builds shared 3D maps from color and depth cameras and keeps the maps consistent when robots meet again. It is the first to show Gaussian splatting SLAM working across multiple robots in large outdoor spaces.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The global-consistency claim relies on the untested assumption that each submap is internally drift-free, so the rigid transform in Eq. 15 cannot correct intra-submap distortion.","rationale":"The central claim is about global consistency at scale, not just tracking accuracy. The paper's only mechanism for correcting drift after loop closure is pose-graph optimization over submap origins followed by rigid application of the optimized origin pose to all Gaussians (Eq. 15). Therefore the correctness of the final map depends on the internal rigidity of each submap. The paper gives no evidence for this. The local submap optimization (Section III-B.3) optimizes Gaussians against keyframe poses that come from the tracking module; if those poses contain gradual drift, the Gaussian map is fit to a drifted trajectory, so the submap itself becomes non-rigidly distorted. Because all metrics in Section IV are either global trajectory errors (ATE) or training-view rendering metrics, they do not detect this failure mode. This is not a disagreement with the field's consensus; it is an internal gap between the stated global-consistency claim and the rigid-body mechanism used to implement it. I agree with the reader's weakest_assumption on this point. The proposed check—comparing intra-submap relative error with inter-submap ICP residuals—would settle whether the assumption actually holds on the Kimera-Multi data. If it fails, the 'globally consistent' claim should be weakened to 'globally aligned at submap granularity.' The reader's CONDITIONAL verdict already captures this, so no verdict change is needed.","tokens_in":13266,"tokens_out":7644,"duration_ms":84948,"concrete_test":"On the Kimera-Multi Outside-1/Outside-2 sequences, compute, for each submap after global optimization, the relative pose error between the first and last keyframe inside the submap, and compare it with the inter-submap loop-closure ICP residual used to accept the adjacent loop closure (Eq. 8 with thresholds tau_f and tau_rho). If the intra-submap relative RMSE is comparable to or larger than the inter-submap alignment RMSE, the single-rigid-transform assumption in Eq. 15 is violated and the global map contains internal distortion. Report this as a new table alongside Table III.","verdict_should_be":"UNCHANGED","load_bearing_attack":"GRAND-SLAM's central claim is globally consistent large-scale multi-agent mapping via loop closures. The mechanism that turns optimized pose-graph results into a global map is Eq. 15, which applies one SE(3) transform to every Gaussian in a submap. This is valid only if each submap is internally consistent after the local optimization in Section III-B.3. That optimization fits Gaussian parameters to keyframe poses supplied by the tracking module; if those poses have drifted within a submap, the optimizer will bake the drift into the Gaussian positions rather than remove it. Nothing in Section IV measures intra-submap consistency: ATE is global and can improve even when submap interiors are warped, and the rendering tables report training-view PSNR/SSIM/LPIPS, which do not expose cross-submap discontinuities. Without an intra-submap consistency check, the paper's headline claim that GRAND-SLAM produces globally consistent maps is supported only by an untested rigidity assumption.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"GRAND-SLAM proposes a multi-agent RGB-D 3D Gaussian Splatting SLAM system built from submap-based local optimization, NetVLAD keyframe retrieval, ICP-based loop closure registration, and pose-graph optimization that updates the global map by rigidly transforming submaps. The paper reports state-of-the-art tracking and rendering results on the Multiagent Replica indoor dataset and the Kimera-Multi Outdoor dataset, with the abstract claiming 28% higher PSNR than existing methods on Replica and 91% lower multi-agent tracking error on Kimera-Multi. It includes ablations with and without loop closure and comparisons against ORB-SLAM3, Gaussian-SLAM, MonoGS, MAGiC-SLAM, CP-SLAM, Swarm-SLAM, and CCM-SLAM, together with qualitative renderings on the outdoor dataset.","tokens_in":13423,"tokens_out":6678,"duration_ms":66326,"significance":"If the central claims hold, the paper is a useful step: it is among the first attempts at multi-agent RGB-D Gaussian splatting SLAM that explicitly integrates inter- and intra-robot loop closures, and it demonstrates the approach on a large-scale outdoor benchmark. The method is clearly motivated, uses an explicit representation that supports rigid-body map correction, and the ablation study isolates the contribution of loop closure. However, the evidence does not yet establish the headline property of global consistency: the submap-rigidity assumption at the core of the map-update step is untested, the rendering evaluation is limited to training views, and some headline averages mix complete and partial baseline runs. With additional consistency metrics and cleaner statistical comparisons, the contribution would be significant for the SLAM and 3D reconstruction communities.","major_comments":[{"comment":"The global-consistency claim rests on the assumption that every submap is internally drift-free. Section III-B.3 optimizes Gaussian parameters against keyframe poses supplied by the tracking module; if those poses drifted within a submap, the optimizer can bake that drift into Gaussian positions rather than remove it. Equation (15) then applies one SE(3) transform to all Gaussians in a submap, so it cannot correct intra-submap distortion. The paper does not measure intra-submap consistency anywhere: the ATE numbers are global trajectory metrics, and the rendering tables report training-view PSNR/SSIM/LPIPS, which do not expose discontinuities at submap boundaries. Please add a quantitative intra-submap or cross-boundary consistency evaluation (for example, alignment error between overlapping submap surfaces after applying Eq. 15) or an analysis showing that the local submap optimization removes pose drift.","section":"Section III-E, Eq. (15); Section III-B.3"},{"comment":"The headline numerical claims are weakened by the inclusion of partial runs in the averages. MAGiC-SLAM and Gaussian-SLAM are reported as failing partway through the Outside 2 Agent 2 traverse, yet their entries for that run (10.50 m and 7.66 m in Table III) are included in the per-method averages, and the abstract's '91% lower multi-agent tracking error' is computed from those averages. Re-run the baselines to completion with robust failure handling, or report only complete runs and separate failed runs explicitly. In addition, all tables report single-run metrics without error bars or significance tests; the state-of-the-art claims would be much stronger with variance estimates over repeated runs or multiple dataset splits.","section":"Section IV-B, Tables III and IV"},{"comment":"Rendering quality is evaluated only on training views. Training-view PSNR can be high even when the global map is inconsistent, because each submap is fit to its own observed frames and the metric never requires rendering across submap boundaries. To support the claim of globally consistent photorealistic maps, report novel-view synthesis metrics on held-out frames and rendering quality at submap overlap boundaries, where misalignment would appear as ghosting or duplication.","section":"Section IV-C, Tables IV and V"},{"comment":"The pose-graph formulation is underspecified. Equation (9) places loop-closure constraints between submap-origin transforms T_g_{a,l}, but the graph in Eq. (12) is defined over keyframe poses {T_i} with tracking constraints and loop-closure constraints, and Eq. (13) solves for poses {T*_i}. The paper does not state how the optimized submap transform T_g_{a,l} used in Eqs. (14)-(15) is recovered from the optimized keyframe poses, nor whether tracking constraints inside a submap participate in the global optimization. This matters because Eq. (15) transfers the pose-graph result to every Gaussian; please clarify the node set, the state variables, and the exact extraction of T_g_{a,l}.","section":"Section III-E, Eqs. (9)-(15)"}],"minor_comments":[{"comment":"The baseline is labeled 'Gaussian SLAM [12]', but reference [12] is ESLAM; the intended citation is Gaussian-SLAM [17].","section":"Table III"},{"comment":"There is a typo: 'SW ARM-SLAM' should be 'Swarm-SLAM'.","section":"Section IV-A.3"},{"comment":"The symbol dmax is used both for the submap-split translation threshold and for the ICP distance threshold in Eq. (8); please use distinct names to avoid ambiguity.","section":"Section III-B.1 and Section III-D"},{"comment":"The caption is hard to parse ('this example demonstrates the renders of a scene ... optimization (left) and after rotating with respect to the origin (right)'); please rephrase it to describe the visual comparison more clearly.","section":"Figure 3"},{"comment":"The implementation section mentions two RTX 3090 GPUs but reports no runtime, memory, or communication cost; a scalability claim would benefit from timing or bandwidth measurements.","section":"Section IV-A.2"},{"comment":"For the Multiagent Replica rendering comparison, CP-SLAM reports very low PSNR and high depth L1 on some sequences (e.g., A-1); please state whether CP-SLAM completed all sequences or whether those entries are partial or failed runs.","section":"Table V"}],"recommendation":"major_revision","confidential_remarks":"The 'first' claim in Section II-E should be treated carefully: 'large-scale' is supported only by a 1.85 km subset of Kimera-Multi, and the comparison set of multi-agent Gaussian SLAM systems is small. Also, the paper does not currently release code, which makes it difficult to verify the handling of baseline failures and the exact pose-graph extraction described in the major comments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague— What you should know: this is the first paper I've seen that takes multi-agent 3D Gaussian splatting SLAM outside the lab. The system combines submap-based local optimization, NetVLAD place recognition, ICP refinement, and pose graph optimization to close loops across and within robots, and it reports results on Kimera-Multi outdoor, where the leading multi-agent Gaussian baseline (MAGiC-SLAM) largely fails. If it works as claimed, that's a genuine system-level advance, and the local-frame tracking idea is a sensible way to avoid gradient imbalance far from the origin. The novelty claim looks believable: MAGiC-SLAM is indoor-only, and I don't know of another outdoor multi-agent Gaussian SLAM with loop closures. The paper is worth a serious referee.\n\nNow the soft spots, in order. The global consistency claim rests on a rigidity assumption that is never tested: each submap is internally drift-free, so Eq. 15 can rigidly transform every Gaussian by the optimized origin pose. If local optimization bakes drift into Gaussian positions—which is what fitting Gaussian parameters to drifted keyframe poses would do—loop closures correct submap origins but not interior warping. The paper gives no intra-submap consistency metric. ATE is global and can improve even with warped interiors; training-view PSNR/SSIM/LPIPS don't expose cross-submap discontinuities. This is the load-bearing gap, and it is fixable: report per-submap consistency or alignment error before and after pose graph optimization.\n\nThe evaluation has smaller but real problems. All tables are single runs, no error bars or significance tests. On Kimera-Multi, Gaussian-SLAM and MAGiC-SLAM fail partway through one traverse; the averages include partial runs (marked with asterisks, but still averaged in), which flatters the comparison. The rendering tables report training-view metrics, not novel-view synthesis, so the photorealistic mapping claim is weaker than it looks. No code or data are released, and the many thresholds—d_max, theta_max, tau_sim, ICP thresholds, information matrices—are not analyzed for sensitivity. None of this is a circularity problem; the measured quantities are benchmark outcomes, not fits to those outcomes. But reproducibility and robustness are genuinely open.\n\nOn the plus side, the architecture is coherent, the baselines are the right ones, and the system actually runs in the regime the paper claims is new. The central claim is plausible. I just would not take 'globally consistent' as established until the intra-submap consistency and full-sequence metrics are added and code is released.\n\nRecommendation: send to peer review, but expect major revisions. This is a solid systems contribution in need of stronger evidence, not a paper to desk reject.","headline":"A genuinely new multi-agent Gaussian SLAM system for large outdoor scenes, with a plausible architecture but evaluation gaps—no code, single runs, and an untested rigidity assumption about submaps—so it deserves serious review rather than desk rejection.","tokens_in":14053,"tokens_out":4526,"would_cite":true,"duration_ms":44579,"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":"GRAND-SLAM claims to be the first large-scale RGB-D Gaussian-splatting SLAM system that integrates inter- and intra-robot loop closures, reporting state-of-the-art tracking and rendering in indoor and outdoor multi-agent settings.","keywords":["multi-agent SLAM","3D Gaussian splatting","RGB-D SLAM","loop closure","pose graph optimization","submaps","photorealistic reconstruction","outdoor mapping"],"falsifier":"Run a single-agent outdoor sequence that revisits the same place inside one submap, apply a correct loop closure, and compare the rendered geometry at the revisit before and after pose-graph optimization: visible residual misalignment inside that submap would show that the rigid-submap assumption does not hold in practice. Alternatively, measure the submap's internal color-depth render consistency on a long traverse; if it degrades with distance even when loop-closure alignments at submap boundaries are accurate, the central premise is weakened.","tokens_in":12976,"feed_emoji":"🗺️","tokens_out":9812,"duration_ms":91752,"temperature":0.7,"pith_summary":"GRAND-SLAM is a collaborative SLAM system built on 3D Gaussian splatting. Each robot keeps its map in small local pieces called submaps, made of tiny 3D Gaussian blobs, and a server finds overlaps between different robots' submaps and between separate visits by the same robot. Those overlaps become constraints in a pose-graph optimization, and after optimization each submap is moved rigidly into one global frame. The paper's central claim is that this is the first multi-agent RGB-D Gaussian-splatting SLAM approach to scalably integrate inter- and intra-robot loop closure, and that it achieves state-of-the-art tracking, 28% higher PSNR than existing methods on the indoor Replica set, and 91% lower multi-agent tracking error than existing multi-agent methods on the large-scale outdoor Kimera-Multi set. If the claim holds, robot teams with only color-plus-depth cameras could rapidly build photorealistic, globally consistent maps of large outdoor areas.","feed_headline":"Robot teams map outdoor scenes with loop-closed Gaussian splats","feed_subtitle":"The first Gaussian-splatting SLAM to close loops between robots reports a 91 percent cut in tracking error.","key_machinery":"The central object is the submap: a bounded collection of anisotropic 3D Gaussians (small 3D blobs whose shape can differ along each axis), each with mean, covariance, opacity, and color, defined relative to a local frame whose first keyframe is the identity. The key identity is Eq. 15, which after pose-graph optimization applies one rigid transform to the whole submap: $\\boldsymbol{\\mu}^{(g)} = R\\boldsymbol{\\mu}^{(l)} + t$ and $\\Sigma^{(g)} = R\\Sigma^{(l)}R^{\\top}$. This machinery is what makes large-scale multi-agent mapping tractable: submaps stay small enough to optimize locally, loop closures add only relative pose constraints between submap origins, and global consistency is achieved by moving whole submaps rather than re-optimizing every Gaussian.","core_discovery":"The paper's central claim is stated directly: GRAND-SLAM is the first large-scale RGB-D Gaussian splatting SLAM approach that leverages inter- and intra-robot loop closure for drift reduction. The system builds a submap from an RGB-D stream, optimizes that submap locally against color and depth renderings, detects loop candidates by matching keyframe descriptors, refines each candidate with dense geometric alignment, and solves a pose graph that jointly corrects all agent trajectories. After the pose-graph solve, every Gaussian in a submap is transformed by the same rigid 3D pose, which converts global map consistency into a discrete pose problem instead of a per-Gaussian optimization. On the outdoor Kimera-Multi dataset the paper reports an average trajectory error of 4.99 m versus 60.79 m for the strongest prior multi-agent Gaussian method, and on the indoor Multiagent Replica set it reports 0.25 cm average error with 41.35 dB PSNR, about 28% higher than the next method.","pith_inferences":["The rigid-submap step implies that residual drift inside a submap can never be repaired by submap-origin corrections; a natural extension is to split or non-rigidly deform a submap when internal misalignment is detected.","The design suggests an ablation the paper does not run: sweeping the loop-closure acceptance thresholds should reveal the point at which inter-agent constraints stop improving and start biasing the global map.","If the local-optimization trick is the source of the gain, it could transfer to other dense SLAM representations, since world-frame rotation optimization suffers from the same gradient imbalance in any representation."],"forward_implications":["Teams of robots using only RGB-D cameras can produce photorealistic, globally consistent maps of kilometer-scale outdoor areas without LiDAR seeding.","Loop closures enter the graph only after passing fitness and RMSE quality gates, so the global optimization should be robust to false-positive place matches.","Because each submap is optimized locally and moved rigidly, per-agent computation stays bounded as the environment grows; adding agents adds constraints, not map-wide re-optimization.","The same system covers both indoor and outdoor regimes, matching or beating prior multi-agent methods on small-scale scenes while generalizing to large-scale ones."],"supporting_citations":[{"why":"supplies the explicit 3D Gaussian splatting scene representation and differentiable rasterizer used for the whole mapping pipeline.","marker":"[15]"},{"why":"introduces submap-based RGB-D Gaussian SLAM, whose submap construction the paper adapts and extends.","marker":"[17]"},{"why":"is the leading multi-agent Gaussian SLAM baseline that GRAND-SLAM compares against and must outperform in tracking and rendering.","marker":"[38]"},{"why":"provides the large-scale real-world outdoor RGB-D multi-agent dataset used to demonstrate scalability.","marker":"[43]"},{"why":"supplies the keyframe descriptor used for loop-closure candidate retrieval in place recognition.","marker":"[44]"},{"why":"provides the factor-graph pose-graph optimization solver used to compute globally consistent submap poses.","marker":"[45]"}],"fun_headline_variants":["First multi-robot Gaussian SLAM for large outdoor maps","Robot teams cut tracking error 91% with loop-closed Gaussian maps","Outdoor multi-agent SLAM: loop closures reduce error by 91%","GRAND-SLAM: first Gaussian splat SLAM to close loops across robots","Multi-robot Gaussian SLAM hits 91% lower error outdoors"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that each submap is internally drift-free after local optimization, so a single rigid transform applied to the whole submap is enough to make the global map consistent; if local tracking leaves distortion inside a submap, correct loop closures still leave the final map misaligned.","fun_headline_variants_meta":{"raw":{"variants":["First multi-robot Gaussian SLAM for large outdoor maps","Robot teams cut tracking error 91% with loop-closed Gaussian maps","Outdoor multi-agent SLAM: loop closures reduce error by 91%","GRAND-SLAM: first Gaussian splat SLAM to close loops across robots","Multi-robot Gaussian SLAM hits 91% lower error outdoors"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000774,"raw_usage":{"total_tokens":3420,"prompt_tokens":938,"completion_tokens":2482,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":2386}},"tokens_in":554,"tokens_out":2482,"duration_ms":16974,"temperature":1.0,"reasoning_tokens":2386,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T18:41:21.434162+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a single-agent outdoor sequence that revisits the same place inside one submap, apply a correct loop closure, and compare the rendered geometry at the revisit before and after pose-graph optimization: visible residual misalignment inside that submap would show that the rigid-submap assumption does not hold in practice. Alternatively, measure the submap's internal color-depth render consistency on a long traverse; if it degrades with distance even when loop-closure alignments at submap boundaries are accurate, the central premise is weakened.","supporting_citations":[{"cited_title":"3d Gaussian splatting for real-time radiance field rendering","cited_arxiv_id":null,"evidence_quote":"supplies the explicit 3D Gaussian splatting scene representation and differentiable rasterizer used for the whole mapping pipeline."},{"cited_title":"Magic-slam: Multi-agent gaussian globally consistent slam,","cited_arxiv_id":null,"evidence_quote":"is the leading multi-agent Gaussian SLAM baseline that GRAND-SLAM compares against and must outperform in tracking and rendering."},{"cited_title":"Resilient and distributed multi-robot visual SLAM: Datasets, experiments, and lessons learned,","cited_arxiv_id":null,"evidence_quote":"provides the large-scale real-world outdoor RGB-D multi-agent dataset used to demonstrate scalability."},{"cited_title":"Netvlad: Cnn architecture for weakly supervised place recognition,","cited_arxiv_id":null,"evidence_quote":"supplies the keyframe descriptor used for loop-closure candidate retrieval in place recognition."},{"cited_title":"Factor graphs and GTSAM: a hands-on introduction,","cited_arxiv_id":null,"evidence_quote":"provides the factor-graph pose-graph optimization solver used to compute globally consistent submap poses."}],"review_version":2}