{"id":"7c4c8577-3bca-4ffa-a621-70faadc9b2d4","arxiv_id":"2412.02249","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":11,"one_line_summary":"A centralized multi-robot planner uses semantic segmentation and 3DGS surface uncertainty to generate reconstruction tasks, plus a global-local mode and improved K-means assignment to coordinate robots.","lead":"The paper combines 3D Gaussian splatting, open-vocabulary semantic segmentation, and a hierarchical task assignment scheme so several robots can jointly reconstruct an indoor scene. The authors report higher reconstruction quality and lower planning time than two existing multi-robot methods in simulated scenes.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Objectness score used as completeness proxy is uncalibrated; the semantic gating that drives reconstruction task generation may misdirect planning and undermine the reported gains.","rationale":"I read the paper as a systems-integration contribution: combine Gaussian-SLAM/ASH, OVIR-3D, uncertainty-cached 3DGS, and hierarchical multi-robot planning. Given that framing, the strongest claim (best quality among planning methods, superior planning efficiency among multi-robot methods) rests on the semantic-guided reconstruction task generation being the right signal to route viewpoints toward incomplete objects. My main concern is not the lack of code or error bars (important but secondary) but a missing calibration link: the objectness score computed as a softmax over ScanNet200 logits is a classification confidence, and the paper provides no evidence that this confidence correlates with geometric incompleteness of the reconstructed instance. The fixed thresholds Cmin/Cmax and the temperature lambda_e=50 magnify this problem: softmax confidence is highly sensitive to the choice of lambda_e and to the relative CLIP score distribution, which can vary across object categories and scenes. Because V3 (surface-only) in Table I is markedly worse than V5/V6, the reported gains are contingent on the semantic gate selecting the right instances. This is precisely the reader's weakest assumption, so I agree. I propose a concrete test using the sim environments' ground-truth meshes to directly measure geometric completeness per instance and compare with C_1^i, plus a threshold sweep. If the test shows no monotonic relation, the central claim should be revised to a more modest, scene-specific claim. I recommend keeping the CONDITIONAL verdict: the work is promising and internally coherent, but this calibration gap must be closed before the headline claims can be accepted.","tokens_in":12405,"tokens_out":7019,"duration_ms":76048,"concrete_test":"In the QaLdn, Nicut, and Oyens sim scenes, compute the proposed objectness score C_1^i for every OVIR-3D instance during the online reconstruction, and compare it to a direct geometric completeness measure for the same instance, e.g., the fraction of ground-truth mesh points of that object within 5 cm of the reconstructed surface (or reconstructed-point coverage of the object's mesh). Then sweep Cmin in {0.1, 0.2, 0.3} and Cmax in {0.5, 0.6, 0.7} and re-run the full pipeline for each setting, reporting PSNR/Acc/Comp/Recall and TGP. If C_1^i does not monotonically track geometric completeness, or if the best threshold set varies across scenes by more than 0.1, the semantic gating is not a reliable driver of the reported gains and the fixed thresholds in Sec. IV-A.2 need to be justified or replaced.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central novelty is semantic-guided reconstruction task generation (Sec. III-C, Algorithm 1). In 'Reconstruction instance generation', the objectness score C_1^i is the maximum softmax probability over 200 ScanNet200 classes after scaling logits by lambda_e=50. This score is used as a proxy for reconstruction completeness: instances with C_1^i > Cmax=0.6 are treated as already well reconstructed and excluded; those with C_1^i < Cmin=0.2 are treated as CLIP noise and excluded. But the maximum softmax probability measures classification confidence of the CLIP feature, not geometric completeness of the 3D instance. A partially scanned chair back can produce a highly confident 'chair' label; a fully scanned low-texture object, such as a uniformly colored wall, may yield an ambiguous, low-confidence label. The thresholds are fixed across three scenes with no sensitivity analysis, and the paper provides no calibration evidence linking C_1^i to reconstruction completeness. Since Table I shows that switching from surface-uncertainty-only (V3) to semantic guidance (V5/V6) is what produces the quality improvements, the entire claimed advantage depends on this unvalidated semantic completeness signal. If the gating misclassifies instances, reconstruction tasks are generated for already-complete regions or withheld from incomplete ones, and the method likely degrades to exploration-only or surface-only behavior, losing both quality and planning-efficiency gains.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a centralized multi-robot autonomous 3D reconstruction framework based on 3D Gaussian splatting (3DGS), with three main components: perception (3DGS, ASH geometry, OVIR-3D open-vocabulary segmentation), semantic-guided reconstruction task generation, and hierarchical planning with global-local mode assignment and an improved K-means task clustering. The central claims are that integrating online open-vocabulary semantic segmentation with 3DGS surface uncertainty focuses view sampling on incomplete objects, and that the proposed mode/task assignment achieves both high reconstruction quality and low planning time. Experiments are conducted in three simulated indoor scenes (QaLdn, Nicut, Oyens) with four robots, comparing against two multi-robot baselines (ACAMS and MS3DSR) and several ablations (V1-V6), plus a qualitative real-robot deployment.","tokens_in":12701,"tokens_out":3547,"duration_ms":40817,"significance":"If the results hold, the paper makes a useful contribution by demonstrating that semantic segmentation can be combined with 3DGS uncertainty in a multi-robot planning loop to avoid local minima and improve reconstruction quality without expensive point-cloud completion or model databases. The systematic ablation (V1-V6) is a strength, and the comparison against two existing multi-robot methods is informative. The real-robot demonstration, while qualitative, adds credibility. The main caveat is that the semantic completeness signal (the objectness score) is used as a load-bearing gating mechanism without calibration or sensitivity analysis, so the validity of the reported gains depends on an unvalidated proxy. If that issue is resolved, the paper would likely be of interest to the robotics and reconstruction communities.","major_comments":[{"comment":"The objectness score C_1^i, defined as the maximum softmax probability over ScanNet200 classes after scaling by lambda_e=50, is used as a direct proxy for reconstruction completeness: instances with C_1^i>Cmax=0.6 are treated as already well reconstructed, and those below Cmin=0.2 as noise. This is a key load-bearing assumption because Table I shows that switching from surface-uncertainty-only (V3) to semantic guidance (V4-V6) produces substantial quality gains. However, no calibration evidence is provided linking classification confidence to geometric completeness. A partially scanned object can have a high-confidence label, while a fully scanned low-texture object may have ambiguous, low-confidence CLIP features. The fixed thresholds are used across all three scenes without sensitivity analysis. Please provide an evaluation of how C_1^i correlates with actual completeness (e.g., using ground-truth geometry in the simulated scenes), or at least a threshold sweep (e.g., varying Cmin and Cmax) to show the results are not artifacts of the chosen values.","section":"Section IV-A.2, Section IV-B"},{"comment":"The experimental protocol appears to involve a single run per configuration, with no reported variance, multiple seeds, or statistical testing. Since several comparisons (e.g., V5 versus V6 in QaLdn: PSNR 21.89 vs 22.75; Nicut: 22.84 vs 24.28; Oyens: 23.35 vs 24.81) are clear, but other differences (e.g., Acc values) are small, the lack of error bars makes it difficult to assess whether the improvements are significant. Please report means and standard deviations over at least three independent runs, or justify why variance is negligible.","section":"Section IV-B, Table I"},{"comment":"The efficiency claim \"superior planning efficiency compared to existing multi-robot methods\" is based on TGP (total planning time), but the path length P.L. is not consistently better: in QaLdn and Nicut, Ours has longer P.L. than MS3DSR (167.77 vs 121.90 and 131.18 vs 120.16), and only marginally better than ACAMS in some scenes. Moreover, it is unclear whether ACAMS and MS3DSR were run under identical conditions (same number of robots, same maximum planned views, same trajectory execution length, same underlying 3DGS/ASH representations). Without this detail, the comparison may not be apples-to-apples. Please clarify the evaluation protocol and also discuss the trade-off between planning time and path length rather than calling both \"efficiency\" interchangeably.","section":"Section III-D, Eq. (7)"},{"comment":"The improved K-means objective mixes different physical quantities without stated weights: D_r is a sum of Euclidean distances (meters), while (N^M_r - N)^2 and |D_r - D| are dimensionless counts and distances, respectively. No weighting coefficients are defined, so the relative importance of movement cost versus load balancing is unspecified. In addition, the manuscript does not explain how the optimization in (7) is performed (e.g., iterative assignment with restarts, a local search over centroids, or an exact solver). Please specify the algorithm, including how the two terms are normalized or weighted, and how the number of clusters is chosen.","section":"Section IV-A.2"}],"minor_comments":[{"comment":"The heading \"Sing-robot view path planning\" contains a typo; it should be \"Single-robot view path planning\".","section":"Section III-D.2"},{"comment":"In Fig. 2, the label \"Uncers\" appears to be a fragment; it should be \"Uncertainty\" or \"Instance uncertainty\" for clarity.","section":"Fig. 2"},{"comment":"The notation for reconstruction tasks is inconsistent: the text defines T_rec = {T_rec_1, T_rec_2, ..., T_rec_Nrec}, while Algorithm 1 outputs T_rec as a set of T_rec_i, but line 7 uses v_rec_i and line 8 assigns T_rec_i <- v_rec_i. Please unify the notation.","section":"Section III-C.2"},{"comment":"The real-robot experiment reports only qualitative results and no quantitative reconstruction metrics or comparison with baselines; it would be helpful to state explicitly that this is a feasibility demonstration and to provide quantitative measures if available.","section":"Section IV-D"},{"comment":"The claim of being \"the first 3DGS-based centralized multi-robot autonomous 3D reconstruction framework\" should be substantiated with a broader literature search; the related work only cites GS-Planner as the single-robot 3DGS method, but there may be other concurrent or recent multi-robot 3DGS systems that should be discussed.","section":"Section I"}],"recommendation":"major_revision","confidential_remarks":"The paper presents a promising framework and a well-structured ablation, but the central semantic-gating mechanism (objectness score as completeness proxy) needs rigorous validation before the claims can be accepted. The efficiency comparison also needs clearer protocol details. The novelty claim about being 'first' should be checked against the latest literature, as concurrent work on multi-robot 3DGS may exist. I recommend major revision with a focus on calibration and experimental rigor."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's my take. The paper is a competent multi-robot active reconstruction system built around 3DGS. What's new is the specific combination: a global 3DGS map, online OVIR-3D semantic segmentation to gate reconstruction tasks, and a two-tier planning scheme with global-local mode assignment and an improved K-means for task clustering. The ablations (V1-V6) are thoughtfully designed and show each component pulls its weight; the comparison against ACAMS and MS3DSR favors the proposed system on both reconstruction quality and planning time.\n\nThe biggest weakness is the semantic completeness proxy. The objectness score is the max softmax probability over ScanNet200 after temperature scaling. That measures how confidently the CLIP feature matches a category, not how complete the geometry is. A half-scanned chair can still be confidently labeled 'chair'; a fully scanned plain wall may be ambiguous. The fixed thresholds (Cmin=0.2, Cmax=0.6) are given without any sensitivity analysis or calibration evidence. Because the quality gains appear when semantic guidance is turned on (V5/V6 vs V3/V4), this unvalidated gating is load-bearing. If the thresholds misbehave in other scenes, the method could generate tasks for already-finished objects or ignore incomplete ones, and the claimed gains vanish.\n\nAlso, results are single-run metrics for each scene with no error bars. Given planning randomness, that is thin evidence, especially against methods with inherent variance. No code or data release further limits reproducibility. The real-robot demo is qualitative and on two robots, which is fine as a sanity check but not evidence for the headline numbers.\n\nMinor point: 'superior planning efficiency' is based on TGP (planning time); the path length is sometimes longer than MS3DSR, though quality is better. That trade-off should be stated.\n\nOverall, the system is coherent and the engineering is solid. The stress-test concern about the objectness score is real and should be addressed experimentally. This paper is for people working on active multi-robot reconstruction and 3DGS-based planning. It deserves a serious referee, but I would ask for threshold sensitivity analysis, multiple runs with variance reporting, and ideally code before accepting. My vote is major revision.","headline":"A competent multi-robot 3DGS reconstruction system whose semantic-gating assumption is the load-bearing weakness.","tokens_in":13274,"tokens_out":3393,"would_cite":true,"duration_ms":34628,"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":"This paper proposes a centralized multi-robot 3D reconstruction framework that uses 3D Gaussian splatting (3DGS) alongside online open-vocabulary semantic segmentation, claiming the first system of its kind.","keywords":["3D Gaussian splatting","multi-robot reconstruction","autonomous exploration","open-vocabulary semantic segmentation","view planning","surface uncertainty","task assignment","indoor scene reconstruction"],"falsifier":"Run the pipeline in an indoor scene containing object categories not covered by ScanNet200, and record which instances pass the objectness filter during the first few reconstruction rounds. If many incomplete objects receive top-1 softmax scores below $C_{\\min}=0.2$, or many already-complete objects score above $C_{\\max}=0.6$, then the generated reconstruction tasks are misdirected. A threshold sweep on a held-out scene that changes final reconstruction quality by more than the reported margins would demonstrate that the claim is scene-specific rather than general.","tokens_in":12162,"feed_emoji":"🤖","tokens_out":4550,"duration_ms":45303,"temperature":0.7,"pith_summary":"The paper claims that multi-robot autonomous 3D reconstruction can be made faster and more accurate by focusing each robot's scanning on semantically identifiable objects that are still incomplete. It presents a centralized framework built around 3D Gaussian splatting, where an online open-vocabulary segmenter identifies object instances in the partial point cloud, and the surface uncertainty from 3DGS pinpoints which parts of those instances are poorly reconstructed. The claimed result is the highest reconstruction quality among the compared planning methods, with planning time reduced to a small fraction of prior multi-robot systems. The significance is that large indoor scenes, which are slow for a single robot to cover, could now be reconstructed by teams that avoid wasting views on already-finished regions or on noisy uncertainty hotspots.","feed_headline":"Robot teams scan scenes faster with semantic guidance","feed_subtitle":"A 3DGS-based planner focuses view sampling on incomplete objects, cutting planning time while raising reconstruction quality.","key_machinery":"The load-bearing mechanism is the instance-uncertainty-guided reconstruction task generator. Surface points from occupied hash voxels carry a cached 3DGS loss as their uncertainty; an open-vocabulary segmenter (OVIR-3D) clusters those points into instances; a softmax over ScanNet200 similarity scores with temperature $\\lambda_e=50$ produces an objectness score, and thresholds $C_{\\min}=0.2$ and $C_{\\max}=0.6$ discard instances that are either too noisy or already well reconstructed. For each remaining instance, points of interest are selected greedily by uncertainty with a minimum spacing of $d_{\\mathrm{POI}}=1.2\\,\\mathrm{m}$, candidate viewpoints are sampled in empty space, and the view maximizing information gain $g(v)=\\sum_k \\sigma_k e^{-0.5 d_{v,k}}$ (where $\\sigma_k$ is the surface uncertainty and $d_{v,k}$ the distance to a visible surface point) becomes a reconstruction task. This machinery converts a scattered uncertainty field into a small set of focused, semantically grounded scan targets, which is what the paper argues avoids local optima and redundant views.","core_discovery":"The central claim is that coupling 3DGS surface uncertainty with open-vocabulary instance segmentation yields a better next-view signal than surface uncertainty alone, and that separating mode assignment from task assignment makes the NP-hard multi-robot planning problem tractable. Concretely, the method projects per-frame color and depth losses into occupied voxels to get a surface uncertainty field, feeds the surface points to an online open-vocabulary 3D instance segmenter, computes an objectness score for each instance against the ScanNet200 vocabulary, keeps only instances with intermediate scores, and samples viewpoints around the highest-uncertainty points of those instances. Robots are then split into exploration and reconstruction modes by a global task ratio plus local task counts, and tasks within each mode are clustered with a modified K-means that balances both the number of tasks and the internal travel distance per robot. In experiments across three indoor scenes, the paper reports higher PSNR, lower accuracy and completion error, and higher recall than the compared single- and multi-robot planners, while cutting per-round planning time from tens of seconds to a few seconds.","pith_inferences":["A natural extension would be to make the objectness thresholds adaptive per scene or per object category, since fixed $C_{\\min}$ and $C_{\\max}$ values are the most scene-sensitive part of the pipeline.","The same instance-uncertainty signal could be used in single-robot exploration, replacing frontier-only coverage objectives with a semantic-completeness objective that explicitly prioritizes incomplete objects.","For known environments, swapping the open-vocabulary segmenter for a faster closed-set detector could reduce latency while retaining the semantic guidance, at the cost of generality.","The centralized design's communication load is acknowledged by the authors as future work; a distributed variant would need to share only task centroids and instance labels rather than the full map and uncertainty field."],"forward_implications":["If the central claim holds, robotic teams can reconstruct large indoor scenes with fewer total views, because scans are concentrated on incomplete objects rather than spread across all high-uncertainty noise.","The separation of mode assignment from task assignment removes the repeated TSP solving that dominates planning time in prior multi-robot planners, making the approach scale better as the number of tasks grows.","The improved K-means objective, which balances both task count and internal travel distance, should prevent the long single-robot paths that cause energy-limited robots to truncate their tours and leave parts of the scene unscanned.","The method is demonstrated on real robots, suggesting that the combination of 3DGS uncertainty, ASH geometry, and semantic guidance can run online in a centralized multi-robot setting.","The fixed objectness thresholds and the use of the ScanNet200 vocabulary imply a direct dependence on the segmenter's open-vocabulary quality, so improvements in that segmenter would translate directly into better reconstruction task generation."],"supporting_citations":[{"why":"Supplies the online open-vocabulary 3D instance segmentation that identifies reconstruction instances from the partial surface point cloud.","marker":"[13]"},{"why":"Supplies the loss-caching strategy that projects per-frame loss into world space to define the 3DGS surface uncertainty field.","marker":"[8]"},{"why":"Provides the Gaussian-SLAM base representation that the global 3DGS map is trained with.","marker":"[22]"},{"why":"Provides the parallel spatial hashing geometry (ASH) used to extract surface points and prune noisy Gaussians.","marker":"[2]"},{"why":"Supplies the frontier-exploration plus surface-reconstruction task formulation and viewpoint sampling strategy that this work extends.","marker":"[24]"},{"why":"Serves as the key multi-robot baseline whose mode-switching MDMTSP formulation and approximate solver motivate the improved assignment strategy.","marker":"[4]"},{"why":"Serves as the multi-robot next-best-view baseline that is compared against in the experiments.","marker":"[6]"},{"why":"Supplies the incremental frontier exploration and ESDF-based travel cost used for exploration tasks and trajectory optimization.","marker":"[26]"}],"fun_headline_variants":["Robot teams cut planning time with semantic 3D mapping","Gaussian splatting guides multi-robot 3D scene reconstruction","Semantic-aware robot fleets map large scenes faster","Multi-robot planner uses semantics to boost 3D reconstruction quality"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole approach depends on the open-vocabulary segmenter returning trustworthy objectness scores for partially scanned objects in real time; if the top-category probability is unreliable or the fixed thresholds are scene-dependent, reconstruction tasks get aimed at the wrong regions and the claimed quality and efficiency gains disappear.","fun_headline_variants_meta":{"raw":{"variants":["Robot teams cut planning time with semantic 3D mapping","Gaussian splatting guides multi-robot 3D scene reconstruction","Semantic-aware robot fleets map large scenes faster","Multi-robot planner uses semantics to boost 3D reconstruction quality"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000787,"raw_usage":{"total_tokens":3475,"prompt_tokens":951,"completion_tokens":2524,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":567,"completion_tokens_details":{"reasoning_tokens":2453}},"tokens_in":567,"tokens_out":2524,"duration_ms":18185,"temperature":1.0,"reasoning_tokens":2453,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-11T23:40:26.390629+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the pipeline in an indoor scene containing object categories not covered by ScanNet200, and record which instances pass the objectness filter during the first few reconstruction rounds. If many incomplete objects receive top-1 softmax scores below $C_{\\min}=0.2$, or many already-complete objects score above $C_{\\max}=0.6$, then the generated reconstruction tasks are misdirected. A threshold sweep on a held-out scene that changes final reconstruction quality by more than the reported margins would demonstrate that the claim is scene-specific rather than general.","supporting_citations":[{"cited_title":"Ovir-3D: Open-vocabulary 3D instance retrieval without training on 3D data","cited_arxiv_id":null,"evidence_quote":"Supplies the online open-vocabulary 3D instance segmentation that identifies reconstruction instances from the partial surface point cloud."},{"cited_title":"ASH: A modern framework for parallel spatial hashing in 3D perception","cited_arxiv_id":null,"evidence_quote":"Provides the parallel spatial hashing geometry (ASH) used to extract surface points and prune noisy Gaussians."},{"cited_title":"Autonomous implicit indoor scene reconstruction with frontier exploration","cited_arxiv_id":null,"evidence_quote":"Supplies the frontier-exploration plus surface-reconstruction task formulation and viewpoint sampling strategy that this work extends."},{"cited_title":"Asyn- chronous collaborative autoscanning with mode switching for multi- robot scene reconstruction","cited_arxiv_id":null,"evidence_quote":"Serves as the key multi-robot baseline whose mode-switching MDMTSP formulation and approximate solver motivate the improved assignment strategy."},{"cited_title":"A multirobot system for 3-d surface reconstruc- tion with centralized and distributed architectures","cited_arxiv_id":null,"evidence_quote":"Serves as the multi-robot next-best-view baseline that is compared against in the experiments."},{"cited_title":"Fuel: Fast UA V exploration using incremental frontier structure and hierarchical planning","cited_arxiv_id":null,"evidence_quote":"Supplies the incremental frontier exploration and ESDF-based travel cost used for exploration tasks and trajectory optimization."}],"review_version":1}