{"id":"baf0efca-739a-4187-8655-1da3debf41bd","arxiv_id":"2504.16782","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"Graph2Nav merges 2D image scene graphs into a global 3D graph with object relations using SLAM, and those relations improve real robot object-search time with an LLM planner.","lead":"Graph2Nav turns live camera and LiDAR data into a 3D map of objects and their relationships, such as a table beside chairs, indoors and outdoors. A robot using this map with a language-model planner found backpacks and bottles faster in real tests.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Navigation improvement may be an artifact of trial order: Section V-C does not state that maps are reset or conditions randomized between paired trials.","rationale":"The reader's weakest assumption concerned the 3D merge procedure for relations, where single-frame relations are inserted into the global graph without multi-view consistency checks. That is a relevant technical concern, but the more load-bearing issue is the validity of the navigation experiment itself. The paper's primary claim is that object relations improve search efficiency, supported by Table III. The protocol for these paired trials is under-specified: no map reset between trials is mentioned, no randomization is stated, and no repeated runs or error bars are provided. Under these conditions, the observed search-time reduction could reflect trial order or environmental familiarity rather than the object-relation edges. This directly undermines the central claim. The reader noted small-scale evidence and internal inconsistencies, which align with this concern, but did not identify the trial-order confound explicitly. Thus my agreement with the reader is partial. The data-count inconsistency in Section V-B (67 total objects vs. 105 summed from the listed per-environment counts) further erodes confidence in the accuracy tables, but the navigation confound is the decisive problem. Given that the central claim is unsupported by the currently reported evidence, the appropriate verdict is REJECT rather than CONDITIONAL; the authors would need to provide a properly controlled navigation experiment to restore the claim.","tokens_in":9744,"tokens_out":5908,"duration_ms":55886,"concrete_test":"Re-run the six search scenarios with the SLAM map and scene graph fully reset before every trial, randomize the order of the No Relations and With Relations conditions per scenario, and repeat each condition at least five times. Compare the distributions of search times; if the improvement vanishes or falls within run-to-run noise, the claimed navigation benefit is not supported.","verdict_should_be":"REJECT","load_bearing_attack":"The central claim that object relations improve search efficiency rests entirely on Table III (Section V-C). For each scenario, the robot performs two trials: one without relations and one with relations from Graph2Nav. The paper never states that the SLAM map and 3D scene graph are reset between these paired trials, nor that the order of conditions is randomized or counterbalanced. If the with-relations trial always follows the no-relations trial, the robot enters the second trial with prior knowledge of the environment, enabling faster search independent of any object-relation edges. Even if the map is reset, LLM stochasticity and the absence of repeated trials make a 18–35% gap over six non-randomized pairs insufficient to establish causation. This is not a minor omission: it is the only direct evidence for the paper's headline claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"Graph2Nav proposes a real-time framework for building 3D object-relation scene graphs from RGB and LiDAR data. A modified PSGFormer predicts 2D panoptic scene graphs per keyframe; LIO-SAM provides poses and semantically labeled voxel clusters form 3D object nodes; per-frame 2D relation graphs are merged into a global 3D layered graph. The authors evaluate object centroid accuracy in four indoor/outdoor environments, relation accuracy on 30 indoor and 17 outdoor human-annotated relations, and integrate the graph with SayNav on a UGV for six object-search scenarios, reporting 18.1-34.8% search time reductions when object relations are used.","tokens_in":9910,"tokens_out":4203,"duration_ms":37651,"significance":"The contribution is potentially significant: it is among the first demonstrations of a real robot using a 3D scene graph with object-object relations to ground an LLM planner in both indoor and outdoor settings, and the design of leveraging 2D panoptic scene graph training data avoids the direct 3D relation training data constraint. The paper's strengths include a clear system architecture, use of an external PSG baseline and SayNav as a prior planner, and an ablation of relations within the same Graph2Nav graph rather than comparing to unrelated methods. However, the evidence for the headline claim is thin: six single-run paired trials without randomization/reset details, and small hand-annotated relation sets. The accuracy claims also contain a numerical inconsistency in object counts. If the experimental protocol is clarified and the relation and navigation results are strengthened, the paper would be a useful contribution.","major_comments":[{"comment":"The central claim that object relations improve search efficiency rests entirely on Table III, but the protocol description does not state whether the SLAM map and 3D scene graph are reset between the paired 'No Relations' and 'With Relations' trials, nor whether the order is randomized or counterbalanced across the six scenarios. If the with-relations trial always follows the no-relations trial, the robot benefits from prior exploration independent of any relation edges; even with resets, six single runs per condition without standard deviations or a statistical test cannot establish causation. Please report the exact protocol (map reset, order, number of runs, randomization), and ideally add repeated trials and a paired test.","section":"Section V-C, Table III"},{"comment":"Relation accuracy is measured on only 30 indoor and 17 outdoor hand-labeled relations, with no inter-annotator agreement or sampling protocol described. The outdoor comparison is 16/17 versus 11/17, i.e., a difference of five relations; this is too sparse to support the claim of outdoor relation-generation generalization. Report a larger annotation set, annotation guidelines, agreement metrics, and confidence intervals or significance tests.","section":"Section V-B, Table II"},{"comment":"The global 3D scene graph merge accepts relations from a single 2D keyframe without geometric or multi-view verification: 'New nodes are added to G with the corresponding edges from Gi' and updating existing nodes are the only merge steps described. A false relation detected from one oblique viewpoint is therefore inserted permanently into the 3D graph, which can inflate both the relation accuracy in Table II and the navigation benefit in Table III. Please specify an edge-verification mechanism (e.g., voting across viewpoints, geometric relation classifiers, or an uncertainty threshold) and report how many edges come from single versus multiple observations.","section":"Section III-C"},{"comment":"The ground-truth object counts are internally inconsistent: the text states 'total 67 measured objects' and then lists 32 (cafeteria) + 38 (lab) + 22 (courtyard) + 13 (parking lot) = 105, and it refers to 'these three environments' while four environments are named. Since the localization errors in Table I depend on these counts, please correct the numbers and clarify per-environment totals so that the results are reproducible.","section":"Section V-B, Table I"}],"minor_comments":[{"comment":"In the overview sentence, '2D papnoptic scene graph generation' contains a typo; it should be 'panoptic'.","section":"Section III"},{"comment":"The word 'descibe' should be 'describe'.","section":"Section IV"},{"comment":"The phrase 'Data Distribution Serivce' should be 'Data Distribution Service'.","section":"Section V-C"},{"comment":"The equation 'P r(G|I))' has an unbalanced parenthesis; the probability notation should be introduced more carefully.","section":"Section III-B"},{"comment":"The word 'fulfull' should be 'fulfill', and the opening sentence of the conclusion repeats the abstract nearly verbatim; consider tightening.","section":"Section VI"}],"recommendation":"major_revision","confidential_remarks":"The paper is within scope and the system is a real integration, but the headline navigation result is supported by very small, possibly confounded trials. I would encourage the editor to request raw per-trial data and a precise protocol description before final acceptance. No concerns about citation behavior."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version: this paper is a real but uneven integration of 2D panoptic scene graph prediction with LiDAR SLAM to build a 3D layered scene graph with intra-level object relations, and it demonstrates the result on a real UGV with an LLM planner. That combination is new as far as I know, and the accuracy tables are plausible. The real-world deployment is a step forward. But the headline navigation result rests on a paired-trial design that is not described carefully enough to rule out an order effect, and the quantitative support is thinner than the positive claims imply.\n\nThe core idea is clean: don't train a 3D scene graph network; lift a 2D panoptic scene graph (PSGFormer) into 3D using LIO-SAM poses and merge per-frame graphs. The layered structure (site / region / object) plus same-level object relation edges is what distinguishes it from Kimera, Hydra, and similar real-time 3D scene graph systems. The paper is honest about using existing components, and the citation pattern is fine. The real-robot search experiments, even with the confound, show a complete system working in a previously unseen lab.\n\nConcerns, in rough order of severity. First, Table III: six scenarios, each with a no-relations and a with-relations trial. The text never says whether the SLAM map and scene graph are reset between the pair, nor whether order is randomized or counterbalanced. If the with-relations trial always came second, the robot has already explored the room, and the improvement could come from memory rather than from relation edges. That is the central claim, so it needs repeated, randomized trials. Second, the relation ground truth is tiny: 30 indoor and 17 outdoor hand-picked pairs. The 94% outdoor accuracy is 16/17, so one error matters a lot. Third, internal inconsistencies: the object count text sums to 105, not the stated 67, and the text says 'three environments' while the table has four. These are fixable but they undercut trust. Fourth, the 3D merge carries over each frame's edges with no geometric or multi-view check; a false 'on top of' seen once is inserted into the global graph permanently.\n\nThe math is simple—the only equation is a factorization—and the components are clearly attributed. Who gets value: robotics researchers working on 3D scene graphs and LLM-based navigation will find this a useful systems demonstration. It deserves a serious referee, not a desk reject, because it fills a real gap and the flaws are addressable.\n\nRecommendation: send to peer review with a request for major revision—clarify the experimental protocol, add repeated trials and statistics, release data and code, and either add a multi-view consistency check or discuss the limitation. Conditional accept is the right ceiling.","headline":"A genuine real-robot integration of 2D panoptic scene graphs into 3D for LLM-based navigation, but the headline navigation gain is confounded by an under-specified paired-trial design.","tokens_in":10432,"tokens_out":4348,"would_cite":false,"duration_ms":36304,"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":"Object relationships cut robot search time by up to 35 percent","keywords":["3D scene graph","object-relation graph","panoptic scene graph","robot navigation","SLAM","semantic mapping","large language model planning","object search"],"falsifier":"Place two objects that are physically separated but appear stacked from one oblique viewpoint, drive the same sensor-equipped robot around them, and inspect the final global graph: if the 'on top of' relation appears even though the objects are not in contact, the single-frame merge assumption is producing false permanent edges.","tokens_in":9569,"feed_emoji":"🤖","tokens_out":7006,"duration_ms":60069,"temperature":0.7,"pith_summary":"Graph2Nav claims that a robot can build a useful 3D scene graph in real time by lifting 2D panoptic scene graphs into 3D with SLAM, and that the resulting object-relation edges make navigation more efficient. The paper tests this indoors and outdoors on a real unmanned ground robot, measuring both object localization and relation-labeling accuracy. When the relations are fed to an LLM-based planner during object search, the robot finds targets between 18.1% and 34.8% faster than when the same graph lacks relations. The significance is practical: it suggests robots can get spatial common sense from a small set of object relations without needing large 3D-labeled training datasets.","feed_headline":"Object relationships cut robot search time by up to 35 percent","feed_subtitle":"A real-time 3D scene graph built from camera and lidar data helps a voice-commanded robot find objects indoors and outdoors.","key_machinery":"The load-bearing machinery is a 2D-to-3D lift-and-merge pipeline. At each SLAM keyframe, a panoptic scene graph transformer produces a 2D graph of object masks, labels, and relations; using calibrated pixel-to-point correspondences, the masks are projected into the 3D point cloud, Bayesian updates assign consistent semantic labels to voxels, Euclidean clustering groups voxels into object nodes, and each frame's graph is translated into 3D and merged into a global layered scene graph. The merge step is what lets noisy single-view relations become temporally consistent graph edges without learning directly from 3D data.","core_discovery":"The paper's central claim is that a 3D object-relation scene graph can be generated online, for both indoor and outdoor scenes, by combining a per-frame 2D panoptic scene graph network with a pose-graph SLAM mapping system, and that the graph's relation edges directly improve downstream search navigation. Per keyframe, the network outputs object masks, class labels, and pairwise relations; pixel-to-point correspondences project them into 3D, Bayesian voxel updates reconcile labels, Euclidean clustering forms object nodes, and the per-frame graph is merged into a global layered graph. This lift-and-merge route sidesteps the lack of labeled 3D scene graph training data by reusing 2D panoptic scene graph datasets with 133 object classes and 56 relation types. In the paper's measurements, the merged graphs locate representative objects more accurately than single-image baselines, label indoor and outdoor relations correctly in 83-94% of annotated cases, and, when the relations are included in an LLM-based planner's prompts, reduce real-robot search times by 18-35% across six trials.","pith_inferences":["If a multi-view consistency check were added before an edge enters the global graph, false relations from a single oblique viewpoint would likely be filtered, which could push relation accuracy above the reported 83-94% and make the navigation gain more robust.","The same lift-and-merge construction could work with open-vocabulary or larger relation sets, because the graph does not depend on the specific 133/56 taxonomy; any per-frame relation head that can output labels could feed the merged graph.","The navigation gain measured here is for a small-object search task; a natural extension is to test whether relation edges also speed up coverage, exploration, or manipulation planning, where object layout matters more directly.","The reported search-time reduction with relations is consistent with the planner using relations to skip redundant waypoints, for example visiting one table to observe two nearby chairs; a direct plan-length comparison would tell whether the savings come mainly from fewer stops."],"forward_implications":["Including object-relation edges in the planner's scene graph prompt shortens measured search times by 18.1% to 34.8% over the same graph without edges, across six real-robot trials.","Graph2Nav's merged 3D graphs reduce object centroid localization error by 11.6% to 37.1% compared with using a single 2D segmented image.","Relation labeling in the global graph is more accurate than in single 2D frames: 83.3% vs 56.7% indoors and 94.1% vs 64.7% outdoors on the annotated test pairs.","Because the graph is built from keyframes of the robot's own SLAM trajectory, the representation is available online during exploration, not precomputed before the mission.","The same 133-class, 56-relation taxonomy covers indoor and outdoor settings, so the framework transfers without retraining on 3D scene graphs."],"supporting_citations":[{"why":"Supplies the one-stage panoptic scene graph network and the 133-class, 56-relation training dataset that Graph2Nav adapts with a lighter backbone.","marker":"[20]"},{"why":"Supplies the pose-graph SLAM whose keyframe poses and loop closures align the image, point cloud, and merged graph geometry.","marker":"[23]"},{"why":"Supplies the LLM-based planner that serves as the navigation integration target and as the no-relations baseline for the search-time comparisons.","marker":"[16]"},{"why":"Supplies the low-level path planning and control stack used to execute high-level steps during the robot trials.","marker":"[24]"},{"why":"Defines panoptic segmentation, the pixel-level labeling that the 2D graph network and 3D point labeling rely on.","marker":"[22]"}],"fun_headline_variants":["3D scene graph speeds robot object search by 35%","Real-time 3D scene graph improves robot navigation efficiency","Object relations in 3D graph boost robot search speed","Graph2Nav: 3D object relations accelerate robot search","Robot finds objects 35% faster with 3D relation graph"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Every relation detected in one camera frame is assumed to hold in the global 3D graph, because new edges are added directly when their nodes appear, with no multi-view or geometric check that the relation is still true from other viewpoints.","fun_headline_variants_meta":{"raw":{"variants":["3D scene graph speeds robot object search by 35%","Real-time 3D scene graph improves robot navigation efficiency","Object relations in 3D graph boost robot search speed","Graph2Nav: 3D object relations accelerate robot search","Robot finds objects 35% faster with 3D relation graph"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000515,"raw_usage":{"total_tokens":2507,"prompt_tokens":962,"completion_tokens":1545,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":578,"completion_tokens_details":{"reasoning_tokens":1460}},"tokens_in":578,"tokens_out":1545,"duration_ms":10732,"temperature":1.0,"reasoning_tokens":1460,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:55:16.761909+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Place two objects that are physically separated but appear stacked from one oblique viewpoint, drive the same sensor-equipped robot around them, and inspect the final global graph: if the 'on top of' relation appears even though the objects are not in contact, the single-frame merge assumption is producing false permanent edges.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the one-stage panoptic scene graph network and the 133-class, 56-relation training dataset that Graph2Nav adapts with a lighter backbone."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the pose-graph SLAM whose keyframe poses and loop closures align the image, point cloud, and merged graph geometry."},{"cited_title":"Rajvanshi, K","cited_arxiv_id":null,"evidence_quote":"Supplies the LLM-based planner that serves as the navigation integration target and as the no-relations baseline for the search-time comparisons."},{"cited_title":"Macenski, F","cited_arxiv_id":null,"evidence_quote":"Supplies the low-level path planning and control stack used to execute high-level steps during the robot trials."},{"cited_title":"Kirillov, K","cited_arxiv_id":null,"evidence_quote":"Defines panoptic segmentation, the pixel-level labeling that the 2D graph network and 3D point labeling rely on."}],"review_version":1}