{"id":"70c72d7f-84bc-443e-9419-0d8a85233796","arxiv_id":"2507.05718","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A cooperative multi-modal SLAM framework for ISAC systems improves radio map accuracy, user localization, and beam management in CARLA simulations.","lead":"This paper combines radio-based mapping with camera and motion-sensor data so a 6G base station can build a radio map, locate users, and steer beams. In simulated indoor and street scenes, the hybrid system reports up to 60 percent lower mapping error and about 37 percent better user localization than single-sensor approaches.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Beam-management claim depends on evaluating the predicted SE of Eq. (26) with channel estimates that omit the very interference the algorithm tries to mitigate; the IUI-rejection step is therefore not shown to do anything.","rationale":"The reader's verdict is CONDITIONAL with the weakest assumption being camera-to-UE association. That is a real concern but the association step is at least explicitly gated behind the D_th threshold and its failure mode degrades gracefully to radio-IMU localization, as the paper notes. Moreover, the localization gain is presented over a limited time window (15–40 s) and is partially acknowledged. The beam-management SE claim, by contrast, is stated as a headline result (36% in the conclusion) and is the part of the paper that most directly connects sensing to communication. The paper never states that \\hat{H}_j(t) is used only as a proxy, and it never validates the predicted-SE decision rule against true SE in simulation. That makes the SE improvement figure unsubstantiated in a way that a single simulation change could check. This is a different weakness from the association issue, although both are instances of simulation evidence not fully supporting the payoff. Having made this check, if the predicted-SE rule does track true SE, the verdict would still be CONDITIONAL (simulation-only, no code, thresholds unspecified); hence I do not propose changing the verdict, only flagging a sharper unresolved assumption to test.","tokens_in":23541,"tokens_out":1731,"duration_ms":17542,"concrete_test":"Recompute Algorithm 4's decision rule using the true channel H_j(t) (known in simulation) instead of the estimated channel \\hat{H}_j(t) in Eq. (26). Specifically, at every beam-selection step, compute SE = Σ_j log_2(1 + |w_j^H H_j f_j|^2 / (Σ_{i≠j}|w_j^H H_j f_i|^2 + σ^2)). If the selected beams from Algorithm 4 do not also improve true SE relative to the 'Tracking' baseline that always uses u_{j,opt}=u_{j,1}, then the 36% figure is an artifact of predicting SE with the estimated channel.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The sensing-aided beam management claim of a 36% SE gain over 'Tracking' rests on Algorithm 4, whose interference mitigation step (lines 6-14) evaluates the predicted SE in Eq. (26) using the estimated channel \\hat{H}_j(t) from Eq. (24), constructed only from the selected prior paths (angles and gains from the radio map). For the interference term in Eq. (26), the predicted SE sums |w_j^H \\hat{H}_j(t) f_i(t)|^2 over i≠j. But if \\hat{H}_j(t) has no component in the direction of UE i's beam f_i(t) whenever the selected prior paths for UE j differ from those for UE i, this term is essentially zero or unrelated to the actual interference experienced by UE j. The actual downlink interference depends on the true channel H_j(t), not on the estimated channel \\hat{H}_j(t) built from the same prior paths the algorithm is choosing. Thus the decision to switch from u_{j,opt}=u_{j,1} to a lower-gain path is justified only if the predicted SE is a faithful proxy for true SE; the paper provides no analysis showing that. The claim is therefore not internally inconsistent but is under-supported: the 36% gain could simply reflect that the predicted SE in Eq. (26) awards points to the management scheme for changing beams, even if true SE would not improve. This is the most load-bearing concern because the other two contributions (mapping and localization) are more directly validated by OSPA/eUE metrics, while the beam-management headline is the communication-layer payoff of the whole framework.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes a multi-modal SLAM framework for multi-user mmWave ISAC systems, combining cooperative radio SLAM, camera and IMU fusion, and sensing-aided beam management in one pipeline. The authors formulate a Bayesian factor-graph model that separates UE motion, local map transition, measurement update/data association, and global map update, and implement it as a two-stage algorithm (initialization without a global map, then refinement with global-map feedback). Localization fuses the radio/IMU SLAM estimate with YOLO/stereo-camera detections through a nearest-neighbor association gate and an offline-trained pixel-dependent error model. Beam management generates per-UE path priors from the global radio map and UE positions, switches selected paths based on a predicted spectral efficiency criterion intended to reduce inter-user interference, and then performs codebook-based tracking. Simulations in an indoor 10x10 m room and an outdoor CARLA street report radio-map OSPA improvements up to 60%, a windowed localization error reduction of 37.5%, and downlink SE gains of 36% and 149% over tracking and sweeping baselines.","tokens_in":23900,"tokens_out":17591,"duration_ms":176847,"significance":"The main strength of the paper is the breadth of the integration: a single Bayesian framework supports three concrete algorithmic outputs (map, location, beam selection), and the simulations compare against reasonable external baselines (single-user SLAM, AA and GCI fusion, tracking and sweeping beam management) rather than only against ablated versions of the proposed method. The outdoor CARLA evaluation with a real YOLOv8 detector and an offline-trained error lookup table is a concrete, falsifiable validation setup, and the paper is candid about the limitations of vision-only localization. If the results hold, the framework is a useful demonstration that SLAM-derived sensing can be converted into communication-layer gains in multi-user scenarios. The main reservations are that all quantitative claims rest on single synthetic scenarios without confidence intervals, the headline localization improvement is confined to a short time window, and the beam-management decision rule is not analyzed beyond the chosen scenario; these are addressed in the major comments.","major_comments":[{"comment":"The abstract claims that the proposed system 'enhances localization accuracy by 37.5%' without qualification, but Section VII-B reports this figure only for the time interval t in (15, 40) s in the outdoor scenario, when vehicles are near the middle of the street where the stereo camera is most accurate; the text states that multi-modal localization outperforms the other two SLAM algorithms only when t in (15, 40) s, implying the advantage does not persist outside this window. Please report the full-horizon average localization error with confidence bounds for the multi-user and multi-modal curves, and qualify the headline claim in the abstract and conclusion with the window over which it holds.","section":"Abstract; Section VII-B and Fig. 9(b)"},{"comment":"The switching step in Algorithm 4 (lines 6-14) maximizes the predicted spectral efficiency of Eq. (26), evaluated with the estimated channel H_j(t) built in Eq. (24) from the prior paths selected for UE j. Because the interference term in Eq. (26) sums |w_j^H H_j(t) f_i(t)|^2 over i not equal to j, it is near zero whenever UE j's selected paths and UE i's chosen beam have disjoint angular supports; in the extreme case of disjoint visibility sets, the interference term is identically zero and the switching logic can only react to the UE's own path gains. The 36% gain in Fig. 6(b) is computed with the true SE of Eq. (6) as stated in Section VII, which mitigates the concern that the reported number itself is a predicted-SE artifact, but the paper still provides no analysis of how well the predicted SE tracks the true SE after a switch, no ablation isolating the contribution of the switching step from the tracking codebooks, and no discussion of the degenerate disjoint-visibility case. Please add such an analysis or ablation and state the conditions under which the predicted-SE proxy is reliable.","section":"Section VI, Algorithm 4, Eqs. (24)-(26)"},{"comment":"The claimed localization gain rests on two assumptions that are not quantified. First, the nearest-neighbor association in Eq. (21) is never analyzed for misassociation probability, yet assigning the wrong camera detection to a UE corrupts the fused estimate in Eq. (22); the gating by the threshold D_Th does not by itself guarantee correct association when multiple objects are present. Second, the error lookup table in Eq. (20) is trained offline and validated only by a mean deviation of 0.036 m in CARLA, and is assumed to describe the vision error at all pixel locations and scenarios, including the online deployment conditions; Section VII-B itself shows that vision-only localization is discontinuous because of YOLO missed detections, so the camera branch is intermittently inactive. Please quantify the sensitivity of the localization improvement to misassociation rates, to the threshold D_Th, and to lookup-table mismatch, or the attribution of the 37.5% gain to the camera fusion branch remains unsupported.","section":"Section V, Eqs. (20)-(22); Section VII-B"}],"minor_comments":[{"comment":"All headline percentages (60%, 37.5%, 36%, 149%) are reported without confidence intervals or error bars even though 1,000 Monte Carlo runs are performed; adding standard-error bands to Figs. 4-6 and 9 would substantiate the comparisons.","section":"Section VII"},{"comment":"The OSPA curves for scenarios 2 and 3 are omitted 'due to space limits'; the claim that the algorithm maintains comparable radio map accuracy under local feature birth/death should be supported by the actual curves or by a quantitative statement of the maximum degradation relative to Scenario 1.","section":"Section VII-A, footnote 4"},{"comment":"The outdoor beam management analysis is stated to be omitted 'due to similar results', but no figure or numerical summary is given; the abstract's claim of significant outperformance in 'both indoor and outdoor environments' would benefit from at least a brief outdoor SE comparison.","section":"Section VII-B"},{"comment":"The parameter r_cut is used as a pruning/confidence threshold in Eqs. (14)-(17) but is described as a 'data association distance threshold' in Section VII-A, while Algorithm 1 uses d_th for the association distance; the parameter naming should be made consistent in the simulation description.","section":"Section IV-A and Section VII-A"},{"comment":"The pruning criterion in Eq. (16) uses a decay factor 2^(t_cand - t) with no stated rationale or time-unit convention, and the sentence 'we assume the radio map states as Q(0) = empty, P1:J = empty' appears to mean P_{1:J}(0) = empty; a clarification of both would help.","section":"Section III and Section IV-A"},{"comment":"Algorithm 2 lines 7 and 17 refer to updates 'similar to (16)', but (16) is the pruning criterion; the intended reference appears to be the fusion equations (15a)-(15e). Several minor typos also occur, such as inconsistent spacing 'D VIS-RF&IMU' versus 'DVIS-RF&IMU' in and around Algorithm 3, and interchangeable 'VA2', 'V A2', and 'V A 2' in Section VII.","section":"Algorithms 1-2 and throughout"}],"recommendation":"major_revision","confidential_remarks":"The manuscript fits the scope of an IEEE communications journal, and the three headline claims are at least defensible given the simulation evidence; my major_revision recommendation is driven by the gap between the abstract's unconditional wording and the windowed/conditional results in Section VII, and by the unvalidated predicted-SE decision rule in Algorithm 4. The reader's strongest concern about the SE evaluation does not land fully as stated, since Fig. 6(b) is evaluated with the true SE of Eq. (6); however, the underlying question about the decision-rule proxy remains open. The reuse of the authors' prior work ([1] and [23]) for the SLAM and beam-tracking components is appropriately cited; the novelty is in the integration, which is legitimate but should be presented as such."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This is a competent integration of prior radio-SLAM and beam-management work into a cooperative multi-modal framework. What is actually new is the end-to-end coupling of multi-user radio SLAM, camera/IMU fusion, and map-based IUI-aware beam selection. The two-stage map construction with feature birth/death handling is a concrete improvement over the single-stage baselines, and the offline-trained vision error look-up table is a practical touch. The paper is also honest: it explicitly says the refinement stage adds only marginal localization gain, and it shows vision-only localization is discontinuous due to YOLO missed detections.\n\nThe soft spots are in the evidence. All headline numbers come from CARLA, with no code or data released. The 37.5% localization improvement is averaged only over t in (15,40) seconds, a favorable window, and the vision association is nearest-neighbor with no treatment of mismatches; if YOLO misses or misdetects, the fusion in Eq. (22) is corrupted. Several thresholds are given as fixed values without sensitivity analysis. The most load-bearing concern is the beam management: Algorithm 4 decides to switch paths based on the predicted SE in Eq. (26), which is computed from the same estimated channel \\hat{H}_j(t) built from the selected prior paths. When two UEs choose different paths, the predicted interference term in Eq. (26) can be near zero by construction, so the IUI-rejection step may be optimizing a proxy that doesn't track true interference. The simulated 36% SE gain is genuine in the sense that the reported SE uses the true channel, but the paper does not explain why the proxy is faithful, which matters for real-world transfer. This is an under-supported claim, not an internal inconsistency, but it deserves a direct answer.\n\nThis paper is for readers working on ISAC, radio SLAM, or vision-aided mmWave beam management. They will find a clear framework, useful baselines, and a reasonable simulation study. It deserves a serious referee: send it to review. Ask for code or data, confidence intervals, threshold sensitivity, and an analysis or comparative simulation showing the predicted SE tracks true SE under mismatched association.","headline":"Solid integration of known radio-SLAM and beam-management pieces, but the evidence is simulation-only and the IUI-rejection claim needs a closer look at its proxy.","tokens_in":24441,"tokens_out":3075,"would_cite":false,"duration_ms":34924,"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":"A unified multi-modal SLAM loop ties cooperative radio map building, camera/IMU fusion, and beam management together, claiming up to 60% better maps, 37.5% sharper localization, and a 36% spectral-efficiency gain in mmWave ISAC systems.","keywords":["integrated sensing and communication","multi-user SLAM","radio map construction","beam management","multi-modal localization","mmWave","inter-user interference","virtual anchor"],"falsifier":"Run the outdoor CARLA scenario with a second vehicle crossing between the true UE and the camera so that the nearest detection in Eq. (21) is the wrong car; if the fused localization error rises above non-fused multi-user SLAM for more than occasional frames, the camera-fusion claim fails. A simpler check is to recompute the reported 37.5% localization gain while varying the association threshold DTh across 0.5 to 2 meters and see whether the gain survives.","tokens_in":23339,"feed_emoji":"📡","tokens_out":4679,"duration_ms":52257,"temperature":0.7,"pith_summary":"The paper is trying to show that one feedback loop can serve sensing and communication: multiple UEs jointly build a radio map, the map improves their own positions, and the map-plus-positions tell the base station which beams to use. A sympathetic reader should care because 6G millimeter-wave beam management currently wastes overhead on sweeping and is blind to which reflectors matter, while camera and IMU data are usually treated as separate from radio SLAM. The paper claims a unified Bayesian treatment of cooperative multi-user SLAM, with a two-stage map-building algorithm, is robust enough to handle dynamic visibility and features that appear or disappear, and that the same sensing outputs transfer directly into beam priors that cut inter-user interference. If true, this is a path to making sensing a communication asset rather than an add-on.","feed_headline":"Cooperative mmWave SLAM cuts map error by 60 percent","feed_subtitle":"A multi-modal loop of radio, camera, and IMU data also lifts localization accuracy and cuts inter-user beam interference.","key_machinery":"The load-bearing object is the global radio map Q(t), a set of virtual anchors — mirror images of the base station in reflective walls — each with a location, covariance, visibility vector across UEs, confidence, and last-update timestamp. The argument works through a two-stage Bayesian factor-graph SLAM where UEs run belief-propagation particle filters locally, the base station fuses uploaded local feature sets into Q(t) with nearest-neighbor association and variance-weighted updates, and UEs then download Q(t) as prior features to refine their own maps. The camera branch contributes a pixel-position-dependent variance lookup table, and beam management converts map features visible to each UE into ordered path priors (angles, search range, gain), selects paths while enforcing a minimum angular separation, and builds tracking codebooks around those priors.","core_discovery":"The paper proposes a single Bayesian framework in which multiple UEs run local radio SLAM from millimeter-wave angle estimates and IMU predictions, upload local virtual-anchor features to the base station, and then download a refined global radio map that in turn improves local estimation. The base station's stereo camera detections are fused with SLAM and IMU outputs through a pixel-dependent error model to produce UE locations, and those locations plus the map are converted into per-UE beam priors. The claimed payoff is a 60% reduction in radio-map OSPA error over single-user SLAM, a 37.5% improvement in UE localization accuracy over single-modal SLAM, and a 36% higher spectral efficiency than beam tracking with the same measurement budget.","pith_inferences":["The paper leaves open whether the nearest-neighbor camera-to-UE association in Eq. (21) survives dense scenes; a natural extension is probabilistic data association or track-based association, which would make the vision branch more robust when a non-UE object appears closer than the true UE.","The offline-trained pixel-error lookup table assumes the camera pose and scene geometry stay fixed; a testable extension is to re-estimate the table periodically or condition it on scene depth, because an outdated table would bias the fused estimate exactly where the camera is most needed.","Because the global radio map carries timestamps and visibility vectors, the same machinery could plausibly support predictive beam handover as UEs move between cells or around blockages, a use case the paper does not explore."],"forward_implications":["If the central claim holds, radio maps can be reconstructed collaboratively by ordinary UEs in deployment, reducing the need for site surveys or dedicated sensing infrastructure.","The map and localization outputs supply beam priors, so beam training can concentrate measurements on a narrow angular region rather than sweeping the full codebook, keeping overhead constant while improving resolution.","The inter-user interference mitigation step implies that spectral efficiency can improve as the number of UEs grows, in contrast to conventional beam sweeping where denser UEs suffer more overlapping beams.","Multi-modal fusion compensates for each sensor's individual failure mode: YOLO missed detections and IMU drift are bounded by the other modalities, so localization remains continuous in time.","The two-stage design can adapt to feature birth, feature death, and heterogeneous fields of view, suggesting the map does not need to be re-initialized when UEs lose or gain visibility of reflectors."],"supporting_citations":[{"why":"Supplies the belief-propagation multipath SLAM method that the local SLAM component builds on.","marker":"[11]"},{"why":"Provides the arithmetic-averaging and generalized-covariance-intersection map fusion baselines that the proposed algorithm is compared against.","marker":"[9]"},{"why":"Contributes the prior-information generation and beam-tracking approach that the sensing-aided beam management extends.","marker":"[23]"},{"why":"Defines the YOLO object-detection framework used by the stereo camera branch for multi-object localization.","marker":"[41]"},{"why":"Provides the CARLA driving simulator used to generate indoor and outdoor scenarios and the stereo camera measurements.","marker":"[44]"},{"why":"Supplies the pre-trained YOLOv8 model applied for vehicle detection in the outdoor scenario.","marker":"[45]"},{"why":"Supplies the measured channel-path angle noise variance that sets the radio measurement model.","marker":"[12]"}],"fun_headline_variants":["Multi-modal SLAM slashes mmWave map error 60%","Cooperative SLAM lifts UE localization accuracy 37.5%","ISAC beam priors from SLAM boost spectral efficiency 36%","Radio-camera-IMU fusion sharpens mmWave SLAM maps"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole vision gain rests on the assumption that the closest YOLO-detected object to each SLAM-tracked UE is that UE and that the offline-trained pixel-error table remains valid; if a nearer non-UE object or an invalid variance estimate attaches to the UE, the fused position is corrupted rather than improved.","fun_headline_variants_meta":{"raw":{"variants":["Multi-modal SLAM slashes mmWave map error 60%","Cooperative SLAM lifts UE localization accuracy 37.5%","ISAC beam priors from SLAM boost spectral efficiency 36%","Radio-camera-IMU fusion sharpens mmWave SLAM maps"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000179,"raw_usage":{"total_tokens":1298,"prompt_tokens":938,"completion_tokens":360,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":554,"completion_tokens_details":{"reasoning_tokens":283}},"tokens_in":554,"tokens_out":360,"duration_ms":4051,"temperature":1.0,"reasoning_tokens":283,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:19:45.275994+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the outdoor CARLA scenario with a second vehicle crossing between the true UE and the camera so that the nearest detection in Eq. (21) is the wrong car; if the fused localization error rises above non-fused multi-user SLAM for more than occasional frames, the camera-fusion claim fails. A simpler check is to recompute the reported 37.5% localization gain while varying the association threshold DTh across 0.5 to 2 meters and see whether the gain survives.","supporting_citations":[{"cited_title":"A belief propagation algorithm for multipath-based SLAM,","cited_arxiv_id":null,"evidence_quote":"Supplies the belief-propagation multipath SLAM method that the local SLAM component builds on."},{"cited_title":"5G mmWave cooperative positioning and mapping using multi- model PHD filter and map fusion,","cited_arxiv_id":null,"evidence_quote":"Provides the arithmetic-averaging and generalized-covariance-intersection map fusion baselines that the proposed algorithm is compared against."},{"cited_title":"Joint beam management and SLAM for mmwave communication systems,","cited_arxiv_id":null,"evidence_quote":"Contributes the prior-information generation and beam-tracking approach that the sensing-aided beam management extends."},{"cited_title":"You only look once: Unified, real-time object detection,","cited_arxiv_id":null,"evidence_quote":"Defines the YOLO object-detection framework used by the stereo camera branch for multi-object localization."},{"cited_title":"CARLA: An open urban driving simulator,","cited_arxiv_id":null,"evidence_quote":"Provides the CARLA driving simulator used to generate indoor and outdoor scenarios and the stereo camera measurements."},{"cited_title":"YOLO by ultralytics,","cited_arxiv_id":null,"evidence_quote":"Supplies the pre-trained YOLOv8 model applied for vehicle detection in the outdoor scenario."},{"cited_title":"Angle- based SLAM on 5G mmWave systems: Design, implementation, and measurement,","cited_arxiv_id":null,"evidence_quote":"Supplies the measured channel-path angle noise variance that sets the radio measurement model."}],"review_version":1}