{"id":"ef54be41-6c99-4173-a3a3-fc4383071b2e","arxiv_id":"2505.12384","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":3.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":0,"one_line_summary":"A comparative survey and embedded benchmark concluding that semantic geometric SLAM is more practical for real-time deployment than NeRF- or Gaussian-splatting-based semantic SLAM.","lead":"This paper surveys semantic SLAM methods and benchmarks a subset of them on an NVIDIA Jetson AGX Orin embedded computer. It concludes that geometric semantic SLAM is more practical on constrained hardware than NeRF- and Gaussian-splatting-based semantic SLAM.","discovery_kind":"review","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table IV's FPS/power rankings conflate offline-mask preprocessing with real-time segmentation, so the geometric-SLAM viability conclusion rests partly on externalized semantic cost.","rationale":"The reader's weakest_assumption identified the preprocessing asymmetry and the TUM-average concern, and I agree these are the right issues. I focus on the preprocessing asymmetry as the single most load-bearing concern because it directly undermines the quantitative basis of the paper's headline conclusion: Table IV is the only embedded resource-utilization measurement in the paper, and VDO-SLAM's favorable numbers come from moving segmentation off-device. The paper itself admits this in Section IV-C, so the issue is not hidden; it is a missing end-to-end measurement rather than an internal contradiction. The concern is load-bearing because without VDO-SLAM's 20.9 FPS / 10.9 W entry, the geometric column of Table IV is much weaker: RDS-SLAM variants run at 4.8 and 10.7 FPS, and Panoptic-SLAM at 2.7 FPS. Dynamic-VINS (24.9 FPS, 12 W, online YOLOv3) does provide genuine support for the conclusion, so the paper is not wrong in direction; it is wrong in the strength of the evidence it presents. I would keep the verdict CONDITIONAL, matching the reader, because the concern is addressable: add an end-to-end measurement for VDO-SLAM, report sequence-level ATE, and temper the conclusion to name the specific systems that actually support it. The GS-vs-NeRF efficiency claim is also unsupported by the paper's own experiments (no NeRF system was run on the Jetson), but that claim is secondary to the main conclusion and can be fixed by rewording. No fraud or misconduct is implied; the authors transparently describe their protocol, which is exactly why the missing end-to-end measurement is identifiable and fixable.","tokens_in":31530,"tokens_out":3003,"duration_ms":32447,"concrete_test":"Rerun VDO-SLAM on the NVIDIA Jetson AGX Orin with its Mask R-CNN segmentation executed online (or, minimally, measure Mask R-CNN inference time separately and add it to VDO-SLAM's per-frame time), then recompute the FPS and power columns of Table IV. Also report per-sequence ATE RMSE for every TUM RGB-D dynamic sequence (fr3/sitting_*, fr3/walking_*) instead of a single average, and state exactly which sequences are included in the Table II average. If VDO-SLAM's end-to-end FPS drops below roughly 10 FPS or its power rises substantially, and/or if the per-sequence ATE on fr3/walking_xyz is several centimeters, the conclusion that geometric SLAM is currently the most viable embedded choice would need to be re-scoped to specific systems and specific dynamic conditions.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central claim that Semantic-aware Geometric SLAM is the most viable option for embedded deployment is supported primarily by Table IV, but that table mixes two incomparable conditions. Section IV-C explicitly states that VDO-SLAM's segmentation is performed offline as a preprocessing step and that its second-place resource numbers are 'only possible because it only needs to load pre-computed segmentation masks.' The FPS and power figures for VDO-SLAM (20.9 FPS, 10.9 W) therefore exclude the dominant cost of semantic extraction, which on the Jetson AGX Orin is precisely the cost that matters for embedded deployment. The conclusion then highlights 'systems such as VDO-SLAM and Dynamic-VINS' as evidence that selective semantic processing keeps costs manageable, but VDO-SLAM's measured performance is not end-to-end. The GS systems also consume preprocessed semantics, so their poor FPS (0.013) is not the source of unfairness; the problem is that the geometric-vs-GS comparison gets its best geometric numbers from a system whose semantic pipeline ran off-device. A secondary but related issue: Table II reports a single ATE average over unstated TUM sequences, with an ORB-SLAM2 baseline of 1.0 cm. Typical ORB-SLAM2 ATE on the hardest sequence, fr3/walking_xyz, is around 7-10 cm, which suggests the most difficult walking sequences may be excluded from the average. If so, the accuracy advantage claimed for semantic geometric methods on dynamic scenes is not demonstrated on the regimes where they are supposed to matter. Dynamic-VINS, which runs YOLOv3 online, remains a legitimate real-time geometric system, so the conclusion may survive correction; but the quantitative ranking that makes the claim convincing is currently contaminated by an unmeasured preprocessing cost.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript surveys three families of semantic visual SLAM (geometric, NeRF, and Gaussian Splatting), then reports an experimental comparison of a selected subset (RDS-SLAM, VDO-SLAM, Dynamic-VINS, Panoptic-SLAM, GS3LAM, SGS-SLAM) on an NVIDIA Jetson AGX Orin using ATE, mIoU, RAM, power, and FPS metrics. The central conclusion is that Semantic-aware Geometric SLAM currently provides the most viable solution for real-time embedded deployment, while NeRF- and 3DGS-based methods are too resource-intensive for constrained platforms. The paper also argues that GS-based systems are more computationally efficient than NeRF-based ones, and it discusses future directions such as hardware-software co-design.","tokens_in":31942,"tokens_out":5899,"duration_ms":60578,"significance":"If accepted, the central conclusion would give the embedded SLAM community a clear research direction: focus on geometric pipelines with selective semantics rather than dense neural scene representations. The paper's strengths are its concrete resource measurements on the Jetson AGX Orin, the containerized experimental setup, the per-component timing breakdown in Figure 9, and the broad coverage of recent literature. The conclusion is plausible, but the support has load-bearing gaps: Table IV mixes end-to-end and off-device semantic processing, Table II averages over an unspecified TUM sequence subset, and the NeRF-versus-GS efficiency comparison is not backed by any Orin measurements of NeRF systems.","major_comments":[{"comment":"Table IV is the main evidence for the paper's central claim, but it compares systems under different semantic-preprocessing conditions. As the text states, VDO-SLAM consumed precomputed Mask R-CNN masks in an offline preprocessing step, so its reported 20.9 FPS and 10.9 W exclude the semantic extraction cost; Section II-A further notes that VDO-SLAM is only applicable to pre-recorded benchmarks. The conclusion in Section V cites VDO-SLAM alongside Dynamic-VINS as evidence that selective semantic processing keeps costs manageable, yet VDO-SLAM's measurements are not end-to-end. Please provide end-to-end numbers with on-device segmentation, or explicitly add the segmentation time and energy to VDO-SLAM's totals, and flag in Table IV which systems perform semantics in real time.","section":"Section IV-C, Table IV"},{"comment":"Table II reports a single average ATE over TUM RGB-D sequences, but the sequence subset is not stated. The ORB-SLAM2 baseline of 1.0 cm ATE RMSE is far below typical values on the dynamic walking sequences, which strongly suggests that the hardest sequences are excluded from the average. Since the paper claims that semantic geometric methods improve accuracy in dynamic environments, the table should either list per-sequence ATE for all TUM sequences used or clearly state the subset and justify it; otherwise the accuracy advantage is not demonstrated on the regime that motivates semantic SLAM.","section":"Section IV-B, Table II"},{"comment":"The conclusion that \"GS-enhanced SLAM systems typically offer better computational efficiency than NeRF-based methods\" is not supported by the experiments in this paper, because no NeRF-based system was run on the Jetson AGX Orin: Section IV-C states that NIS-SLAM and vMAP could not be executed due to insufficient resources, and the NeRF entries in Tables II and III are literature values from [5] that were obtained on different platforms. Either run representative NeRF systems on the same Orin setup, or restrict the conclusion to what Table IV can support, namely geometric systems versus the two GS systems.","section":"Section V, with Section IV-C and Table III"},{"comment":"The power figures in Table IV are load-bearing for the embedded-deployment recommendation, but the manuscript does not describe how power was measured (e.g., wall meter, on-module sensors, averaging window, or load conditions). Without this methodology, even the end-to-end comparisons that remain after the VDO-SLAM issue are hard to interpret; please add a short measurement-protocol paragraph.","section":"Section IV-A, Section IV-C"}],"minor_comments":[{"comment":"RDS-SLAM is described as being built on the ORB-SLAM2 framework, but Table I and Section II-A state ORB-SLAM3; please reconcile this inconsistency.","section":"Section IV, first paragraph"},{"comment":"The sentence \"For these experiments, two widely used datasets TUM RGB-D and Replica\" lacks a verb; rephrase to something like \"For these experiments, we use two widely used datasets: TUM RGB-D and Replica.\"","section":"Section IV-B, first sentence"},{"comment":"The bullet symbol (•) is used in several columns without a legend; please state that a bullet means \"yes\" or otherwise explain the notation.","section":"Table I"},{"comment":"The caption contains a typo: \"an be incorporated\" should read \"can be incorporated.\"","section":"Figure 6 caption"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is better characterized as an experimental comparison with survey elements rather than a pure survey; the central recommendation is reasonable but needs methodological tightening before publication. I saw no concerns about attribution or excessive self-citation: reference [16] is used as background only and does not drive the conclusions."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: worth engaging, but the headline claim rests on more Jetson runs than on a fully controlled comparison. The genuinely new piece is the Orin resource table, not the survey narrative.\n\nWhat it does well: the field is organized into geometric/NeRF/GS families with a readable map of how semantics enter each pipeline. The experiments are real: Dockerized systems actually run on an AGX Orin, with FPS, power, and RAM reported. That is more than most surveys do. The paper also openly discloses that VDO-SLAM uses precomputed segmentation masks offline, which is honest. The conclusion that semantic geometric SLAM is currently the practical embedded choice is plausible and probably correct.\n\nSoft spots, in proportion: the comparison is not apples-to-apples. VDO-SLAM's 20.9 FPS excludes the segmentation cost entirely, so its second-place resource numbers are not end-to-end deployment costs. The GS systems also consume preprocessed semantics, but their FPS is so low that the geometric-vs-GS ranking likely survives; still, the quantitative table overstates the geometric advantage. The TUM ATE average is underspecified: an ORB-SLAM2 baseline of 1.0 cm implies the hardest walking sequences are excluded, which weakens the dynamic-scene accuracy argument just where semantics supposedly matter. There are no error bars or repeated runs. The \"first survey\" claim is false given the cited [17], and Figure 1's caption contradicts the text. The NeRF-versus-GS efficiency claim appears without Orin runs of any NeRF system. These are all fixable.\n\nWho this is for: embedded robotics practitioners wanting a compact comparison of semantic SLAM families and a starting point for deployment studies. The survey narrative breaks little ground, but the benchmark is a useful first step. I would send it to peer review with the expectation of moderate revision: correct the novelty claim, specify the TUM sequences, release the Docker setup, and rerun with repeated trials and clearly separated preprocessing costs. The central direction holds up.","headline":"A useful but uneven embedded semantic SLAM survey whose central claim is plausible yet rests on a benchmark with incomparable preprocessing conditions.","tokens_in":32377,"tokens_out":1420,"would_cite":true,"duration_ms":17290,"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":"Semantic geometric SLAM is the only family that runs near real time on embedded hardware.","keywords":["semantic SLAM","embedded systems","NeRF","3D Gaussian Splatting","resource utilization","Jetson AGX Orin","dynamic environments","semantic segmentation"],"falsifier":"Re-running the benchmark with all systems required to compute semantic segmentation live on the Jetson AGX Orin, and reporting per-sequence accuracy on the dynamic RGB-D benchmark including the hardest walking sequence (fr3/walking_xyz), would settle the ranking: if Gaussian-splatting systems then approach real time while geometric systems stall, or if the geometric advantage disappears on the hardest sequences, the conclusion fails.","tokens_in":1653,"feed_emoji":"🤖","tokens_out":1702,"duration_ms":86617,"temperature":0.7,"pith_summary":"Robots that build maps with object labels—semantic SLAM—must fit accuracy, memory, and power budgets when deployed on embedded processors. The paper compares the three main architecture families—geometric, NeRF-based, and 3D Gaussian Splatting—on a single embedded platform, the Jetson AGX Orin, measuring localization error, segmentation quality, memory, speed, and energy. Its central finding is that semantic geometric SLAM is currently the most viable family for real-time embedded deployment, with Dynamic-VINS reaching 24.9 FPS at 12 W, while Gaussian-splatting systems run at 0.013–0.014 FPS and NeRF systems could not be run at all. NeRF and Gaussian Splatting still offer the richest semantic maps, so the paper frames the conclusion as a call for efficiency work and algorithm–hardware co-design rather than a permanent verdict.","feed_headline":"Geometric semantic SLAM tops embedded speed test","feed_subtitle":"On a Jetson AGX Orin, Gaussian-splatting SLAM runs at 0.013 FPS while Dynamic-VINS reaches about 25 FPS.","key_machinery":"The argument is carried by a resource-utilization benchmark on a single embedded platform, the Jetson AGX Orin, that measures trajectory accuracy (ATE RMSE, the root-mean-square error between estimated and ground-truth trajectories), semantic quality (mIoU, the average overlap between predicted and true masks), memory, power, and speed across the three architecture families, using a dynamic RGB-D benchmark for geometric accuracy and the Replica dataset for semantic quality. The decisive mechanism is the semantic-integration mode: systems that apply segmentation only to keyframes or run it live, versus systems that consume precomputed semantic masks, determine whether semantics can stay within real-time budgets. These measurements are what separate the geometric family's near-real-time performance from the neural-representation families' high semantic detail.","core_discovery":"The paper's central claim is that, on the Jetson AGX Orin, semantic geometric SLAM is currently the only architecture family that balances accuracy, memory, power, and speed well enough for real-time embedded deployment: Dynamic-VINS runs at 24.9 FPS with 8.29 GB RAM and 12 W, and VDO-SLAM at 20.9 FPS with 6.62 GB and 10.9 W, while the two Gaussian-splatting systems, GS3LAM and SGS-SLAM, run at 0.013–0.014 FPS with over 16 GB RAM and over 15 W even when consuming precomputed semantic masks. NeRF-based semantic SLAM systems could not be executed on the Orin under the paper's setup, and their reported accuracy on the Replica dataset, while sub-centimeter, does not close the embedded-deployment gap. The paper also observes that Gaussian-splatting methods are more efficient than NeRF methods while delivering comparable reconstruction quality and semantic consistency, but still far from real time on embedded hardware.","pith_inferences":["A natural extension is to rerun the comparison with hardware-accelerated neural-network inference, since the paper deliberately used unoptimized inference; such acceleration could change the ranking if geometric systems benefit more from faster segmentation.","A fairer embedded-deployment metric would count segmentation cost in the end-to-end loop for every system, since the current protocol gives precomputed-mask systems an advantage and the geometric-vs-Gaussian gap could narrow or widen under uniform treatment.","The dynamic-scene comparison is currently only geometric; a dynamic benchmark with semantic annotations could test whether Gaussian-splatting systems' high mIoU degrades when objects move and are occluded."],"forward_implications":["Embedded semantic SLAM research should prioritize geometric pipelines with lightweight, keyframe-only semantic modules, since those delivered the only near-real-time results on the Orin.","Gaussian-splatting SLAM needs memory and compute reductions of more than an order of magnitude before real-time embedded deployment; at 0.013–0.014 FPS it is two orders of magnitude below real-time.","Panoptic semantics are not yet viable on embedded hardware, with Panoptic-SLAM reaching only 2.7 FPS, so the current embedded trade-off favors object-level or lightweight semantic segmentation.","NeRF-based semantic SLAM, despite sub-centimeter accuracy on Replica, could not be run on the Orin in this study, leaving its embedded feasibility unproven.","Algorithm–hardware co-design, including accelerators and lightweight model variants, is the paper's proposed path for making dense neural representations embeddable."],"supporting_citations":[{"why":"Supplies RDS-SLAM, a keyframe-only geometric system whose SegNet and Mask-RCNN variants anchor the accuracy and resource tables.","marker":"[18]"},{"why":"Supplies VDO-SLAM, a geometric system with precomputed semantic masks whose 20.9 FPS result supports the embedded viability claim.","marker":"[19]"},{"why":"Supplies Dynamic-VINS, the embedded-designed system that runs live segmentation and reaches 24.9 FPS at 12 W.","marker":"[20]"},{"why":"Supplies GS3LAM, a Gaussian-splatting system with the highest mIoU but 16.56 GB RAM and 0.013 FPS.","marker":"[22]"},{"why":"Supplies SGS-SLAM, a Gaussian-splatting system with dynamic-object handling that still runs at 0.014 FPS.","marker":"[24]"},{"why":"Provides the NeRF and Gaussian-splatting accuracy and mIoU values the paper could not reproduce on the Orin.","marker":"[5]"},{"why":"Provides the dynamic RGB-D sequences and ground-truth trajectories used for the trajectory accuracy comparison.","marker":"[143]"},{"why":"Provides the semantically annotated indoor scenes used to measure mIoU on Replica.","marker":"[144]"}],"fun_headline_variants":["Embedded SLAM: geometric wins, Gaussian splatting lags","Jetson test: semantic SLAM trade-offs laid bare","Gaussian-splat SLAM hits 0.013 FPS on embedded board","Only geometric semantic SLAM fits real-time embedded use"],"cache_read_input_tokens":34432,"weakest_assumption_plain":"The central comparison assumes that systems given precomputed semantic masks can be ranked for embedded deployment on the same footing as systems that must run segmentation live, and that the average over the chosen RGB-D dynamic sequences fairly represents dynamic-scene performance.","fun_headline_variants_meta":{"raw":{"variants":["Embedded SLAM: geometric wins, Gaussian splatting lags","Jetson test: semantic SLAM trade-offs laid bare","Gaussian-splat SLAM hits 0.013 FPS on embedded board","Only geometric semantic SLAM fits real-time embedded use"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000262,"raw_usage":{"total_tokens":1618,"prompt_tokens":988,"completion_tokens":630,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":604,"completion_tokens_details":{"reasoning_tokens":555}},"tokens_in":604,"tokens_out":630,"duration_ms":5520,"temperature":1.0,"reasoning_tokens":555,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T20:34:28.592514+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-running the benchmark with all systems required to compute semantic segmentation live on the Jetson AGX Orin, and reporting per-sequence accuracy on the dynamic RGB-D benchmark including the hardest walking sequence (fr3/walking_xyz), would settle the ranking: if Gaussian-splatting systems then approach real time while geometric systems stall, or if the geometric advantage disappears on the hardest sequences, the conclusion fails.","supporting_citations":[],"review_version":1}