{"id":"1f03c64f-1399-42e9-95d8-00e2bbf7b1b9","arxiv_id":"2501.08672","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A LiDAR-inertial-visual SLAM system that maintains a global Gaussian map in CPU memory and optimizes only a sliding window of Gaussians on the GPU, achieving real-time mapping and odometry on an embedded platform.","lead":"GS-LIVO combines LiDAR, camera, and inertial data with a 3D Gaussian map to run real-time odometry and photorealistic mapping on resource-limited robots. It reports map updates above 10 Hz indoors on an NVIDIA Jetson Orin NX, much faster than prior Gaussian SLAM systems, while using far less GPU memory.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Sliding-window-only Gaussian optimization is never stress-tested on a revisit; without loop-closure evidence, the global-consistency part of the real-time SLAM claim remains unsupported.","rationale":"The reader's weakest_assumption is the same one I identify as most load-bearing: optimization is restricted to the current FoV, retired Gaussians are not re-optimized, and no experiment exercises a revisit or loop closure. I agree the paper is a credible engineering contribution and the sliding-window mechanism plausibly delivers bounded GPU cost and real-time operation on the reported sequences. My concern does not overturn the paper; it narrows what the current evidence supports. The paper should either add a revisit/loop-closure experiment or soften the SLAM/global-consistency framing to odometry with a local Gaussian map. Because this is an addressable gap in evidence rather than a demonstrated failure, the appropriate verdict remains CONDITIONAL, and I do not change the reader's recommendation.","tokens_in":21883,"tokens_out":9220,"duration_ms":92939,"concrete_test":"Run GS-LIVO on a trajectory that revisits a previously mapped region, for example a custom indoor loop with MoCap ground truth or a MARS-LVIG sequence with a return. At the revisit, render the stored global Gaussians from the current estimated pose, compute PSNR/SSIM against the current camera image, and compare odometry RMSE before and after the loop. As a control, temporarily re-insert the revisited old Gaussians into the sliding window and repeat the same metrics; if the control substantially improves PSNR or reduces post-loop RMSE, the permanent retirement of old Gaussians is the cause, and the global-consistency claim fails.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central real-time SLAM claim depends on the assumption that optimizing only the Gaussians currently in the sliding window is sufficient for a globally consistent map. Section II.C explicitly retires Gaussians that leave the field of view by copying them into the global hash-indexed octree, and Section II.D's IESKF contains no loop-closure mechanism; once retired, those Gaussians are never re-optimized when revisited. The experiments in Section III (CBD03, HKU01, HKisland03, HKairport01, Playground01/02, landmark01) are forward trajectories without a measured return to a previously mapped area, so there is no evidence about how the map behaves when old and new observations of the same place disagree. Since pose drift from the odometry is permanently baked into the stored Gaussian positions, the paper supports real-time odometry with a local Gaussian map, but not yet the stronger claim of a globally consistent SLAM map over long, revisiting trajectories.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"GS-LIVO proposes a LiDAR-inertial-visual odometry and mapping system in which the map is represented by 3D Gaussians stored in a hash-indexed octree, with only the Gaussians in a sliding window optimized on the GPU. The paper claims real-time high-frequency map updates (over 10 Hz indoors and around 3 Hz outdoors) on an NVIDIA Jetson Orin NX, competitive odometry accuracy compared with FAST-LIVO, R3LIVE, and LVI-SAM, and reduced memory and optimization time compared with 3D-GS, SplaTAM, and MonoGS. Experiments are reported on subsets of the FAST-LIVO2 and MARS-LVIG datasets plus three indoor sequences, with PSNR for rendering quality and RMSE for odometry accuracy.","tokens_in":22016,"tokens_out":5592,"duration_ms":51638,"significance":"If validated, the contribution is genuinely useful: the memory-bounded split between a RAM-resident global octree and a VRAM-resident sliding window is a credible engineering solution for scalable Gaussian SLAM, and the IESKF visual update that renders the Gaussian map to form photometric residuals is a sensible extension of FAST-LIVO2. The paper also promises open-source release of algorithms, hardware designs, and CAD models, which would help the community. However, the current evidence is not sufficient to support the central claims: the outdoor RMSE numbers contradict each other, the PSNR evaluation is measured on the same frames used for optimization, and no experiment tests global consistency after revisiting a previously mapped area. The architecture is plausible, but the evaluation needs substantial strengthening.","major_comments":[{"comment":"The outdoor accuracy claim is internally inconsistent and therefore not verifiable. The text states that for outdoor scenarios 'our method demonstrates superior accuracy with an RMSE of 0.042m', while Table III reports 0.58 m for HKisland03 and 0.63 m for HKairport01 for GS-LIVO, and Fig. 7 shows errors around 0.54 m RMSE. The reported numbers need to be reconciled, and the paper should state whether the 0.042 m value is a typo, a different sequence, or a different metric. No variance or repeated-run statistics are provided for any of the RMSE values.","section":"Section III.B.2, Table III"},{"comment":"The global-consistency claim of the SLAM system is not supported by any experiment. Section II.C retires Gaussians that leave the field of view by copying them into the global hash-indexed octree, and Section II.D's IESKF contains no loop-closure mechanism, so retired Gaussians are never re-optimized when the robot revisits a location. All odometry experiments in Section III are forward trajectories without a measured return to a previously mapped area. Please add a revisit or loop-closure experiment, or explicitly restrict the claims to local odometry with a local Gaussian map.","section":"Section II.C, Section II.D, Section III"},{"comment":"The PSNR comparison in Table II measures fit to the training views rather than novel-view synthesis. The Gaussians are optimized against the same images Ik used to compute PSNR through Eq. (9), and no held-out view synthesis metric is reported. The 'roaming perspective' images in Fig. 5 are qualitative only. Please report PSNR, SSIM, or LPIPS on held-out viewpoints or on frames not used in the optimization.","section":"Section II.B.2, Section III.B.1, Table II"},{"comment":"The experimental configuration is described inconsistently. Section III.B.1 states that indoor scenes use a root voxel size of 0.03 m and outdoor scenes use 1.0 m, while Section III.B.2 states that the octree configuration is (0.06 m, 2 layers) for indoor and (0.5 m, 2 layers) for outdoor environments. These are materially different parameter choices and must be reconciled in a single parameter table so that the results in Tables II and III are reproducible.","section":"Section III.B.1 vs. Section III.B.2"}],"minor_comments":[{"comment":"The phrase 'as demonstrated in Tab. 7' refers to a nonexistent table; it should refer to Table III.","section":"Section III.B.2"},{"comment":"The column header 'Dur./ms↑' uses an upward arrow, although lower duration is better; the arrow direction is inconsistent with Table II's 'Dur./s↓'.","section":"Table IV"},{"comment":"The text lists the outdoor sequences as HKisland03 and HKairport01, but Fig. 12 labels include 'HKisland01' and 'HKairport03'; these labels should be corrected.","section":"Section III.A, Fig. 12"},{"comment":"The scale parameters sδ, sy, and sz in Eq. (2) are not fully defined; sδ is called a hyper-parameter, but the reader needs to know how sy and sz are obtained from the LiDAR downsampling or voxel level.","section":"Section II.B.1, Eq. (2)"},{"comment":"The sentence 'where ex is the normal vector on x-axis' is unclear; presumably ex is the unit vector of the x-axis, and the notation should be stated explicitly.","section":"Section II.B.1, Eq. (3)"},{"comment":"The entries marked ✗ for S3GS on HKU01, HKisland03, and HKairport01 are not explained; please state why the comparison is unavailable for those sequences.","section":"Table II"}],"recommendation":"major_revision","confidential_remarks":"For the editor: the paper is a plausible systems contribution, but the submitted version contains a direct contradiction in the headline outdoor RMSE, a broken table reference, and no evidence for the global-consistency aspect of the SLAM claim. These issues are fixable with additional experiments and corrections, so I recommend major revision rather than rejection. The claim of being 'the first' Gaussian-based SLAM on an embedded platform should also be checked carefully against prior embedded NeRF/3D-GS systems before publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The genuinely new thing here is the system architecture: a CPU-resident hash-indexed octree for the global Gaussian map, a contiguous GPU sliding window of the Gaussians currently in view, and incremental host-device transfers so only the visible subset is optimized. That is a sensible way to bound GPU memory, and the ablation with and without the sliding window shows real runtime and memory gains. The embedded deployment on the Jetson Orin NX, with plans to release code and CAD, is a concrete and checkable claim. The IESKF odometry that uses rendered Gaussians as a measurement model is also a reasonable extension of FAST-LIVO2, and the photometric Jacobians in the appendix are worked out in enough detail to be followed.\n\nThat said, the evidence is not as clean as the claims. There is an internal contradiction in the outdoor RMSE: the text reports 0.042 m for HKisland03 and HKairport01, while Table III reports 0.58 m and 0.63 m for the same sequences. There is also a broken reference to \"Tab. 7\" in Section III.B.2, which looks like it should point to Table III. These are easy to fix, but they make it hard to trust the quantitative story without a careful rerun.\n\nMore substantively, the PSNR numbers are computed on the same frames used to optimize the Gaussians, so they measure fitting, not novel-view synthesis. The paper claims high-fidelity mapping, but without held-out views or at least a cross-validation split, the rendering-quality comparison to SplaTAM and MonoGS is weakened. There are also no error bars or multiple runs, and the closest prior systems—Gaussian-LIC and LiV-GS—are cited but not compared against.\n\nThe stress-test note holds up. Section II.C retires Gaussians that leave the field of view, and Section II.D's IESKF has no loop-closure mechanism; once retired, those Gaussians are never re-optimized when revisited. The evaluated sequences (CBD03, HKU01, HKisland03, HKairport01, Playground01/02, landmark01) are forward trajectories without a measured return to previously mapped territory. So the paper supports real-time odometry with a locally maintained Gaussian map, but not yet the stronger claim of a globally consistent SLAM map over long, revisiting trajectories. That is an important scoping issue, not necessarily a fatal one.\n\nThe paper is worth a serious referee. The core engineering contribution is real and likely useful to the robotics community, and the problems are addressable in revision: fix the numbers, add novel-view metrics, run a revisit sequence, and either add the missing baselines or justify their omission. I would engage with this work and would cite it for the sliding-window architecture, but I would not yet cite its global-consistency claims.","headline":"Real-time Gaussian mapping on an embedded GPU is a genuinely useful integration, but the paper's own evidence undercuts its headline numbers and the global-consistency claim is untested on revisits.","tokens_in":22634,"tokens_out":1677,"would_cite":true,"duration_ms":18877,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A LiDAR-inertial-visual odometry system keeps 3D Gaussian map updates running in real time on resource-constrained embedded hardware.","keywords":["3D Gaussian Splatting","SLAM","LiDAR-inertial-visual odometry","multi-sensor fusion","sliding window optimization","embedded systems","octree","photometric rendering"],"falsifier":"Run the system on a trajectory that leaves a mapped area, travels away, and later returns: if the rendered old Gaussians visibly disagree with the new camera image, or if the odometry jumps when re-observing known structure, then sliding-window-only optimization is not sufficient for global consistency.","tokens_in":21619,"feed_emoji":"🗺️","tokens_out":9550,"duration_ms":80853,"temperature":0.7,"pith_summary":"GS-LIVO tries to show that 3D Gaussian Splatting, a photorealistic scene representation that normally demands large GPU budgets, can be embedded in a tightly coupled LiDAR-inertial-visual odometry pipeline that runs in real time on a resource-constrained robot computer. The central claim is that by storing the global Gaussian map in CPU memory and optimizing only a sliding window of Gaussians inside the current field of view, map updates can be kept at over 10 Hz indoors and about 3 Hz outdoors while odometry accuracy stays competitive with established multi-sensor SLAM systems. The paper reports per-frame pipeline latencies below 90 ms on a desktop and about 48 ms on a small ARM-based embedded computer, with GPU memory use between 1.2 GB and 9.5 GB depending on scene scale. If the claim holds, photorealistic dense maps become usable online for navigation and planning rather than only for offline reconstruction.","feed_headline":"Real-time 3D Gaussian map updates now work on embedded robot hardware","feed_subtitle":"A multi-sensor mapping system keeps photorealistic map refresh above 10 Hz indoors by optimizing only visible Gaussians.","key_machinery":"The load-bearing mechanism is the sliding window of Gaussians: a contiguous CPU buffer holding the Gaussians of the current field of view, mirrored on the GPU and updated incrementally frame to frame through a spatial hash table. When a voxel leaves the field of view, its optimized parameters are copied back into the global hash-indexed octree in RAM; when a new voxel enters, it is appended to the contiguous buffer and transferred to the GPU. This keeps host-to-device transfers proportional to newly visible voxels rather than to the entire map, and lets the Adam-based photometric optimization operate only on visible Gaussians. A second mechanism is the hash-indexed octree itself, which indexes the global map by spatial hash keys over sparse volumes and supports multiple voxel levels so that detail can adapt to indoor and outdoor scene scales.","core_discovery":"The paper's discovery is that the practical bottleneck in Gaussian-based SLAM is the scope of map optimization, not the Gaussian representation itself. Restricting photometric optimization to a sliding window of currently visible Gaussians, while persisting all retired Gaussians in a hash-indexed octree in CPU memory, keeps per-frame GPU work and memory nearly constant as the map grows. The visual measurement model renders the current Gaussian window at the LiDAR-inertial predicted pose, compares the rendered image with the real camera image, and feeds the photometric residual and its Jacobians into an iterative error-state Kalman filter that also ingests LiDAR point-to-plane residuals and IMU propagation. The paper reports indoor trajectory RMSE of about 5 mm, outdoor aerial RMSE of 0.58 m, and a total pipeline time of 48.3 ms on the embedded platform, and presents the system as the first real-time Gaussian-based SLAM with online map updates on such hardware.","pith_inferences":["A natural extension the paper does not pursue is re-activating retired Gaussians when the robot revisits an area, which would add a loop-closure mechanism for the global map; without it, long-revisit global consistency remains an open question.","Because the visual measurement model only requires a rendered image and a photometric residual, the same sliding-window architecture could in principle wrap any differentiable radiance-field representation, not just 3D Gaussian Splatting.","The reported outdoor rendering quality (around 15 dB PSNR) is noticeably lower than indoor quality, suggesting large-scale outdoor scenes stress the fixed-level octree; a testable improvement is adaptive voxel sizing driven by range, texture, or structural complexity.","The integrated autonomous navigation demonstration indicates the Gaussian map can be projected to occupancy grids for planning; a natural next experiment is measuring how much map update rate, as opposed to pose accuracy, limits closed-loop control performance."],"forward_implications":["At the reported update rates, the Gaussian map can be consumed online for occupancy-grid generation, obstacle avoidance, and novel-view rendering during a mission, not after it.","Because only field-of-view Gaussians occupy GPU memory while the global map lives in CPU RAM, environments larger than GPU memory can be mapped as long as system RAM or swap space is available.","Sequential IESKF updates let the visual rendering residual contribute a pose update with covariance that propagates to later IMU and LiDAR updates, making the odometry more robust to brief loss of one sensor.","Root voxel size, octree depth, image resolution, and sliding-window capacity form a direct trade-off between update frequency and map detail, allowing the same system to be configured for indoor and large-scale outdoor operation."],"supporting_citations":[{"why":"Supplies the differentiable 3D Gaussian representation and alpha-blending renderer used for the map and visual measurement model.","marker":"[25]"},{"why":"Supplies the tightly-coupled iterated Kalman filter odometry backbone and the evaluation sequences used for validation.","marker":"[8]"},{"why":"Supplies the patch-based photometric visual measurement model that the Gaussian rendering model replaces.","marker":"[7]"},{"why":"Supplies the photometric residual and camera-pose Jacobians for Gaussian rendering that the visual update reuses.","marker":"[51]"},{"why":"Is the RGB-D Gaussian SLAM baseline whose slow map updates motivate the sliding-window optimization.","marker":"[33]"},{"why":"Supplies the surface normal estimation used to orient planar Gaussians during LiDAR-camera initialization.","marker":"[56]"},{"why":"Serves as a real-time LiDAR-inertial-visual odometry baseline for accuracy and runtime comparison.","marker":"[5]"},{"why":"Serves as a factor-graph-based LiDAR-visual-inertial baseline for comparison.","marker":"[54]"},{"why":"Supplies the aerial dataset with RTK ground truth used to validate large-scale outdoor odometry.","marker":"[9]"}],"fun_headline_variants":["Gaussian SLAM optimized for embedded robots with sliding-window updates","Real-time Gaussian mapping now possible on resource-limited hardware","Multi-sensor fused Gaussian SLAM achieves real-time on Jetson Orin","Sliding-window optimization keeps Gaussian map updates fast and light","LiDAR-inertial-visual fusion powers first embedded real-time Gaussian SLAM"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The argument stands or falls on the assumption that optimizing only the Gaussians inside the current field-of-view sliding window is sufficient for a globally consistent map and accurate odometry over long trajectories, so Gaussians that leave the window never need to be re-optimized.","fun_headline_variants_meta":{"raw":{"variants":["Gaussian SLAM optimized for embedded robots with sliding-window updates","Real-time Gaussian mapping now possible on resource-limited hardware","Multi-sensor fused Gaussian SLAM achieves real-time on Jetson Orin","Sliding-window optimization keeps Gaussian map updates fast and light","LiDAR-inertial-visual fusion powers first embedded real-time Gaussian SLAM"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001037,"raw_usage":{"total_tokens":4427,"prompt_tokens":1067,"completion_tokens":3360,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":683,"completion_tokens_details":{"reasoning_tokens":3268}},"tokens_in":683,"tokens_out":3360,"duration_ms":22809,"temperature":1.0,"reasoning_tokens":3268,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:20:11.052181+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the system on a trajectory that leaves a mapped area, travels away, and later returns: if the rendered old Gaussians visibly disagree with the new camera image, or if the odometry jumps when re-observing known structure, then sliding-window-only optimization is not sufficient for global consistency.","supporting_citations":[{"cited_title":"Fast-livo2: Fast, direct lidar-inertial-visual odometry,","cited_arxiv_id":null,"evidence_quote":"Supplies the tightly-coupled iterated Kalman filter odometry backbone and the evaluation sequences used for validation."},{"cited_title":"Fast-livo: Fast and tightly-coupled sparse-direct lidar-inertial-visual odometry,","cited_arxiv_id":null,"evidence_quote":"Supplies the patch-based photometric visual measurement model that the Gaussian rendering model replaces."},{"cited_title":"R3live: A robust, real-time, rgb-colored, lidar- inertial-visual tightly-coupled state estimation and mapping package,","cited_arxiv_id":null,"evidence_quote":"Serves as a real-time LiDAR-inertial-visual odometry baseline for accuracy and runtime comparison."}],"review_version":1}