{"id":"e536ed1e-e591-4562-8417-62bbe832c4cc","arxiv_id":"2501.13876","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"A LiDAR-inertial-visual odometry system skips most camera frames when the laser gives enough constraints, reducing runtime and memory on edge hardware for a small accuracy loss.","lead":"A robot navigation system now decides how often to use camera images based on whether the laser scanner is getting enough information, cutting computing load and memory on cheap ARM chips. It keeps most of the accuracy of the best previous system, losing about 3 cm on a standard benchmark, and runs in real time on a roughly one hundred dollar processor.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Translation-only degeneracy metric may miss rotational degeneracy in the Hilti sequences where accuracy drops most (Stairs, Cupola, Attic), undermining the claimed ~3 cm average trade-off.","rationale":"The reader's weakest_assumption correctly identified the degeneration detection and fixed threshold as the fragile premise. My analysis sharpens it: the paper's own data supply internal evidence of failure. The translation-only metric (Eq. 3) explicitly discards rotational constraints based on a rarity assumption in Section IV-A.1. The three sequences with the largest accuracy drops are independently described as rotation-heavy with sparse structure, the exact regime where rotational degeneracy can coexist with translational sufficiency. If the system skips visual frames in those regimes because sigma_min is high, the 'similar robustness' claim fails, and the 3 cm average is a misleading summary of a bimodal error distribution. This is more load-bearing than the abstract's numerical overstatements (Tables II and III imply roughly 24% runtime and 32% memory reductions, not 33% and 47%), because a reporting correction would not change the system, whereas the degeneration blind spot affects the central mechanism's correctness. I therefore keep the reader's CONDITIONAL verdict, now with a more specific condition: demonstrate that the selector does not skip needed visual frames in rotation-degenerate intervals, or restrict the robustness claim accordingly.","tokens_in":10339,"tokens_out":6939,"duration_ms":63074,"concrete_test":"Instrument the system to log per-frame sigma_min from Eq. (3), the number of visual frames selected, and per-axis pose error on the Hilti Construction Stairs, Cupola, and Attic to Upper Gallery sequences. If sigma_min remains above 0.07 during intervals where ATE grows to 0.15-0.22 m while FAST-LIVO2 maintains <0.07 m, and visual frames are being skipped in those same intervals, the translation-only criterion has failed to detect rotational insufficiency. A complementary offline check: compute the smallest eigenvalue of the full 6-DoF LiDAR Hessian for the same scans and compare it with the translation-only sigma_min; large divergence in the rotation-heavy intervals would confirm the metric's blind spot and invalidate the claimed universal threshold.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The adaptive visual frame selector's entire benefit rests on the LiDAR degeneracy signal from Eq. (3): the smallest normalized singular value of the translation-only plane-normal matrix. Section IV-A.1 justifies ignoring rotational DoFs by asserting that rotational-only degeneration with well-constrained translation is 'exceedingly rare.' Yet the paper's own Table I shows the three largest accuracy regressions versus FAST-LIVO2 occur in Construction Stairs (0.170 vs 0.016 m), Cupola (0.220 vs 0.121 m), and Attic to Upper Gallery (0.180 vs 0.069 m). In Section V-B.1 the authors describe such scenes as involving 'intense rotations and sparse structural information.' In these conditions LiDAR rotational constraints can be weak even when translational constraints from diverse plane normals remain sufficient; sigma_min then stays above the fixed 0.07 threshold, so the selector skips visual frames precisely when they are needed to constrain rotation. The reported 2.9 cm average RMSE increase conceals 6-15 cm per-sequence degradations, suggesting the degeneration detector's blind spot, not just random tuning noise. If this concern holds, the core mechanism removes the wrong frames in the scenarios where multi-sensor fusion is intended to provide robustness.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper extends the authors' FAST-LIVO2 LiDAR-inertial-visual odometry system for resource-constrained platforms. The proposed system adds a LiDAR-degeneration-aware adaptive visual frame selector that reduces the frequency of visual updates when a translation-only LiDAR constraint metric (the smallest normalized singular value of the plane-normal outer product) is above a threshold, and it replaces the single large unified map with a smaller local unified map plus a long-term visual map. The paper reports results on 16 Hilti 2022/2023 sequences, runtime and memory comparisons with FAST-LIVO2 on x86 and ARM platforms, qualitative private-dataset tests, and an ablation of the long-term visual map on the MARS-LVIG HKIsland03 sequence. The claimed net effects are a 33% per-frame runtime reduction, a 47% memory reduction, and a 3 cm average RMSE increase relative to FAST-LIVO2.","tokens_in":10614,"tokens_out":7016,"duration_ms":61421,"significance":"If the reported trade-off is accurate, the system is a useful engineering step for deploying direct LIVO on edge hardware. The paper does several things well: it evaluates on a standardized public benchmark, compares against multiple state-of-the-art systems, reports per-module runtime and memory figures, and includes an ablation for the long-term visual map. The adaptive visual frame selection idea is reasonable and the use of the Hilti benchmark limits benchmark overfitting. However, the headline efficiency numbers are not reproduced by the tables, and the degeneracy detector's translation-only metric is not validated against the specific sequences where accuracy degrades most. No code or data artifacts accompany the submission, so the quantitative contributions require additional verification.","major_comments":[{"comment":"The abstract's headline claims of a 33% per-frame runtime reduction and a 47% memory reduction are not supported by the reported data. From Table II, the mean total runtime reductions are (75.87−57.82)/75.87 ≈ 24% on ARM and (35.66−25.99)/35.66 ≈ 27% on x86; from Table III, the mean memory reduction is (2.5−1.7)/2.5 = 32%. The authors should either recompute the headline numbers from the tables or identify the specific subset of sequences and measurement conditions that yield 33% and 47%.","section":"Abstract; Section V-B.2; Section V-B.3; Tables II–III"},{"comment":"The adaptive visual frame selector rests on the claim that LiDAR degeneration can be inferred from the translation-only point-to-plane Jacobian A, with rotational DoFs ignored because translation-well-constrained rotational degeneration is said to be 'exceedingly rare' (Section IV-A.1). However, the largest RMSE regressions relative to FAST-LIVO2 in Table I occur in Construction Stairs (0.170 vs 0.016 m), Cupola (0.220 vs 0.121 m), and Attic to Upper Gallery (0.180 vs 0.069 m), which the paper associates with intense rotations and sparse structural information. In these scenes, the translation-only sigma_min from Eq. (3) can remain high while rotational constraints are weak, causing the selector to drop visual frames that are needed. Please provide per-sequence evidence that sigma_min correctly tracks the availability of rotational constraints in these sequences, or extend the degeneracy metric to the full SE(3) Jacobian and show that the fixed 0.07 threshold still selects appropriate frames.","section":"Section IV-A.1, Eq. (3); Table I"},{"comment":"The parameters of the central mechanism are presented as a consistent configuration, but in fact they are tuned per scene family: keyframe thresholds are set to 1 m–60° for indoor and 2 m–60° for outdoor sequences, while the map edge lengths (200 m/800 m) and sliding thresholds (20 m/100 m) are fixed. The paper provides no sensitivity analysis for the degeneration threshold 0.07 or for tau_predefined in Eq. (4), so it is unclear whether the reported efficiency-accuracy trade-off is robust to these choices. Please report results for perturbed thresholds (e.g., 0.05, 0.10 and nearby tau values) on a few representative Hilti sequences.","section":"Section V-A.3 and Eq. (4)"}],"minor_comments":[{"comment":"The text states that the LiDAR module experienced a slight increase in runtime due to code optimizations, but Table II shows the opposite: the Ours LiDAR times are lower than FAST-LIVO2 on both ARM (53.83 vs 56.50 ms) and x86 (23.36 vs 25.05 ms). Please correct the text or the table.","section":"Section V-B.2 and Table II"},{"comment":"The normalization operation in Eq. (3) is not defined. In addition, the singular values of nest(nest)^T are the squared singular values of A; please state explicitly whether the threshold is applied to normalized singular values of A or of A^T A.","section":"Eq. (3)"},{"comment":"The paper promises that the private datasets will be open-sourced, but no repository URL, dataset release statement, or code availability is provided. Since the claims concern a concrete system built on FAST-LIVO2, releasing code and configuration files would materially improve verifiability.","section":"Section I, Contribution 3"},{"comment":"The table is labeled as reporting 'mean and standard error' but the paper does not state whether these statistics are computed across sequences, across repeated runs, or both; please clarify the measurement protocol.","section":"Table II"},{"comment":"There are several typos and minor wording issues: 'fwollowed' in the opening of Section V, 'improves' in the Introduction, 'datast' in Section V-E, and 'consist' in Section IV-B.1. Also, in Fig. 9, 'Aivia LiDAR' should presumably be 'Livox Avia LiDAR'.","section":"Various"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the scope of a robotics systems venue, and the use of the authors' own FAST-LIVO2 as the baseline is appropriate. The main reasons for major revision are the unsupported headline efficiency numbers and the unvalidated translation-only degeneration metric, both of which are central to the paper's claims. The issues appear fixable within the manuscript's scope."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a practical efficiency paper on top of FAST-LIVO2, and the core idea is sound enough to deserve a real review. The two new modules—a degeneration-aware adaptive visual frame selector and a hybrid local/long-term visual map—are genuine extensions that address a real deployment problem. The experiments are the strongest part: 16 Hilti sequences, private dark/tunnel data, and real-time runs on a ~$100 ARM board, with per-frame costs around 37 ms. The efficiency gains over FAST-LIVO2 are consistent across the tables, and the ablation on the MARS-LVIG sequence shows the long-term visual map buys real accuracy (0.85 to 0.39 m RMSE) for modest memory.\n\nNow the soft spots. First, the abstract's headline numbers don't match the tables. Table II gives a runtime reduction of about 24% on ARM and 27% on x86, not 33%. Table III gives memory reduction around 32%, not 47%. That's a 10–15 point gap, too big to hand-wave. Second, and more substantive, is the degeneracy detector. The metric in Eq. (3) looks only at translation constraints from plane normals. The paper waves off rotational degeneracy as 'exceedingly rare,' but the largest per-sequence accuracy losses vs FAST-LIVO2 are in Construction Stairs (0.170 vs 0.016), Cupola (0.220 vs 0.121), and Attic to Upper Gallery (0.180 vs 0.069)—scenes the paper itself describes as involving intense rotations and sparse structures. If sigma_min stays above the 0.07 threshold while rotational constraints are weak, the selector skips visual frames exactly when they'd help, and the 3 cm average hides 6–15 cm local degradations. The paper doesn't show sigma_min traces or selected-frame ratios for those sequences, so this concern isn't answerable from the text. Third, the 0.07 threshold is hand-set and tested on only two LiDAR scenarios; calling it 'universal' is overreach. Fourth, no code or data are released, despite the abstract promising open-sourced datasets, and Tables I and III lack error bars for RMSE and memory.\n\nNone of this is fatal. The central efficiency-memory-accuracy trade-off is credible, and the adaptive selection idea is worth building on. But the authors should correct the headline numbers, add per-sequence degeneracy diagnostics, and test a richer degeneracy signal (e.g., including rotational observability) before the claims are taken at face value.\n\nSend it to review. A good referee will push on the degeneracy detector, and the paper will come out stronger for it.","headline":"Useful incremental efficiency work on FAST-LIVO2, but the headline savings are overstated and the rotation-blind degeneracy metric is the real soft spot.","tokens_in":11082,"tokens_out":3872,"would_cite":true,"duration_ms":34145,"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":"The paper claims that selectively skipping camera frames when LiDAR constraints are strong cuts per-frame runtime by about a third and memory by nearly half on the Hilti benchmark, with accuracy within 3 cm of the full system.","keywords":["LiDAR-inertial-visual odometry","degeneration-aware frame selection","error-state iterated Kalman filter","hybrid local and long-term map","resource-constrained platforms","edge computing","FAST-LIVO2","Hilti dataset"],"falsifier":"Take a handheld or robot sequence inside a large textured sphere or on a turntable with no planar structure, run both FAST-LIVO2 and this system, and compare drift: if $\\tilde{\\sigma}_{\\min}$ from Eq. (3) remains above 0.07 while most camera frames are skipped and the trajectory error grows well beyond the reported 3 cm gap, the degeneracy signal has missed an unconstrained direction and the central claim fails in that regime.","tokens_in":10181,"feed_emoji":"🤖","tokens_out":6946,"duration_ms":55074,"temperature":0.7,"pith_summary":"The paper tries to establish that the full-measurement LiDAR-inertial-visual odometry pipeline FAST-LIVO2 can be made much cheaper to run without sacrificing robustness, by spending computation only when the sensor actually needs help. Its central device is a degeneracy detector that watches a normalized singular value of the LiDAR plane-normal matrix; when that value is high, the environment already constrains the pose and the system can skip most camera frames. A second device is a map split into a small local unified LiDAR-visual map plus a sparse long-term visual map, so historical visual observations survive without keeping a huge dense map. On the Hilti benchmark the paper reports a 33 percent per-frame runtime cut and 47 percent lower memory use against FAST-LIVO2, at the cost of about 3 cm more RMSE, and demonstrates real-time operation on a roughly $100 ARM board. The significance, if correct, is that tightly coupled LIVO becomes deployable on cheap edge hardware for drones and handheld robots.","feed_headline":"Adaptive frame skipping cuts odometry runtime 33%, memory 47%","feed_subtitle":"A degeneration-aware selector and hybrid map run FAST-LIVO2 in real time on a $100 ARM board.","key_machinery":"The load-bearing machinery is a degeneration-aware adaptive visual frame selector built on the smallest normalized singular value of the translation-only plane-normal matrix $\\tilde{\\sigma}_{\\min}$ (Eq. (3)). The paper argues that rotational degeneracy with well-constrained translation is rare, so translational constraint sufficiency can stand in for overall LiDAR health. When $\\tilde{\\sigma}_{\\min}$ falls below a fixed threshold (0.07) for consecutive frames, the selector keeps every camera image; otherwise it thins images to keyframes with an adaptive threshold $\\tau = \\sqrt{3}\\,\\tilde{\\sigma}_{\\min}\\,\\tau_{\\text{predefined}}$ (Eq. (4)), spending visual computation in inverse proportion to LiDAR constraint strength. The second mechanism is the hybrid map: a small unified local map (200 m edge) for dense LiDAR-visual matching, with points that leave its boundary transferred to a sparse long-term visual map (800 m edge) that preserves historical visual observations. Together they make visual update cost and map memory scale with need rather than with sensor rate.","core_discovery":"The paper's claim is that the expensive habit of updating the ESIKF with every available camera frame is unnecessary whenever the LiDAR scan is already well constrained. By computing the smallest normalized singular value of the translation-only plane-normal matrix $\\tilde{\\sigma}_{\\min}$ from each recombined scan and comparing it to a fixed threshold, the system declares LiDAR degeneration only when that value stays below threshold for consecutive frames; in normal conditions the camera updates are thinned to keyframes whose spacing is itself scaled by $\\tilde{\\sigma}_{\\min}$ through $\\tau = \\sqrt{3}\\,\\tilde{\\sigma}_{\\min}\\,\\tau_{\\text{predefined}}$ (Eq. (4)). This converts the visual update from a fixed cost into an adaptive cost that is high only in genuinely degenerate scenes. The accompanying map change splits storage into a compact robocentric unified local map and a larger long-term visual map, with sliding windows at 20 m and 100 m and edge lengths 200 m and 800 m, so history is retained for relocalization without retaining all dense points. The paper reports the resulting trade-off on Hilti '22/'23 as 0.063 m average RMSE versus 0.034 m for FAST-LIVO2, with the 33 percent and 47 percent resource reductions, and an ablation on the HKIsland03 sequence showing the long-term visual map improves RMSE from 0.85 m to 0.39 m.","pith_inferences":["An implicit extension is that the same selector logic could be dropped into other direct or semi-direct LIVO systems that currently consume all images, since the selector only needs the LiDAR plane-normal matrix and keyframe spacing, not FAST-LIVO2-specific code.","The fixed 0.07 threshold is the main generalization risk; a self-calibrating threshold that adapts to each LiDAR's noise and field of view, or a full six-degree-of-freedom degeneracy metric, would be a natural next step the paper does not test.","The hybrid map suggests a tunable memory-accuracy Pareto frontier: shrinking the local map edge length (currently 200 m) trades accuracy for memory continuously, and the paper's numbers bracket only one operating point.","The paper leaves implicit that any environment where LiDAR translation constraints are strong but the pose is still poorly constrained in rotation, such as a large sphere or a rotating platform, is a stress test where the translation-only degeneracy assumption would likely break."],"forward_implications":["On the Hilti benchmark, per-frame total runtime drops from about 35.7 ms to 26.0 ms on x86 and from 75.9 ms to 57.8 ms on ARM; the visual part drops from 10.6 ms to 2.6 ms on x86, so the system meets 10 Hz real-time on the ARM board with about 37 ms per frame in live tests.","Memory averages 1.7 GB versus 2.5 GB for FAST-LIVO2 across Hilti sequences, enabling longer missions before exhausting edge-device RAM.","The adaptive selector keeps accuracy competitive: 6.3 cm average RMSE on Hilti versus 3.4 cm for FAST-LIVO2, and it still beats FAST-LIO2 and most LIVO baselines in the reported comparison.","The long-term visual map is load-bearing for accuracy on large sequences: removing it worsens RMSE from 0.39 m to 0.85 m on HKIsland03 with modest memory overhead.","The system runs in real time on a roughly $100 RK3588 ARM board entirely on CPU, implying LIVO can be deployed on cost- and power-constrained robots."],"supporting_citations":[{"why":"the base system: its ESIKF sequential-update state estimator and unified LiDAR-visual map are the starting point the paper modifies and the main baseline for accuracy, runtime, and memory comparisons.","marker":"[6]"},{"why":"supplies the singular-value-based degeneracy analysis from which Eq. (3) is adapted and normalized.","marker":"[13]"},{"why":"the LiDAR-inertial baseline (FAST-LIO2) that the paper must outperform to show the visual module contributes accuracy.","marker":"[3]"},{"why":"a tightly coupled LIVO baseline with an ikd-tree unified map, used in the Hilti accuracy comparison.","marker":"[8]"},{"why":"provides the adaptive voxel map registration and plane-feature estimation used to update the unified local map.","marker":"[26]"},{"why":"provides the Hilti'22 public benchmark sequences used for accuracy, runtime, and memory evaluation.","marker":"[27]"},{"why":"provides the Hilti'23 public benchmark sequences used for accuracy, runtime, and memory evaluation.","marker":"[28]"}],"fun_headline_variants":["Degeneration-aware skipping: 33% faster, 47% leaner LIVO","Resource-constrained LIVO: 33% speed bump, 47% memory cut","Edge odometry: adaptive visual frames trim runtime 33% and memory 47%","Thin-client LIVO: LiDAR-guided frame selection saves 33% time, 47% memory"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The efficiency gain rests on trusting one number, the smallest normalized singular value of the LiDAR plane-normal matrix, to mean the LiDAR pose is well constrained, and on a fixed threshold of 0.07 working across all LiDAR types; if that number stays high while some pose direction is actually unconstrained, the system skips camera frames it needed and accuracy degrades.","fun_headline_variants_meta":{"raw":{"variants":["Degeneration-aware skipping: 33% faster, 47% leaner LIVO","Resource-constrained LIVO: 33% speed bump, 47% memory cut","Edge odometry: adaptive visual frames trim runtime 33% and memory 47%","Thin-client LIVO: LiDAR-guided frame selection saves 33% time, 47% memory"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000391,"raw_usage":{"total_tokens":2103,"prompt_tokens":1036,"completion_tokens":1067,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":652,"completion_tokens_details":{"reasoning_tokens":972}},"tokens_in":652,"tokens_out":1067,"duration_ms":9244,"temperature":1.0,"reasoning_tokens":972,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:30:02.126895+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a handheld or robot sequence inside a large textured sphere or on a turntable with no planar structure, run both FAST-LIVO2 and this system, and compare drift: if $\\tilde{\\sigma}_{\\min}$ from Eq. (3) remains above 0.07 while most camera frames are skipped and the trajectory error grows well beyond the reported 3 cm gap, the degeneracy signal has missed an unconstrained direction and the central claim fails in that regime.","supporting_citations":[{"cited_title":"On degeneracy of optimization- based state estimation problems,","cited_arxiv_id":null,"evidence_quote":"supplies the singular-value-based degeneracy analysis from which Eq. (3) is adapted and normalized."},{"cited_title":"Fast-lio: A fast, robust lidar-inertial odometry package by tightly-coupled iterated kalman filter,","cited_arxiv_id":null,"evidence_quote":"the LiDAR-inertial baseline (FAST-LIO2) that the paper must outperform to show the visual module contributes accuracy."},{"cited_title":"R 3 live: A robust, real-time, rgb-colored, lidar- inertial-visual tightly-coupled state estimation and mapping package,","cited_arxiv_id":null,"evidence_quote":"a tightly coupled LIVO baseline with an ikd-tree unified map, used in the Hilti accuracy comparison."},{"cited_title":"Hilti-oxford dataset: A millimeter- accurate benchmark for simultaneous localization and mapping,","cited_arxiv_id":null,"evidence_quote":"provides the Hilti'22 public benchmark sequences used for accuracy, runtime, and memory evaluation."},{"cited_title":"The hilti slam challenge dataset,","cited_arxiv_id":null,"evidence_quote":"provides the Hilti'23 public benchmark sequences used for accuracy, runtime, and memory evaluation."}],"review_version":1}