{"id":"abc1957e-f264-498f-a65e-0ea5565b379e","arxiv_id":"2512.01280","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A decentralized LiDAR-swarm planner that maintains near-perfect target visibility by encoding 3-D occlusion as a Spherical Signed Distance Field and coordinating 3-D encirclement via an electrostatic-potential-inspired cost.","lead":"A decentralized swarm of LiDAR-equipped drones cooperatively tracks a moving target while keeping it visible through cluttered 3-D environments. The core inventions are a spherical signed distance field for occlusion, field-of-view costs for mixed LiDAR types, and a Thomson-problem-inspired 3-D formation cost.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The '>99% across all test conditions' headline is computed from the best of four trials per condition; all-trial statistics are needed before the robustness claim is supported.","rationale":"I read the paper's central contribution as a systems demonstration: a decentralized framework with SSDF-based visibility, FOV-aware costs, and 3-D distribution costs. The internal argument is coherent: Algorithms 2-4 are specified, the incremental SSDF update is checked against brute-force computation in Sec. VIII-E with negligible error, and the ablation in Sec. VIII-C shows each component contributes. There is no obvious mathematical inconsistency in the cost formulations or gradient derivations. The strongest weakness is in how the headline 'robust >99% visibility' is evidenced. The authors disclose that Table I uses the best of four trials; then use that table to justify a universal claim. For a robustness/fault-tolerance claim, the relevant statistic is the lower tail or at least the mean over seeds, not the maximum. Without per-trial distributions, the claim 'across all test conditions' cannot be distinguished from trial selection. Real-world experiments are qualitative with no numeric visibility ratio, so they cannot supply the missing distribution. This does not mean the system is ineffective; it means the central quantified assertion is not yet supported. The fix is straightforward: release code and weight values, report all trials, and quantify real-world visibility where possible. If all-trial results confirm >99%, the paper should be accepted; if not, the claim should be scoped. Because the reader already set CONDITIONAL, my stress test does not change that verdict, so I mark UNCHANGED.","tokens_in":26500,"tokens_out":5636,"duration_ms":60247,"concrete_test":"Re-run the Sec. VIII-A benchmark with the same released or described code and map parameters, but record all trials (at least 10 per condition with fixed seeds) and report the full distribution of ϑavg, ϑwrst, γvis, and davg — e.g., median, min, max, and the fraction of trials with γvis ≥ 99%. Make per-trial raw logs available. If the non-best trials all remain above the claimed thresholds (e.g., no trial below 95% γvis), the concern is resolved; if several trials fall below, the '>99% across all test conditions' statement must be weakened to 'best-run' or 'median-run' performance.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing weakness is in the empirical basis for the quantified central claim, not in the planner's internal logic. In Sec. VIII-A, Table I is produced by 'four independent trials, and the trial with the highest ϑavg is selected as the benchmark.' Immediately afterward the paper concludes that the proposed planner 'maintains near-complete swarm visibility (>99%) across all test conditions.' Because the reported ϑavg, ϑwrst, and γvis are conditional on choosing the best draw, they do not establish a consistent >99% performance over repeated runs, nor worst-case robustness. Robustness in the sense used in the abstract is a distributional property; a best-of-four selection is an upper envelope, not a central tendency or lower bound. This is especially consequential for 'robust cooperative tracking' and 'uninterrupted target tracking' (Sec. IX-D), because the real-world sections provide no quantitative visibility statistics at all. If the remaining trials are materially worse, the central claim reduces to 'the best run of the system exceeded 99%,' which is a materially weaker claim. The linear-extrapolation and marked-target assumptions are additional scope limits, but this benchmark-selection issue affects even the paper's own favorable evaluation setting.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a fully decentralized LiDAR-based swarm tracking framework for cooperating aerial vehicles to track a passive target in cluttered 3-D environments. The core technical contribution is a Spherical Signed Distance Field (SSDF) representation of environmental occlusion around the predicted target position, with an incremental update algorithm. On top of this, the authors derive differentiable costs for occlusion avoidance, FOV alignment for heterogeneous LiDAR configurations, tracking-distance regulation, teammate-occlusion avoidance, and a Thomson-problem-inspired 3-D swarm distribution cost. These costs are embedded in a two-stage planner comprising a kinodynamic front-end searcher and a spatiotemporal SE(3) back-end optimizer. The full system is evaluated in simulation against three prior planners and in real-world experiments with heterogeneous LiDAR swarms tracking a drone and a human runner, including dynamic agent join/leave scenarios. The paper claims 'near-complete swarm visibility (>99%) across all test conditions' in simulation and 'uninterrupted target tracking' during membership changes.","tokens_in":26775,"tokens_out":2831,"duration_ms":33135,"significance":"If the central claims hold, this would be a substantial systems contribution to multi-robot target tracking: the SSDF-based visibility metric is differentiable, sensor-agnostic, and supported by an efficient incremental update; the FOV-alignment cost explicitly handles heterogeneous LiDAR geometries; and the electrostatic-potential distribution cost is a novel way to achieve 3-D encirclement rather than the usual 2-D formations. The manuscript also has notable strengths: the derivations of the gradients in Eqs. (13)-(14), (29)-(40) are careful; the ablation study in Sec. VIII-C shows that each proposed cost contributes to the reported performance; and the incremental SSDF update is honestly validated against the brute-force ground truth in Sec. VIII-E, with negligible numerical error. The promise of open-sourced code further increases the potential impact. However, as detailed below, the empirical support for the headline robustness claim is weakened by the benchmark selection protocol and the absence of quantitative real-world visibility statistics.","major_comments":[{"comment":"The benchmark protocol states 'each test case undergoes four independent trials, and the trial with the highest ϑavg is selected as the benchmark.' The paper then concludes that the proposed planner 'maintains near-complete swarm visibility (>99%) across all test conditions.' This conclusion is not supported by the reported statistics: selecting the best of four trials gives an upper envelope, not a central tendency or a worst-case guarantee. The robustness claim in the abstract is distributional, but the table reports only the best draw. Please report all trials (or mean, standard deviation, min, and number of trials) for each condition, and state whether the >99% figure holds for every trial or only for the best one. Without this, the central robustness claim is materially weakened.","section":"Sec. VIII-A, Table I"},{"comment":"The real-world sections contain no quantitative visibility metrics. The text claims 'collisions or occlusions' did not occur, 'full visibility throughout the target's walking and running phases,' and 'uninterrupted target tracking during dynamic membership changes,' but no ϑavg, ϑwrst, γvis, or loss durations are provided. Given that the abstract and conclusion rely on 'robust cooperative tracking' and 'uninterrupted tracking,' quantitative real-world visibility statistics (or at least loss timelines for the join/leave experiment) are necessary to substantiate these claims. Qualitative snapshots are not sufficient for the strength of the stated conclusion.","section":"Sec. IX-B to IX-D"},{"comment":"The planning and evaluation rely on three strong assumptions that should be stated as limitations and preferably stress-tested: (i) target future positions are obtained by linear velocity extrapolation (Eq. 6), so a target maneuvering faster than the prediction model may invalidate the SSDF centers; (ii) the environment is treated as static during the horizon, so moving obstacles are not reflected in the SSDFs; and (iii) the target must be the largest high-reflectivity cluster in every tracker's point cloud (Sec. IX-A). These assumptions are reasonable for the demonstrated scenarios but are scope limits on the claimed robustness. Please discuss them explicitly in a limitations paragraph and, if possible, provide sensitivity experiments for prediction error or missed detections.","section":"Sec. IV-D and Sec. IX-A"},{"comment":"The Ho et al. row for the Walls scenario at 1.0 m/s lists identical values for Swarm Configuration A and Swarm Configuration B (ϑavg=3.514, ϑwrst=1.0, γvis=67.214, davg=1.988). Since the two configurations differ in sensor FOV orientations, identical results are suspicious. Please verify the data and explain whether this is a typo or an actual outcome; if it is the latter, justify why the FOV configuration has no effect on this method. This issue does not by itself overturn the main comparison, but it undermines the benchmark's credibility as reported.","section":"Table I, Walls (1.0 m/s), Ho et al."}],"minor_comments":[{"comment":"'brutal approach' should read 'brute-force approach'.","section":"Sec. IV-C"},{"comment":"Reference [39] contains typos in the original Thomson paper title ('oscilletion', 'Tomson') and should be corrected or quoted accurately.","section":"References"},{"comment":"The caption uses '(j)-(k)' twice for different phases; the in-text references '(j)-(h)' and '(l)-(m)' also appear inconsistent. Please renumber the subfigures with unique labels.","section":"Fig. 18"},{"comment":"The temporal variable substitution (Eq. 41) is stated without derivation. A brief explanation of how the map ensures the equality constraint TΣ=Tp by construction would improve readability.","section":"Sec. VII-D"}],"recommendation":"major_revision","confidential_remarks":"The paper is technically strong in its derivations and system integration, but the empirical validation of the headline claim needs revision. The best-of-four benchmark selection and the lack of real-world quantitative visibility metrics are the main barriers. If the authors can provide all-trial statistics and real-world visibility data, the paper could become a strong candidate for acceptance. I also note a dense cluster of self-citations to the group's own prior work (Swarm-LIO, ROG-Map, map sync, MPC, prior tracker); while not circular in the technical sense, the comparison would be more persuasive if independent baseline implementations or external validation were included."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a genuine systems contribution, and the SSDF-based occlusion cost is a neat idea worth building on. But the headline \"near-complete visibility (>99%) across all test conditions\" is computed from the best of four trials, so the robustness claim is overstated as written.\n\nWhat's actually new: borrowing SSDF from graphics and making it differentiable for tracking is a real adaptation, not an incremental tweak. The incremental 3-D SSDF update is properly validated against brute force (Table II), and the FOV-alignment cost that handles heterogeneous LiDAR configurations fills a real gap; prior swarm trackers assume homogeneous sensors or ignore FOV geometry. The Thomson/logarithmic potential formation cost is a sensible way to get 3-D encirclement instead of the usual 2-D ring. The real-world demos — forest drone tracking, human runner, join/leave — are genuinely difficult and credible at the level of a systems demonstration.\n\nThe math looks careful. Cost functions are smooth, gradients are derived consistently, and the ablation shows each proposed cost contributes. The incremental update is honestly compared to ground truth.\n\nWhere it's soft: the benchmark protocol. Table I says \"the trial with the highest average visibility is selected\" from four independent trials. That turns the reported 100% and 99.7% into an upper envelope, not a distributional claim. The abstract and Section VIII-A conclude \">99% across all test conditions\" — that is not supported unless all trials are shown. The real-world sections say \"full visibility\" and \"uninterrupted\" without quantitative visibility statistics; those claims need numbers or softer wording. Cost weights and thresholds are not reported, which will hamper reproduction even after the promised code release. The system assumes a high-reflectivity marker and linear target extrapolation over the 1.8 s horizon; that's a scope limit, not a fatal flaw, but it should be stated explicitly. The heavy reliance on the group's own stack is understandable, but the strongest baseline being their own prior work makes the benchmark less independent.\n\nVerdict: the planning architecture is sound and the SSDF idea is worth referee time. The fix is mostly honest reporting — give all trials or mean/std, and tone down the robustness claim until then.","headline":"Solid systems contribution with a genuinely useful SSDF visibility formulation, but the headline \">99%\" claim rests on best-of-four runs and needs honest statistics before it supports robustness.","tokens_in":27380,"tokens_out":3231,"would_cite":true,"duration_ms":33613,"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 fully decentralized LiDAR-based drone swarm can keep an agile target in near-constant view by planning around a spherical signed distance field and an electrostatic-inspired formation cost.","keywords":["swarm tracking","decentralized planning","LiDAR","visibility-aware","spherical signed distance field","field-of-view alignment","Thomson problem","aerial robotics"],"falsifier":"A real-world experiment where the target executes a hard 90° turn at 3 m/s into a blind alley, while the planner still predicts straight-line motion; if the swarm's line-of-sight is lost for more than one replanning cycle (≈66 ms), the linear-prediction assumption has been broken and the visibility guarantee fails.","tokens_in":26306,"feed_emoji":"🛸","tokens_out":5900,"duration_ms":50311,"temperature":0.7,"pith_summary":"This paper claims that a swarm of drones, with no central coordinator, can track a fast-moving human or drone through cluttered outdoor spaces while keeping the target in view almost all the time. The key insight is to encode occlusion as a Spherical Signed Distance Field (SSDF) around the predicted target position, so the cost of being in a shadow is a smooth, differentiable angular distance to the nearest visible boundary. Three differentiable costs—SSDF occlusion, field-of-view alignment for heterogeneous LiDARs, and a logarithmic Coulomb repulsion that spreads drones uniformly in 3D—drive a two-stage planner that balances visibility, safety, and swarm coordination. If the claims hold, cooperative aerial tracking becomes practical for cinematography, surveillance, and inspection without a ground station.","feed_headline":"Drone swarm keeps moving target in view 99% of the time","feed_subtitle":"Decentralized LiDAR planners use spherical visibility fields and electrostatic formation costs for near-constant tracking in clutter.","key_machinery":"The central object is the SSDF: a spherical grid around the target storing, for each ray direction and radius, the signed angular distance to the nearest visibility boundary. It is computed by a 2-D spherical distance transform followed by an incremental 3-D update that exploits the monotonic fact that along each ray, occlusion only worsens with distance. The other load-bearing costs are the FOV-alignment term, which keeps the target inside each LiDAR's unique field of view by optimizing both position and yaw, and the swarm distribution cost, a logarithmic variant of the Thomson problem that treats drones as repelling charges on a sphere around the target to achieve uniform multidirectional","core_discovery":"The paper introduces the SSDF as a spatial representation centered at the target, where each direction stores the signed angular distance to the closest visibility boundary, updated incrementally by exploiting the monotonicity of occlusion along rays. It couples this with a general FOV-alignment cost that handles upward- and downward-facing 360° LiDARs and conic Avia LiDARs by jointly optimizing the drone's position and yaw in SE(3). A swarm distribution cost, derived from the Thomson problem of electrons on a sphere, encourages the drones to form a uniform 3D encirclement, preventing mutual occlusion and maximizing angular diversity. These costs are embedded in a hierarchical planner—a kino","pith_inferences":["The SSDF representation is sensor-agnostic: with a depth camera or aggregated point cloud, the same occlusion and FOV costs could drive camera-based swarms, though the FOV geometry would need reformulation.","The benchmark's '>99% visibility' reports the best of four independent trials; a stricter average-of-trials metric would likely show lower but still high visibility, which is a caveat for expectations.","The 1.8 s prediction horizon with linear extrapolation means the method is inherently limited to targets that do not make sharp unpredictable maneuvers; a nonlinear motion model or learned predictor could extend it.","The distribution cost is a heuristic for coverage; a more principled objective might maximize mutual information of target measurements across sensors, but the electrostatic analogy is an efficient, differentiable proxy."],"forward_implications":["A four-drone swarm with mixed LiDAR orientations (upward, downward, conic) can track a drone flying at up to 3 m/s through a forest without losing line-of-sight, as demonstrated in real-world experiments.","The swarm automatically reconfigures its formation when members join or leave, maintaining tracking without a central planner.","Computation stays below 16 ms per replanning cycle for swarms up to 10 drones, suggesting scalability to larger teams on modest onboard hardware.","In simulation benchmarks the planner outperforms methods with fixed 2-D formations and 2-D visibility sectors, especially in dense obstacle fields and with fast target speeds."],"fun_headline_variants":["Swarm LiDAR drones maintain near-perfect target view","Decentralized drone swarm keeps subjects in sight","Aerial swarm tracks targets with 3D visibility optimization","Cooperative LiDAR swarm ensures continuous target visibility","Visibility-aware swarm tracks agile targets in clutter"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The system assumes the target is a passive, marker-wearing object whose next 1.8 seconds of motion are a straight-line extension of its current velocity, and that the environment around the predicted positions is static; when the target maneuvers harder than the linear prediction or a moving obstacle intrudes, the visibility maps are built at the wrong centers with stale occupancy and the near-100% visibility claim no longer holds.","fun_headline_variants_meta":{"raw":{"variants":["Swarm LiDAR drones maintain near-perfect target view","Decentralized drone swarm keeps subjects in sight","Aerial swarm tracks targets with 3D visibility optimization","Cooperative LiDAR swarm ensures continuous target visibility","Visibility-aware swarm tracks agile targets in clutter"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000377,"raw_usage":{"total_tokens":1879,"prompt_tokens":812,"completion_tokens":1067,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":556,"completion_tokens_details":{"reasoning_tokens":1007}},"tokens_in":556,"tokens_out":1067,"duration_ms":8231,"temperature":1.0,"reasoning_tokens":1007,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-03T19:14:42.250044+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A real-world experiment where the target executes a hard 90° turn at 3 m/s into a blind alley, while the planner still predicts straight-line motion; if the swarm's line-of-sight is lost for more than one replanning cycle (≈66 ms), the linear-prediction assumption has been broken and the visibility guarantee fails.","supporting_citations":[],"review_version":1}