{"id":"225e3d6e-1d61-433c-b796-5677b85767de","arxiv_id":"2506.09876","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"Aucamp combines depth-from-focus, a consensus update, and pressure-sensor orientation control into a low-cost underwater multi-robot testbed, but the 'distributed' claim is undercut by the ground-station-centric architecture described in the paper.","lead":"This paper presents Aucamp, a three-robot underwater testbed that estimates distances by scanning the focus of a monocular camera and fuses the estimates with a consensus protocol. The robots are all wired to a ground station that processes their images, which conflicts with the paper's central claim of distributed localization.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The paper's 'distributed and robust' claim is contradicted by its own architecture: Section 2 places a ground station between all robots, processes all camera images there, and links robots by cables, making the platform a star network with a single point of failure.","rationale":"The reader's verdict is REJECT with high confidence, and my analysis supports that rejection for a different reason than the stated weakest_assumption. The most load-bearing flaw is not that the depth model is empirically calibrated or that the consensus proof is omitted, but that the system described in the paper is not distributed at all: a single ground station processes all camera images and relays all inter-robot communication. This contradicts the paper's own motivation for using a distributed protocol and invalidates the core claim of being the first distributed underwater localization platform. Even if the depth model were perfect and Eq. (5) were proven to converge, the platform would remain a centralized star network with a single point of failure. I therefore set verdict_should_be to UNCHANGED because the reader already rejected the paper; my concern reinforces that rejection. I mark agreement_with_reader as disagree because the reader's weakest_assumption field identifies depth calibration and consensus convergence rather than the centralized architecture, although the reader's rationale does mention the architecture as a load-bearing flaw.","tokens_in":16099,"tokens_out":9991,"duration_ms":118650,"concrete_test":"Construct the data-flow diagram from Section 2 and Figure 2: mark every camera-image processing step and every robot-to-robot message relay, and check whether the ground station lies on every such path. Then run the Section 6.2 localization experiment with the ground station removed from both processing and communication (e.g., using on-board image processing and direct robot-to-robot links). If the robots cannot localize while maintaining the reported sub-0.4% error, the 'distributed and robust' claim is unsupported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section 2 states that each robot's control module is connected to a base station computer via cables and that the ground station serves as the communication medium among robots; camera image data is processed by the ground computer to derive object positioning data, which is then relayed to the low-level control module. Section 6 repeats that the three robots are connected to a ground computer that acts as the communication medium, forming a fully connected topology through that computer. This is a star topology, not a distributed network, and the ground station is a single point of failure for both perception and communication. The paper's own motivation in Section 1 argues that a distributed strategy is needed because loss of the central node can collapse the entire platform; the described implementation has exactly that central node. Consequently, the headline contribution that Aucamp is the first multi-robot platform systematically designed for distributed, robust localization is not supported by the presented system; only the numerical update rule in Eq. (5) is distributed, while perception and inter-robot communication are centralized.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper describes Aucamp, an underwater multi-robot platform with three cable-connected robots in a water tank. Each robot carries a low-cost monocular camera with manual focus, four pressure sensors, and eight thrusters. The authors propose monocular depth estimation by sweeping the focus position and locating the Tenengrad clarity peak, then deriving object position via the pinhole model in Eq. (4). A distributed update protocol in Eq. (5) is claimed to propagate target-position estimates through the robot network and converge to the true target position. A PI orientation controller using four pressure sensors is designed and tested under external disturbances. The paper reports a relative localization error below 0.4% and demonstrates applications in underwater 3D reconstruction and creature tracking.","tokens_in":16285,"tokens_out":5786,"duration_ms":67917,"significance":"If the central claims held, the platform would be a useful low-cost testbed for underwater multi-robot localization: the hardware is real and reasonably priced, the clarity-based ranging idea is original, and the pressure-sensor orientation controller is a practical low-cost mechanism with qualitative disturbance-rejection experiments. However, the load-bearing claims are not supported. The system architecture in Section 2 contradicts the headline 'distributed' contribution because all camera images are processed by a ground-station computer connected by cables, making perception and communication centralized. The convergence of Eq. (5) is asserted with a citation to standard consensus literature rather than proved. The depth model is calibrated and then validated on the same setup, making the reported accuracy largely a measure of curve fit. These issues affect the central claims of the paper, not merely its presentation.","major_comments":[{"comment":"The paper's headline claim of distributed localization is contradicted by the implementation it describes. Section 2 states that 'all the control modules of every robots are connected to the base station computer via cables,' that 'camera image data, which is too memory-intensive, is processed by the ground computer to derive object positioning data,' and that the ground station 'serves as a medium for communication among robots.' Section 6 repeats that the three robots are connected to a ground computer that acts as the communication medium, 'thereby forming a distributed network' in a fully connected topology. This is a star network with a single point of failure for both perception and communication, which is exactly the centralized architecture that Section 1 argues against ('the loss of connection with the central node could result in the collapse of the entire platform'). Only the numerical update in Eq. (5) is distributed; the platform as built does not support the claim of being 'systematically designed for distributed localization.'","section":"Section 2 (Control system) and Section 6 (first paragraph)"},{"comment":"The convergence claim 'according to [19], it is easy to prove ... x_i^t -> x*' is not supported by the cited reference. Reference [19] is standard average consensus with fixed or switching topologies and no persistent measurement injection. Eq. (5) is a different algorithm: it has time-varying step sizes alpha_i^t with different schedules depending on membership in the valid set S^t, it injects fresh noisy measurements xhat_i^t at every iteration rather than using them only as initial conditions, and it does not include the self-loop term x_i^t in the consensus sum for agents in S^t. Even under the Gaussian noise assumption in Eq. (7), consensus algorithms do not by themselves converge to the true position x*; one needs a proof, or at least a simulation study, for this specific update rule, including the joint effect of alpha_i^t -> 0, the expanding set S^t, and measurement noise. As written, the central convergence guarantee is asserted, not established.","section":"Section 4.2, Eq. (5)"},{"comment":"The validation of the depth model is circular. The function h(rho) in Eq. (3) is calibrated in Section 6.1 by fitting curves to optimal focus positions measured on a calibration board in the same water tank (Figure 10), and the reported R-squared = 0.99 and RMSE = 3.43 quantify the fit to those same calibration data. The depth estimation result in Figure 11 then uses that same fitted h to produce depths, and the localization experiments in Section 6.2 inherit this calibration. Consequently, the claimed relative error below 0.4% largely reflects how well the calibration board was fit, not how well the method predicts depth for unseen targets, distances, or water conditions. An independent evaluation using hold-out distances, a different object, or different water turbidity is needed before the localization accuracy claim can be accepted.","section":"Section 3.2.2 and Section 6.1, Eq. (3)"},{"comment":"The experimental evidence for the central localization claim is under-reported. The text says only that 'the relative error is less than 0.4%' and shows qualitative plots in Figure 12; no ground-truth measurement procedure, no definition of the relative error (relative to what reference length?), no number of trials, and no error bars are provided. Since the convergence claim of Section 4 is unproved, this quantitative claim carries the entire validation burden and needs to be reported with its full experimental protocol and statistics.","section":"Section 6.2"}],"minor_comments":[{"comment":"There are several typographical errors that should be corrected: 'estimzation' in Section 1.2, 'pensented' in Section 6.3, 'out multi-robot platform' in Section 7, 'Aucamp is a underwater' in Figure 1, and 'All the control modules of every robots' in Section 2.","section":"Throughout"},{"comment":"The notation in Eq. (3) is inconsistent: kappa and k are used interchangeably in the text ('h = k x f / (k x - f) + c'), and f is used both for the optical focal length and as a fitted parameter. Please define all symbols and their units explicitly.","section":"Eq. (3)"},{"comment":"The conditions on alpha_i^t are stated inconsistently: the text says 'for all i not in S^t, alpha_i^t = 0' but then says 'if N_i intersect S^t is empty, alpha_i^t = 1'; these conditions conflict for agents that are outside S^t and have no valid neighbors. Please clarify the exact update rule for all cases.","section":"Section 4.1"},{"comment":"The axes and legends of Figure 14 are not described in the text, and the claimed recovery 'in less than 2 seconds' is not visible from the data presentation alone. Please add labeled axes, units, and a clear indication of the disturbance onset and recovery time.","section":"Figure 14 and Section 6.3"}],"recommendation":"reject","confidential_remarks":"The paper's central contribution is the distributed localization claim, and the implementation described in Section 2 makes that claim untenable as written. The convergence proof and independent depth validation are also missing. The hardware platform and the pressure-sensor orientation controller may be salvageable as a centralized multi-robot localization system, but that would be a substantially different contribution from the one claimed in the title and abstract."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper is a genuine engineering effort: the authors built three identical underwater robots with monocular cameras, pressure sensors, eight thrusters, and a gimbal, and demonstrated cooperative localization of a target with a relative error below 0.4% in a 1.5m tank. The orientation control experiment - forced to instability and recovered to horizontal in under 2 seconds - is a tangible, reproducible result. The integration of Tenengrad clarity-based focus sweeping, LU2Net image enhancement, pressure-based attitude estimation, and a consensus-style update rule is new as a packaged platform; I don't know of another system that combines these particular pieces.\n\nThe central soft spot is the 'distributed' claim. Section 2 says all camera images are sent to a ground station computer over cables, which processes them and relays positions to the robots; the ground station is also the communication medium. That is a star topology with a single point of failure - exactly the architecture the introduction argues against. What is distributed is only the numerical update rule in Eq. (5). The paper should either put perception on the robots (or at least an edge computer per robot) or drop the distributed descriptor.\n\nThe depth estimation is a fitted curve from calibration on a static board in the same tank; it is not an independent prediction method for arbitrary targets. The convergence of the protocol is asserted via a citation, not proven, and the alpha schedule is vaguely specified. The experiments show no error bars, no multiple trials, and no comparison to, say, a centralized estimator or a different depth method.\n\nThese are fixable in revision. The platform is real, the integration is nontrivial, and the paper is honest about many implementation details. But as written, the headline claim overstates what is shown. I would send it to a serious referee for a systems venue, expecting major revision. I wouldn't cite it in my own work until the distributed architecture and depth validation are made credible.","headline":"A genuine underwater multi-robot testbed whose 'distributed' claim is contradicted by its own star-topology architecture, and whose depth model is calibrated rather than independently validated.","tokens_in":603,"tokens_out":1435,"would_cite":false,"duration_ms":37583,"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":"This paper claims that Aucamp, an underwater multi-robot platform built from sub-$50 monocular cameras, achieves distributed global localization with relative error below 0.4 percent by combining clarity-based depth estimation, a…","keywords":["underwater multi-robot platform","monocular camera depth estimation","clarity feature","Tenengrad function","distributed localization","consensus protocol","orientation control","underwater robotics"],"falsifier":"Take a calibrated Aucamp robot in a tank, move a textured target from 0.3 m to 1.5 m in increasingly turbid water, and compare clarity-based depth estimates with ground truth; if the fitted $h(\\rho)$ error grows beyond the reported 0.4 percent relative error or the Tenengrad peak shifts with lighting or turbidity, the central ranging claim fails.","tokens_in":15895,"feed_emoji":"🤿","tokens_out":6987,"duration_ms":63179,"temperature":0.7,"pith_summary":"This paper presents Aucamp, an underwater multi-robot platform that tries to achieve distributed localization with inexpensive monocular cameras instead of costly acoustic or active optical sensors. Each robot estimates distance to a target by sweeping the camera focus and reading the focus position where the Tenengrad clarity score peaks, then converts that distance through the thin-lens equation into a 3D position. A distributed consensus update lets the estimate spread across the robot network, so any single robot seeing the target eventually gives the whole platform a common global fix. A pressure-sensor-driven orientation controller keeps robots horizontal under disturbance, which the paper says is necessary for the camera-based estimates to remain valid. The reported result is a relative localization error below 0.4 percent in tank experiments, and the paper argues this makes low-cost underwater sensing, tracking, and 3D reconstruction practical.","feed_headline":"Underwater robot swarm localizes targets with under 0.4% error","feed_subtitle":"A $50 camera, a clarity-based focus sweep, and shared updates give each robot global target positions.","key_machinery":"The load-bearing mechanism is monocular ranging from the clarity focus sweep. For each image region $R$, the modified Tenengrad function $T(R)$ is single-peaked in focus-motor position $\\rho$; the peak locates the in-focus object distance through the calibrated relation $h(\\rho) = f\\kappa\\rho/(\\kappa\\rho - f) + c$, where $f$ is the equivalent focal length and $\\kappa, c$ are fitted camera constants. Supporting this are two further mechanisms: the distributed update rule (5), an average-consensus-style iteration with an $\\epsilon$-greedy step size that propagates and averages local estimates across the robot graph, and a PI orientation controller that maps four pressure-sensor depth readings through a linear mixing matrix to four central-thruster speeds, keeping the robot level. The camera intrinsic matrix is recalibrated underwater so that refraction effects are absorbed into the model.","core_discovery":"The central claim is that Aucamp is the first underwater multi-robot platform systematically designed for distributed localization with low-cost monocular cameras, and that combining clarity-based depth estimation, a distributed update protocol, and robust orientation control yields accurate global positioning. The depth estimate rests on the empirical relation $h(\\rho) = f\\kappa\\rho/(\\kappa\\rho - f) + c$, fitted once per camera, which maps the focus-motor position at maximum Tenengrad clarity to object distance. Distributed localization follows the update rule $x_i^{t+1} = \\sum_{j \\in N_i \\cap S_t} (1-\\alpha_i^t) w_{ij} x_j^t + \\alpha_i^t \\hat{x}_i^t$, with the step size $\\alpha_i^t$ shrinking like $O(1/t)$, so robots that never see the target still converge to its true position through neighbors. The paper reports that the full system achieves relative error under 0.4 percent and that a robot regains horizontal posture within about 2 seconds after a forced disturbance.","pith_inferences":["A natural testable extension is whether the fitted $h(\\rho)$ calibration transfers from the 1.5 m tank to longer, more turbid open-water ranges; the paper only validates the fit at tank scale.","The convergence argument for the distributed protocol is inherited from consensus theory for a static target; extending the same guarantee to moving targets would require a modified protocol or a new proof, although the tracking demonstration suggests it works empirically.","Because the experiments use tethered robots and a ground computer as the communication medium, a stronger distributed demonstration would run the same update onboard each robot with wireless or acoustic links.","The clarity-focus method may transfer to other scattering media such as fog or smoke, since clarity differences are relative features that survive where absolute image features do not."],"forward_implications":["If the claimed accuracy holds, underwater localization no longer requires expensive sonar or active optical sensors; a sub-$50 camera per robot is enough.","With at least one robot seeing the target, the distributed protocol propagates the estimate to all robots, and they converge to a common position that averages away part of the measurement noise.","The orientation controller keeps the camera stable enough for perception even after external pushes, so the localization pipeline is not limited to perfectly calm water.","The platform can be combined with NeRF for distributed underwater 3D reconstruction and with a tracking algorithm to follow marine organisms, because global target positions are shared across robots.","Because the depth estimate uses relative clarity rather than absolute color features, the same ranging approach can keep working where feature-matching and binocular stereo fail underwater."],"supporting_citations":[{"why":"Provides the average-consensus convergence result that the distributed update protocol relies on to claim all robots converge to the true target position.","marker":"[19]"},{"why":"Supplies the modified Tenengrad clarity operator used to find the optimal focus position for each image region.","marker":"[30]"},{"why":"Supplies the LU2Net real-time underwater image enhancement used to improve clarity-based depth estimation.","marker":"[34]"},{"why":"Provides the thin-lens optics and depth-of-field analysis that justify the monocular distance measurement formula and its error bound.","marker":"[10]"},{"why":"Inspires the use of multiple pressure sensors for robot attitude sensing and disturbance compensation.","marker":"[13]"},{"why":"Gives the camera calibration method used to determine the intrinsic matrix in the imaging model.","marker":"[22]"},{"why":"Establishes that underwater refraction mainly alters the intrinsic matrix and focal length, so recalibration restores the model.","marker":"[12]"},{"why":"Motivates the epsilon-greedy step-size schedule used to balance own measurements and neighbor information.","marker":"[29]"}],"fun_headline_variants":["Underwater robot swarm localizes with low-cost cameras","Distributed underwater robots localize with under 0.4% error","Cheap monocular cameras enable robust underwater robot localization","Low-cost distributed localization for underwater robot teams","Underwater multi-robot platform achieves sub-0.4% error localization"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the clarity-based focus sweep gives accurate object distance through the fitted linear model $h(\\rho) = f\\kappa\\rho/(\\kappa\\rho - f) + c$, and that this calibration, done once on a static board in a tank, remains valid for arbitrary targets, robot motion, and changing water conditions; the paper also assumes, without proving, that the consensus protocol in equation (5) converges to the true target position under the stated $\\alpha$ schedule.","fun_headline_variants_meta":{"raw":{"variants":["Underwater robot swarm localizes with low-cost cameras","Distributed underwater robots localize with under 0.4% error","Cheap monocular cameras enable robust underwater robot localization","Low-cost distributed localization for underwater robot teams","Underwater multi-robot platform achieves sub-0.4% error localization"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000589,"raw_usage":{"total_tokens":2765,"prompt_tokens":950,"completion_tokens":1815,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":566,"completion_tokens_details":{"reasoning_tokens":1732}},"tokens_in":566,"tokens_out":1815,"duration_ms":13269,"temperature":1.0,"reasoning_tokens":1732,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-07T04:38:15.310442+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a calibrated Aucamp robot in a tank, move a textured target from 0.3 m to 1.5 m in increasingly turbid water, and compare clarity-based depth estimates with ground truth; if the fitted $h(\\rho)$ error grows beyond the reported 0.4 percent relative error or the Tenengrad peak shifts with lighting or turbidity, the central ranging claim fails.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the average-consensus convergence result that the distributed update protocol relies on to claim all robots converge to the true target position."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the modified Tenengrad clarity operator used to find the optimal focus position for each image region."},{"cited_title":"LU2Net: A Lightweight Network for Real-time Underwater Image Enhancement","cited_arxiv_id":"2406.14973","evidence_quote":"Supplies the LU2Net real-time underwater image enhancement used to improve clarity-based depth estimation."},{"cited_title":"2008.Engineering optics","cited_arxiv_id":null,"evidence_quote":"Provides the thin-lens optics and depth-of-field analysis that justify the monocular distance measurement formula and its error bound."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Inspires the use of multiple pressure sensors for robot attitude sensing and disturbance compensation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Gives the camera calibration method used to determine the intrinsic matrix in the imaging model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Establishes that underwater refraction mainly alters the intrinsic matrix and focal length, so recalibration restores the model."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates the epsilon-greedy step-size schedule used to balance own measurements and neighbor information."}],"review_version":1}