{"id":"682be5df-0a2c-47aa-835b-3c5b617f279a","arxiv_id":"2504.18153","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":4.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A cooperative model-predictive control framework lets a UAV swarm plan drone paths to minimize position uncertainty when tracking multiple drifting castaways.","lead":"This paper designs a control system that lets a fleet of drones cooperate to keep track of several people drifting at sea. It is a step toward using autonomous drones to find and monitor castaways in search-and-rescue missions.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The NMIP objective minimizes covariance under guaranteed detections inside the FoV, but the paper's own sensor model makes detections probabilistic; planned reductions may be optimistic.","rationale":"The central claim is that the proposed NMIP-MPC yields non-myopic trajectories that minimize tracking error by reducing predicted posterior covariance. For that to hold, the predicted covariance P_{tau+k|tau} in objective (8) must be an accurate model of the estimator's future uncertainty. This requires correct associations, as the reader notes, but also correct treatment of the sensor's stochastic detection process. The manuscript provides a stochastic detection model in Secs. II-C and II-D, yet the planner's b_{ij} variables are purely geometric constraints, and Eq. (12c) creates a pseudomeasurement whenever the target is inside the FoV; the detection probability p_i^tau never enters the NMIP. Thus the planned covariance is a lower bound of the expected posterior covariance under the stated sensor model. The reported 87% average covariance reduction and 0.18 m RMSE therefore do not yet establish the central claim under 'realistic conditions' as claimed in Sec. V. This is not a disagreement with community consensus; it is an internal mismatch between the paper's sensor model and the optimization objective. The data-association concern raised by the reader is closely related but distinct: even with perfect association, the Bernoulli detection dropout remains unmodeled in the planner. I therefore agree with the conditional verdict, with the condition expanded to include planning under or validation against probabilistic detections, rather than association alone.","tokens_in":9900,"tokens_out":8789,"duration_ms":99466,"concrete_test":"Rerun the 300 Monte Carlo simulations with the Sec. II-C Bernoulli draw implemented: for each target inside an agent's FoV at altitude z, drop the measurement with probability 1-p_i(z) using the same alpha_1, alpha_2, and p_min parameters. Compare the realized average covariance trace and RMSE to the reported 87% reduction and 0.18 m figure. If they degrade materially, modify the objective to minimize the p_i-weighted expected covariance and check whether the planned reductions are recovered.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The most load-bearing gap is that the objective (8) is minimized over covariance updates that assume detections are guaranteed whenever a target is geometrically inside the FoV. Constraints (10c)-(10e) define b_{ij} purely from the planned footprint; Eqs. (11c)-(11g) apply a full Kalman correction whenever b=1; and Eq. (12c) fabricates a pseudomeasurement y=C\\hat{p}+n(\\sigma) for every in-FoV target. The observation probability p_i^\\tau from Eq. (5), and the Bernoulli emptiness of H_i^\\tau in Sec. II-C, do not appear anywhere in the NMIP. Under the paper's own sensor model the expected posterior covariance is p_i P_update + (1-p_i) P_prior, which is larger than the deterministic P_update used in Eq. (8). Hence the planned '87% average covariance reduction' and the 0.18 m RMSE from Fig. 6 may be optimistic whenever the UAV flies at altitudes where p_i<1. The simulation setup (Sec. IV-A) states only that Gaussian noise is added to ground truth; it does not say the Bernoulli detection draw is executed, so either the simulator also ignores the paper's detection model or the controller solves the wrong problem. Data association is a separate oracle, acknowledged in Sec. IV-A; even with perfect association this stochastic-detection mismatch persists.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a collaborative NMIP-based MPC framework for a UAV swarm tracking multiple castaways in maritime environments. Each UAV runs local Kalman filters for the targets, fuses estimates via covariance intersection, and solves a receding-horizon nonlinear mixed-integer program that minimizes the sum of predicted posterior covariance traces over a finite horizon. Binary variables encode whether each target is inside a UAV's field of view, and the KF correction is applied accordingly, with pseudomeasurements generated for in-field targets. A deterministic clustering scheme groups targets and assigns UAVs greedily. The numerical section reports Monte Carlo simulations with up to 10 UAVs and 3-5 castaways, an 87% covariance reduction for a two-UAV fleet relative to the one-UAV case, an empirical CDF of inter-agent distances satisfying the safety constraint, and an illustrative three-target/two-UAV scenario with an average RMSE of 0.18 m. The paper concludes that the method consistently and accurately tracks castaways under realistic conditions.","tokens_in":10233,"tokens_out":5428,"duration_ms":54569,"significance":"If the validation gaps were closed, the framework would be a useful contribution to multi-UAV maritime search-and-track: it couples non-myopic planning with KF covariance minimization, explicitly handles limited sensor field of view through mixed-integer constraints, and uses a physically motivated Stokes-drift target model. The formulation is coherent and the simulation effort, including Monte Carlo fleet-size sweeps and collision-avoidance checks, is commendable. However, the paper's central performance claims rest on the controller's own covariance prediction rather than on independent tracking-error evaluation, and the simulation does not exercise the paper's own probabilistic detection model. The acknowledged idealization of data association further limits the claim of 'realistic conditions.' The work is a plausible stepping stone, but the evidence presented is not yet sufficient to support the headline quantitative claims.","major_comments":[{"comment":"The MPC formulation assumes a detection is guaranteed whenever a target is geometrically inside the FoV. Specifically, constraints (10c)-(10e) set b_{ij}=1 based solely on geometry, and constraints (11c)-(11g) apply a full Kalman correction whenever b=1. The probabilistic detection model of Eq. (5), and the possibly empty random set H_i^tau of Section II-C, do not appear anywhere in the optimization or in the pseudomeasurement generation of Eq. (12c). Under the paper's own sensor model, the expected posterior covariance is p_i P_update + (1-p_i) P_prior, which is strictly larger than the deterministic P_update used in Eq. (8). Section IV-A states only that Gaussian noise is added to ground truth and gives no indication that the Bernoulli detection draw is executed. Thus both the controller and the simulator appear to solve a different problem from the one defined by the sensing model, and the reported 87% covariance reduction and 0.18 m RMSE may be optimistic whenever the UAV flies at altitudes where p_i<1. The authors should either incorporate the detection probability into the planning (e.g., as a probabilistic correction or chance constraint) or justify, with evidence, that p_i=1 throughout the simulated flight envelope.","section":"Section III-B, Eqs. (10)-(12) vs. Section II-D, Eq. (5)"},{"comment":"The statement 'Data association was not considered, assuming agents can differentiate between targets' is explicitly acknowledged but is load-bearing for the central claim. Every KF update in Eqs. (11d)-(11g) and every pseudomeasurement in Eq. (12c) presupposes that an incoming measurement is assigned to the correct castaway. In a maritime scene with multiple people, clutter, and intermittent detections, association errors produce wrong innovations that can drastically inflate true tracking error, even if the planned covariance is low. The paper's conclusion that tracking is 'under realistic conditions' is therefore not supported by the current experiments. I recommend adding a data-association module (e.g., nearest-neighbor or JPDA) to the simulation, or at minimum a sensitivity analysis with misassociation rates, before the multiple-target tracking claim can be accepted.","section":"Section IV-A, data association assumption"},{"comment":"The primary scalability result (Fig. 4) plots the average target covariance against fleet size, but this is the same covariance that the MPC minimizes in Eq. (8). Consequently, the reported 87% reduction partly reflects the internal consistency of the planner's own model rather than an independently measured improvement in tracking accuracy. The paper reports a ground-truth-based RMSE only for a single scenario (C=3, N=2), and no comparison is made with a myopic planner, a greedy baseline, a single-agent sweep, or an existing tracking controller. Without such baselines and without RMSE curves across the fleet-size and target-count grid, the 'significant benefits from the multi-agent approach' are not demonstrated. I ask the authors to report RMSE (or another ground-truth error metric) for the full simulation grid and to include at least one baseline controller.","section":"Section IV-B, Fig. 4 and Eq. (8)"}],"minor_comments":[{"comment":"The displayed matrices in Eq. (2) are inconsistent with the declared dimensions: A is stated to be in R^{6x6} and B in R^{6x3}, but the printed blocks imply A in R^{3x12} and B in R^{3x6}. Please correct the block-matrix notation or the dimension statement.","section":"Eq. (2)"},{"comment":"The text says constraint (13) ensures the inter-agent distance is 'at most d_t', while Eq. (14b) enforces d_i >= d_t. The correct inequality should be 'at least d_t'; please fix the wording.","section":"Eq. (13) and Eq. (14b)"},{"comment":"The function r(z_i) in Eq. (12a) is never explicitly defined, and its link to the piecewise observation probability of Eq. (5) is only described as 'mimicking.' A concrete definition (or a statement that r(z) is a calibrated approximation with its own parameters) is needed for reproducibility.","section":"Eq. (12a) and Eq. (5)"},{"comment":"References [5] and [25] appear to be the same publication ('Model Predictive Control For Multiple Castaway Tracking with an Autonomous Aerial Agent,' ECC 2023). Please consolidate the duplicate citation and clarify in the text which results come from that prior paper versus the present manuscript.","section":"References [5] and [25]"}],"recommendation":"major_revision","confidential_remarks":"The manuscript extends the authors' prior ECC 2023 work [5]/[25] to the multi-agent case, but the novelty boundary is not crisply drawn: the sensing-model validation cited as [25] is the same prior work, and the present paper's probabilistic sensing model is not used in the controller or the simulator. This, together with the lack of any baseline comparison, makes the quantitative claims difficult to evaluate. If the authors can amend the experiments to include Bernoulli detections, a data-association mechanism, and RMSE-based baselines, the contribution would be substantially strengthened; in the current form, I would not advise acceptance."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea is a fair extension of the authors' prior single-agent MPC: a sequential NMIP formulation for a UAV swarm that adds target clustering, covariance-intersection fusion, pseudomeasurement updates, and collision constraints. The equations are mostly coherent and the 300-run Monte Carlo setup is reasonable for a simulation-only paper. Credit is due for the deterministic clustering and for reporting collision-avoidance behavior separately.\n\nThe soft spots are real, and one is load-bearing. The MPC objective in Eq. (8) minimizes the trace of the KF covariance under the assumption that every in-FoV target produces a measurement. The binary variable b_{ij} in Eq. (10) is a geometric FoV check; Eqs. (11c)–(11g) apply a full KF correction whenever b=1, and Eq. (12c) fabricates a pseudomeasurement for every such target. But the paper's own sensing model (Sec. II-C/D) says the measurement is probabilistic: the random set H_i^tau is empty with probability 1 - p_i^tau. That p_i^tau never appears in the planning problem. The simulator description (Sec. IV-A) also only mentions additive Gaussian noise, no Bernoulli draw, so either the simulator silently ignores the detection model or the controller solves the wrong problem. The claimed 87% average-covariance reduction and 0.18 m RMSE are therefore optimistic whenever p_i < 1, which is exactly the regime the altitude-dependent model is meant to capture.\n\nData association is a separate acknowledged gap; I would not penalize the paper heavily for that, since it is stated plainly. More damaging is the absence of any baseline comparison. No comparison to the single-agent version, a myopic controller, or a random-walk policy, so you cannot tell how much of the improvement comes from the non-myopic horizon versus the swarm size. The paper also ships no code and omits parameter values needed for reproduction, which is a minor but fixable issue.\n\nThe clustering algorithm is described verbally rather than formally, but I could follow it. The drift model via Stokes' equation is a nice touch. Citation pattern is fine.\n\nWho is this for? Researchers working on SAR tracking with small UAV fleets, especially those building on the authors' earlier ECC work. It would benefit from a serious referee process, but the referee should require the authors to either incorporate the detection probability into the planning or explicitly simulate it, add a baseline, and release code/parameters. The central formulation is worth engaging with, but the evidence as presented does not support the conclusion that the method tracks castaways under realistic conditions.","headline":"Useful multi-agent MPC extension of the authors' own single-agent tracker, but the simulator ignores the paper's own probabilistic detection model, so the headline covariance reductions are likely optimistic.","tokens_in":10733,"tokens_out":2153,"would_cite":false,"duration_ms":25946,"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":"The paper claims that a fleet of camera drones can track drifting castaways by planning flight paths to minimize predicted position uncertainty, with simulations showing 0.18 m average error.","keywords":["Search and Rescue Robots","Model Predictive Control","Multi-agent Systems","Target Tracking","Kalman Filter","UAV Swarm","Castaway Tracking","Maritime Environment"],"falsifier":"Run the same controller in a simulation where two castaways cross paths or come within a few meters, with detections generated by an actual detector rather than noise added to known identities, and compare RMSE and covariance to the paper's 0.18 m and 87 percent figures; if association errors dominate, the planned reductions will not materialize. More directly, an experiment with real cameras over water where targets swap positions would show whether the assumed distinguishability holds.","tokens_in":9721,"feed_emoji":"🚁","tokens_out":4518,"duration_ms":44898,"temperature":0.7,"pith_summary":"This paper tries to establish that a small fleet of camera-equipped drones can keep precise track of several people drifting at sea by planning their flight paths several steps ahead instead of reacting to the latest measurement. The controller, a nonlinear mixed integer program solved in a receding horizon, chooses drone positions and altitudes that are predicted to minimize the Kalman-filter covariance of every castaway's estimated position. The authors argue that this collective, uncertainty-aware planning is what matters: simulated trials with two drones and three castaways produce an average root-mean-square position error of 0.18 meters, and adding a second drone already reduces average estimation covariance by 87 percent compared with a single drone. If the approach carries into real operations, it would let a rescue coordinator send a few drones to hold a drifting group in view with quantified confidence.","feed_headline":"UAV swarm tracking cuts castaway error to 0.18 meters","feed_subtitle":"A receding-horizon plan shrinks predicted position uncertainty; two drones cut covariance by 87 percent","key_machinery":"The load-bearing object is the receding-horizon NMIP in Eqs. (9)-(14), whose objective is the sum over the horizon of the traces of the predicted Kalman covariance matrices of all targets in an agent's assigned cluster. It couples a linear drone dynamics model, a field-of-view model via binary variables, Kalman-filter prediction and update equations with pseudomeasurements and intermittent-observation gating, altitude-dependent measurement noise, and collision-avoidance constraints. This machinery lets the controller optimize not just where to look but what it expects to see, converting sensor placement into an explicit reduction of state uncertainty.","core_discovery":"The central discovery is that the tracking problem can be cast as a single optimization over predicted posterior covariance: at each planning cycle, each agent solves for its control sequence over a short horizon by simulating what Kalman updates would do if future camera measurements arrive, using pseudomeasurements and binary indicators of field-of-view coverage, and then minimizes the sum of covariance traces over the horizon. Because each agent plans sequentially with access to the others' planned trajectories and fuses shared estimates by covariance intersection, the fleet's motions are coordinated without a central optimizer. The paper reports that the resulting trajectories keep castaways inside the sensors' field of view at beneficial altitudes, and that in simulation the method consistently tracks three to five drifting castaways with one to ten agents, with the largest gains coming from the first few agents.","pith_inferences":["The paper leaves implicit that its 0.18 m RMSE should be read as an upper bound on achievable precision: real detections would include association errors, which would feed wrong innovations into the same Kalman equations and likely worsen the reported numbers.","Minimizing covariance trace is a natural proxy for search-and-rescue success, but it is not identical to minimizing expected search time or maximizing survival probability; weighting targets by urgency or detection priority would be a direct extension of the same objective.","Because the planner is built on a Stokes-drift model of castaway motion, its promised reductions are only as good as that drift model; coupling the planner with online drift estimation or measured current data could make the non-myopic plans robust to model mismatch."],"forward_implications":["With two agents, average estimation covariance drops by 87 percent relative to one agent, and a third agent gives a further 10 percent decrease; beyond a certain fleet size for five targets, additional drones add little.","Agents maintain the prescribed safety distance in all simulated fleet sizes from two to ten, so the collaborative planning does not sacrifice collision avoidance.","The controller adapts group assignments: it lets one agent shadow a split-off castaway while another keeps the remaining castaways in view, balancing altitude for coverage versus precision.","Quantified uncertainty from the Kalman filter gives a mission coordinator a predicted confidence measure for where survivors are, not just a point estimate.","Intermittent observations are handled inside the planner itself, so the control law is designed for the reality that castaways will periodically leave the camera's field of view."],"supporting_citations":[{"why":"Supplies the altitude-dependent observation-probability model that the sensing model and measurement-noise constraints adopt.","marker":"[25]"},{"why":"Provides the Kalman-filtering-with-intermittent-observations method used to gate updates with binary field-of-view variables.","marker":"[29]"},{"why":"Provides the Stokes-drift model used to simulate ground-truth castaway motion under wave conditions.","marker":"[23]"},{"why":"Supplies the Kalman filter that underlies the estimator and the planned covariance updates.","marker":"[24]"},{"why":"Gives the distributed model-predictive-control scheme used for sequential planning with neighboring agents' plans.","marker":"[26]"},{"why":"Supplies the covariance-intersection technique used to fuse target estimates when cross-correlations are unknown.","marker":"[27]"},{"why":"Earlier single-agent MPC formulation that this work extends to a collaborative multi-agent setting.","marker":"[5]"}],"fun_headline_variants":["Swarm MPC coordinates drones to shrink multi-target tracking error","UAV swarms use predictive control to track castaways accurately","Collaborative MPC coordinates drone swarm for precise target tracking","Drone swarm coordination reduces uncertainty in castaway tracking","Receding-horizon MPC guides UAV swarms to track castaways"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole estimator assumes that whenever a camera detects a person, the drone knows which castaway that person is, so every measurement is fed to the correct Kalman filter; in a real scene with several people, clutter, and missed detections, mis-assigned measurements would corrupt the very covariance reductions the controller plans.","fun_headline_variants_meta":{"raw":{"variants":["Swarm MPC coordinates drones to shrink multi-target tracking error","UAV swarms use predictive control to track castaways accurately","Collaborative MPC coordinates drone swarm for precise target tracking","Drone swarm coordination reduces uncertainty in castaway tracking","Receding-horizon MPC guides UAV swarms to track castaways"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000748,"raw_usage":{"total_tokens":3299,"prompt_tokens":882,"completion_tokens":2417,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":498,"completion_tokens_details":{"reasoning_tokens":2334}},"tokens_in":498,"tokens_out":2417,"duration_ms":18918,"temperature":1.0,"reasoning_tokens":2334,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T10:23:33.320520+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same controller in a simulation where two castaways cross paths or come within a few meters, with detections generated by an actual detector rather than noise added to known identities, and compare RMSE and covariance to the paper's 0.18 m and 87 percent figures; if association errors dominate, the planned reductions will not materialize. More directly, an experiment with real cameras over water where targets swap positions would show whether the assumed distinguishability holds.","supporting_citations":[{"cited_title":"Model predictive control for multiple castaway tracking with an autonomous aerial agent,","cited_arxiv_id":null,"evidence_quote":"Supplies the altitude-dependent observation-probability model that the sensing model and measurement-noise constraints adopt."},{"cited_title":"Kalman filtering with intermittent observations,","cited_arxiv_id":null,"evidence_quote":"Provides the Kalman-filtering-with-intermittent-observations method used to gate updates with binary field-of-view variables."},{"cited_title":"Theoretical study of drift of small rigid floating objects in wave fields,","cited_arxiv_id":null,"evidence_quote":"Provides the Stokes-drift model used to simulate ground-truth castaway motion under wave conditions."},{"cited_title":"An introduction to the kalman filter,","cited_arxiv_id":null,"evidence_quote":"Supplies the Kalman filter that underlies the estimator and the planned covariance updates."},{"cited_title":"Robust distributed model predictive control,","cited_arxiv_id":null,"evidence_quote":"Gives the distributed model-predictive-control scheme used for sequential planning with neighboring agents' plans."},{"cited_title":"General decentralized data fusion with covariance intersection,","cited_arxiv_id":null,"evidence_quote":"Supplies the covariance-intersection technique used to fuse target estimates when cross-correlations are unknown."},{"cited_title":"Model Predictive Control For Multiple Castaway Tracking with an Autonomous Aerial Agent,","cited_arxiv_id":null,"evidence_quote":"Earlier single-agent MPC formulation that this work extends to a collaborative multi-agent setting."}],"review_version":1}