{"id":"03552238-b57a-4407-8af2-e28522cab213","arxiv_id":"2510.03534","paper_version":5,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Multi-agent deep Q-learning with spatiotemporal Gaussian process regression yields energy-aware multi-day mapping of the Douro plume in simulation, beating several baselines but not on every test month.","lead":"A team trained a multi-agent reinforcement learning controller that guides underwater robot fleets to map the Douro River plume over multiple days, using Gaussian-process salinity estimates and intermittent communication. In Delft3D simulations the learned policies often beat baseline sampling strategies and extend mission endurance, but the paper's own table contains a counterexample to its 'consistently outperforms' claim.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Table I contradicts 'consistently outperforms': on Oct'17 with N=6 the [32] baseline achieves lower MSE (2.04 vs 2.79), and no visible row shows endurance doubling when scaling 3→6.","rationale":"The reader's verdict of CONDITIONAL is appropriate. My stress-test identifies a more specific, internally checkable weakness than the simulator-fidelity concern that the reader lists as the weakest assumption. The paper's own Table I contains a direct counterexample to the 'consistently outperforms' claim: in Oct'17 with N=6, the [32] baseline MSE (2.04) is lower than the proposed method's (2.79). This is a concrete internal inconsistency that would need to be resolved—either by correcting the table, clarifying the evaluation horizon, or softening the claim.\n\nI also checked the endurance claim. Abstract says 'in some instances' endurance more than doubles, but the Contributions section states it as an example ('for example scaling from 3 to 6 vehicles more than doubling mission lifetime') without pointing to a specific table row or figure value. The visible rows of Table I do not support it, and Mar'18 even shows decreased endurance with more agents. This is a second unsupported sub-claim.\n\nI did not find grounds to recommend REJECT: the method is plausible, most test months favor it, and the GP+multi-head DQN architecture is a reasonable contribution. The absence of error bars and code are standard reproducibility weaknesses, not evidence of bad faith. A CONDITIONAL verdict—requiring code release, seeded reruns, and clarified evaluation protocol—is the right level. My concern therefore leaves the reader's verdict unchanged rather than moving it.\n\nI flagged no methodology artifacts: the garbled '/uni...' strings in the extracted Table I are likely PDF encoding issues, but I did not rely on any invisible rows. The contradictory evidence is already present in the visible, intelligible part of the table.","tokens_in":13178,"tokens_out":6882,"duration_ms":61081,"concrete_test":"Request the authors' code/checkpoints and evaluation scripts, then rerun Table I exactly as specified but with a fixed evaluation horizon for all algorithms and at least 5 random seeds per configuration. If any test month/fleet size shows the [32] baseline's mean MSE below Plume-DQN-GP's (as in Oct'17 N=6), the 'consistently outperforms' claim must be revised. Separately, identify the precise η2 setting and figure that produces the 3→6 endurance doubling and report the two endurance values; if no such configuration exists, remove claim (iii) from the Contributions.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim (ii) — \"consistently outperforms baselines and benchmarks in both single- and multi-agent settings\" — is directly contradicted by the paper's own summary table. In Table I (Policy Generalization Across Seasonal Regimes), the row for October 2017 with N=6 reports Ours MSE = 2.79 while the baseline [32] MSE = 2.04, i.e., the baseline is strictly better. Because the table is presented as the evidence for cross-month and cross-year generalization, this single cell undermines the word \"consistently.\"\n\nClaim (iii) — \"scaling from 3 to 6 vehicles more than doubling mission lifetime\" — is also not supported by the visible table. In no row does N=6 endurance exceed twice N=3 endurance; in March 2018 the endurance decreases from 3.5 to 3.1 days. The abstract hedges with \"in some instances,\" but the Contributions section promises a concrete example, and the manuscript does not provide the specific configuration or figure where the doubling occurs.\n\nAdditionally, the evaluation protocol is underspecified: the table lists endurance only for the proposed method, not for baselines, and the comparison horizon for MSE is not defined. Without error bars, seed counts, or a fixed mission duration for all algorithms, the reported numbers—especially a 2.79 vs 2.04 gap—cannot be distinguished from stochastic variation.\n\nThe simulator-fidelity concern raised by the reader is real but secondary; the first-order issue is that the paper's own reported evidence does not uniformly support the headline performance claims.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper addresses long-term mapping of the Douro River plume with a fleet of AUVs. It formulates the problem as minimizing time-averaged MSE over a grid subject to trajectory constraints and ocean-current dynamics (Sec. II). The proposed solution combines a GPR estimator with a separable spatiotemporal kernel fit to historical Delft3D data (Sec. III-A) and a multi-head DQN that selects direction and speed for each AUV (Sec. III-B). A central server communicates intermittently with the AUVs. Experiments in the Delft3D simulator compare against uniform sampling, ideal rotations, an ideal EIBV benchmark [18], and an adaptive Voronoi benchmark [32] (Sec. IV). The authors claim consistent outperformance, generalization across unseen months/years, and endurance gains from scaling the fleet.","tokens_in":13640,"tokens_out":4977,"duration_ms":40225,"significance":"If substantiated, the method would be a meaningful contribution to adaptive sampling and multi-agent reinforcement learning for environmental monitoring. The architecture is sensible, the effort to fit the kernel to empirical correlations is valuable, and the comparison against EIBV and Voronoi baselines is commendable. However, the present evidence is not sufficient to support the central claims: one cell of Table I directly contradicts 'consistently outperforms', no error bars or seed statistics are provided, and the endurance-doubling example is not reported with numeric data. These are correctable with additional experiments and more careful phrasing.","major_comments":[{"comment":"In the October 2017 row with N=6, the proposed method's MSE is 2.79 while the Voronoi baseline [32] achieves 2.04. This directly contradicts Contribution (ii), 'consistently outperforms baselines and benchmarks in both single- and multi-agent settings.' Since Table I is the primary evidence for cross-season generalization, either the claim must be weakened or the table must be augmented with repeated-seed statistics and a statistical test showing the gap is within noise.","section":"Section IV, Table I"},{"comment":"No error bars, seed counts, or variance information are provided for any RL result. Figures 9 and 10 report box plots, but the underlying distributions, numbers of runs, and training seeds are unspecified. Without this, the reported differences (including the 2.79 vs 2.04 gap) cannot be distinguished from stochastic variation. The evaluation protocol should be described: number of training seeds, number of evaluation episodes, how MSE is aggregated over grid and time, and whether endurance values in Table I come from a single run.","section":"Section IV, Multi-agent performance"},{"comment":"The contribution states that scaling from 3 to 6 vehicles more than doubles mission lifetime. Table I (η2=50) shows no such doubling; for example, Oct ’17 endurance goes from 13.0 to 15.6 days, and Mar ’18 endurance decreases from 3.5 to 3.1 days. The only support is the Fig. 11 caption. The authors should report the exact configuration (η2 value, month) where the doubling occurs and provide the numeric data behind that figure, otherwise the endurance-scaling claim is unsupported.","section":"Section IV, Fig. 11 and Table I"},{"comment":"The tuning of baselines is described only qualitatively: 'We empirically tune the exploration-exploitation hyperparameters in [32]' and the threshold for [18] is selected as best from {28,30,32,34}. To ensure fair comparison, the selected hyperparameter values and the criterion for 'best' should be reported. This is load-bearing because the outperformance claims in Contributions (ii) and (iii) rest on these comparisons.","section":"Section IV, Baselines"}],"minor_comments":[{"comment":"The text near Table I contains garbled strings such as '/uni00000015/uni00000016/...' which appear to be rendering artifacts. These must be removed or fixed before publication.","section":"Table I"},{"comment":"Endurance columns are reported only for the proposed method, not for the baselines, so the reader cannot evaluate the endurance comparison. Please include baseline endurance or state explicitly that baselines were run with a fixed mission duration.","section":"Table I"},{"comment":"The memory window M is introduced but its value (24 frames) appears only in the baselines description. Specify M for all experiments in the training setup.","section":"Section III-A"},{"comment":"The temporal kernel parameter T0 is described as the tidal period. Clarify whether T0 is fixed a priori or fitted to the data, and how its value is determined.","section":"Section III-A, Eq. (13)"},{"comment":"The word 'consistently' in the abstract and Contributions overstates the evidence. Consider hedging with 'in the majority of test cases' or similar, until statistical evidence is available.","section":"Abstract and Contributions"},{"comment":"The paper should explicitly remind readers in the abstract and conclusion that all results are internal to the Delft3D simulator and that no in-situ validation is presented; the current phrasing 'generalizes to unseen conditions' is limited to the simulator.","section":"Section IV"}],"recommendation":"major_revision","confidential_remarks":"The manuscript addresses a timely and relevant problem, and the algorithmic idea is plausible. The main issues are overstated performance claims and lack of statistical rigor, both of which are addressable in revision. I do not see a fundamental flaw that would require rejection, but the central claims (ii) and (iii) must be re-supported with proper experiments and calibration."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things you should know. First, this is a genuinely useful system: spatiotemporal GPR combined with a multi-head DQN to control a fleet of AUVs over multiple days, with an explicit energy-reward term. The multi-day, endurance-aware setting is new for plume mapping, and the Delft3D simulations are a credible testbed. Second, the headline claim outruns the evidence. The paper says it 'consistently outperforms' baselines, but Table I shows the Voronoi benchmark [32] beating it on October 2017 with N=6 (2.04 vs 2.79 MSE). That's a direct contradiction of a central claim, not a minor wrinkle.\n\nWhat it does well: the GP kernel is fit to historical data, including a tidal-period term, and the reward function splits global MSE and per-agent credit, which addresses a real credit-assignment issue. The comparisons against an 'ideal' EIBV baseline [18] and a Voronoi benchmark [32] are appropriate, and in most tested months the method is better. The energy-speed tradeoff, with two cruise speeds and a fleet-level penalty, is a sensible way to handle endurance.\n\nWhere it's soft: no error bars or seed counts on any RL result, so we can't tell if the reported gaps are meaningful. The baselines are 'empirically tuned' and the best results shown, which is test-set tuning. The evaluation protocol is underspecified: no fixed mission horizon for MSE, and endurance is reported only for the proposed method, not the baselines. The endurance-doubling claim isn't visible in Table I—no row has N=6 endurance as double the N=3 value, and March 2018 actually decreases. The simulator-fidelity issue is real but secondary; every result is internal to Delft3D, so generalization is within the model, not to the physical ocean.\n\nThe paper is not weak in conception. The formulation is valuable, and the method plausibly works. But the claims as written are stronger than the data. A revision with seed statistics, a fixed evaluation protocol, and a softened or properly supported 'consistently outperforms' claim could make this a solid contribution.\n\nFor you: this is worth bringing to a reading group if you work in marine robotics or adaptive sampling, mainly to discuss what evidence is needed for 'consistent' improvement. I wouldn't cite it until the claims are backed up. But I'd send it to serious peer review—the core idea deserves referee time, and the flaws are fixable.","headline":"The multi-day, energy-aware plume mapping formulation is genuinely useful, but the paper's own Table I contradicts the central 'consistently outperforms' claim.","tokens_in":14102,"tokens_out":2323,"would_cite":false,"duration_ms":19079,"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":"A learned multi-agent policy maps the Douro river plume for days, cutting error roughly in half while more than doubling fleet endurance when scaled from 3 to 6 vehicles.","keywords":["river plume mapping","multi-agent reinforcement learning","autonomous underwater vehicles","Gaussian process regression","long-term environmental monitoring","Delft3D simulation","energy-aware navigation","Douro river"],"falsifier":"Compare the learned policy's real-world salinity maps against independent CTD casts during a high-discharge Douro event. If the map MSE is no better than a simple lawnmower survey, or the fleet runs out of energy far sooner than the simulated 3-6 day missions, the central claim does not transfer to the ocean.","tokens_in":13107,"feed_emoji":"🌊","tokens_out":4331,"duration_ms":35391,"temperature":0.7,"pith_summary":"The paper tries to establish that a small fleet of energy-limited AUVs can keep a useful, low-error map of a fast-moving river plume for multiple days if their headings and speeds are chosen by a central learned controller rather than by fixed coverage patterns or myopic uncertainty sampling. It claims that a multi-agent deep reinforcement learning policy, fed with Gaussian-process estimates of the salinity field, consistently beats both physically unrealizable uniform-random sampling and the best published front-tracking and Voronoi coverage baselines. It further claims that increasing the fleet from 3 to 6 vehicles can more than double mission endurance while improving accuracy, because the policy learns when the slower cruise speed suffices. If true, this is a path to practical long-term autonomous monitoring of dynamic coastal features.","feed_headline":"Learned AUV fleet maps Douro plume for days at roughly half the error","feed_subtitle":"RL coordinates 3-6 underwater vehicles to cut map error and more than double mission endurance","key_machinery":"The load-bearing component is the multi-head Q-network: a shared state-value head feeds two advantage heads, one over eight directions and one over two speed levels, so the server can command heading and speed separately for each AUV. Around it sits a spatiotemporal Gaussian process whose kernel is separable in space (exponential) and time (linear decay plus a 12.5-hour tidal oscillation fitted to historical data), and a reward that adds individual credit to global MSE to fix the credit-assignment problem. Together they convert raw salinity measurements into commands over 30-minute slots with only intermittent, roughly 160-byte uplinks.","core_discovery":"On the paper's own terms, the central claim is that long-horizon plume mapping reduces to a sequential decision problem that a centralized multi-head Q-network can learn to solve, provided the state encodes a spatiotemporal Gaussian-process regression map plus recent trajectory traces and wind, and the reward mixes global MSE, a salinity-contrast term, a fleet-level speed penalty, and per-agent credit for sampling where the previous estimate was wrong. The authors show in Delft3D simulations that this learned policy tracks the Douro plume over days, generalizes to months and years not used in training, and outperforms single-agent EIBV front tracking and multi-agent Voronoi coverage, with fl","pith_inferences":["A direct test the authors do not report: deploy the same policy on real LAUVs during a Douro high-flow event and compare CTD-verified maps against the simulated MSE; that would separate simulator generalization from real-ocean generalization.","Because the method is agnostic to the measured scalar, the same server-AUV loop could map temperature, turbidity, or dye plumes, provided the GP kernel is refit to the target process's correlation scale and period.","The reward decomposition (global MSE + individual credit + fleet speed penalty) is a reusable recipe for other multi-vehicle monitoring tasks where energy and communication are the binding constraints.","The fitted tidal-period kernel hints that encoding known environmental cycles into the surrogate is a cheap way to improve long-horizon estimates; that likely transfers to any tidally forced coastal process."],"forward_implications":["Even a single AUV with this policy keeps the plume in view over days, whereas the EIBV front-tracking benchmark loses track over long horizons.","With 3 or 6 vehicles, the learned policy produces roughly half the mapping MSE of adaptive Voronoi coverage on the March 2018 test month.","Raising the fleet from 3 to 6 vehicles can more than double mission endurance, in addition to improving accuracy.","Varying one reward weight traces an endurance-accuracy frontier, so operators can dial in the desired trade-off.","Generalization across unseen months and years (2016-2018) suggests the policy captures persistent dynamics rather than overfitting one season."],"fun_headline_variants":["RL-driven AUVs map Douro plume for days with lower error","Multi-agent RL maps Douro plume longer and more accurately","AUV fleet learned to double endurance while mapping Douro plume","Douro plume mapped for days via multi-agent reinforcement learning","RL-coordinated AUVs cut plume map error, boost endurance"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The entire result rests on the Delft3D simulation being a faithful stand-in for the real Douro plume; every policy and kernel is trained and tested inside that model, with no at-sea validation.","fun_headline_variants_meta":{"raw":{"variants":["RL-driven AUVs map Douro plume for days with lower error","Multi-agent RL maps Douro plume longer and more accurately","AUV fleet learned to double endurance while mapping Douro plume","Douro plume mapped for days via multi-agent reinforcement learning","RL-coordinated AUVs cut plume map error, boost endurance"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.00062,"raw_usage":{"total_tokens":2694,"prompt_tokens":708,"completion_tokens":1986,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":452,"completion_tokens_details":{"reasoning_tokens":1899}},"tokens_in":452,"tokens_out":1986,"duration_ms":9912,"temperature":1.0,"reasoning_tokens":1899,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-04T11:37:45.293864+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Compare the learned policy's real-world salinity maps against independent CTD casts during a high-discharge Douro event. If the map MSE is no better than a simple lawnmower survey, or the fleet runs out of energy far sooner than the simulated 3-6 day missions, the central claim does not transfer to the ocean.","supporting_citations":[],"review_version":1}