{"id":"d2a2942e-f77f-4068-a0b1-d23ea97aa28b","arxiv_id":"2607.09557","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.5,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"Surrogate CFD drag models enable zero-shot RL control of a 6-DOF AUV that is more efficient, accurate, and robust to payload change than simplified-physics baselines.","lead":"Researchers trained a neural-network approximation of CFD drag forces and used it inside a reinforcement-learning loop to produce a 6-DOF AUV controller that transfers zero-shot to a real vehicle. The CFD-based policy used less energy, finished waypoint tasks faster, and was the only one that still worked after 2 lb of extra weight was added.","discovery_kind":"new_method","skeptic_critique":{"model":"grok-4.5","headline":"The 31%/11%/19% gains rest on a single reward-shaping recipe and modest trial counts; the CFD fidelity claim is secondary to that confound.","rationale":"The reader correctly flags the steady-state/split-velocity CFD assumption as a soft spot, yet that assumption is not the most load-bearing one for the quantitative claim that is actually advertised. The paper’s own tank results already demonstrate large sensitivity of the lower-fidelity models to reward coefficients; therefore the cleanest threat to the 31/11/19 numbers is that they were measured under a single, CFD-friendly reward rather than under a controlled multi-reward comparison. The exclusive 2 lb transfer result is more persuasive and survives this critique, which is why the overall verdict stays CONDITIONAL rather than moving to REJECT. A multi-reward field re-run would cleanly separate “CFD is better hydrodynamics” from “CFD is more robust to reward shaping.” No stronger internal inconsistency was found; the experimental design remains stronger than most sim-to-real AUV papers.","tokens_in":11884,"tokens_out":646,"duration_ms":6364,"concrete_test":"Re-train all three drag models under the identical three reward coefficient sets used in the tank (Cconservative, Cbalanced, Caggressive), deploy each of the nine policies zero-shot on the same field U-pattern (n≥5), and recompute the energy/time/error tables. If the CFD advantage disappears or reverses under any of the other two reward sets, the headline percentages cannot be attributed primarily to hydrodynamic fidelity.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The strongest claim packages three quantitative gains (31% energy, 11% faster, 19% less error) plus exclusive 2 lb transfer success as evidence that SDM/CFD training is what closes the gap. Those percentages come exclusively from the field U-pattern under one fixed reward set (§3.3.1: λ p=5.0, λ q=2.0, λ a=−0.08, etc.) with n=5 trials per policy (Figs. 3–4). The tank reward-shaping sweep (§3.3.2, Fig. 5, A.1) already shows that the same inertia-box model can be made competitive by switching from “Aggressive” to “Balanced” coefficients, while System-ID remains mediocre across all three. Consequently the field deltas may largely reflect that the CFD policy happened to be less brittle to that particular hand-chosen reward rather than that the steady-state split-velocity CFD itself is a more faithful hydrodynamic model. The exclusive 2 lb success is stronger evidence, yet it still uses the single best reward found for each model and only five trials; without a matched multi-reward field ablation the causal attribution to CFD fidelity remains under-supported.","agreement_with_reader":"partial"},"referee_report":{"model":"grok-4.5","summary":"The paper proposes training lightweight MLP surrogate drag models (SDMs) on steady-state OpenFOAM RANS data of a 6-DOF AUV, then using those SDMs inside a highly parallel IsaacSim PPO loop so that zero-shot policies can be obtained in ~15 minutes. Three drag models are compared (MuJoCo-style inertia box, System-ID diagonal coefficients, and the CFD SDM). On a field U-pattern the SDM policy reports 31 % lower thruster effort, 11 % faster waypoint times and 19 % lower cross-track error than the simplified-physics baseline; in a tank payload experiment only the SDM policy transfers under a 2 lb stern weight when domain randomization is applied. The authors further claim that the CFD-based controller is less sensitive to reward-coefficient choices and that its simulation reward curves better predict real-world transfer.","tokens_in":12238,"tokens_out":1001,"duration_ms":12364,"significance":"If the causal attribution to hydrodynamic fidelity holds, the work supplies a practical route for incorporating high-fidelity CFD into RL pipelines without placing a full solver inside the training loop—an important engineering contribution for 6-DOF AUV control. The combination of tank, field (including a scientifically relevant reef site) and payload-transfer experiments is stronger empirical grounding than most contemporaneous AUV-RL papers. Explicit credit is due for the open comparison across three drag models, the quantitative energy/time/error metrics, and the demonstration that only the CFD policy survives a realistic payload change under domain randomization.","major_comments":[{"comment":"§4.1 / Figs. 3–4: the headline 31 % / 11 % / 19 % gains are obtained under a single hand-chosen reward set (§3.3.1) with n=5 trials per policy. The tank reward-shaping sweep (§3.3.2, Fig. 5, A.1) already shows that the inertia-box model can be made competitive simply by switching from “Aggressive” to “Balanced” coefficients. Without a matched multi-reward field ablation (or at least confidence intervals / statistical tests on the n=5 runs), the causal claim that the gains arise from CFD fidelity rather than differential brittleness to that particular reward remains under-supported.","section":null},{"comment":"§3.1.3: the SDM is constructed by sampling linear and angular velocities independently, fitting two separate MLPs, and summing their wrench outputs. The manuscript never quantifies the approximation error of this split relative to a joint (v,ω) CFD evaluation or to measured vehicle wrenches. Because the central scientific claim is that higher-fidelity hydrodynamics close the sim-to-real gap, an error analysis (or at least a residual plot against held-out joint-velocity CFD cases) is load-bearing and currently missing.","section":null},{"comment":"§4.2 / Fig. 6: the exclusive 2 lb transfer success is the strongest evidence for the method, yet it still uses only the single best reward found for each model and five trials. Reporting success rates with binomial confidence intervals (or repeating the payload test under the other two reward sets) would make the robustness claim far more convincing.","section":null}],"minor_comments":[{"comment":"Abstract and §1 claim “first” zero-shot 6-DOF AUV policy trained on CFD surrogates; a short related-work paragraph clarifying the distinction from the 2-D/3-DOF CFD-RL fish papers already cited would strengthen the novelty statement.","section":null},{"comment":"Eq. (1) and the System-ID paragraph: the effective-mass / inertia symbols are never numerically reported; listing the fitted d_i, k_j values (or placing them in the appendix) would aid reproducibility.","section":null},{"comment":"Figure A.2 caption: the red inertia-box curves appear only for a subset of axes; clarifying which components are shown would help the reader assess the claimed cross-coupling.","section":null},{"comment":"Typographical: “PIMPLE Solver” appears twice in Fig. 1 with inconsistent capitalization; “meff,i” in Eq. (1) should be m_eff,i for consistency with later notation.","section":null}],"recommendation":"major_revision","confidential_remarks":"The empirical package is already stronger than many AUV-RL submissions; if the authors can supply even a modest multi-reward field check or a quantitative residual analysis of the split-velocity SDM, the paper would clear the bar for a solid contribution. The current major issues are fixable within a revision cycle."},"author_rebuttal":null,"desk_editor":{"model":"grok-4.5","letter":"The core result is real: they distilled steady-state OpenFOAM RANS data into two light MLPs (linear and angular, then summed), dropped them into IsaacSim + PPO, and got a zero-shot 6-DOF policy that swam a U-pattern on a reef and handled an extra 2 lb stern weight when the box and System-ID policies failed. Training finishes in ~15 min. That combination—CFD fidelity cheap enough for parallel RL plus actual free-swimming hardware—has not been shown before for full 6-DOF AUVs.\n\nWhat they do well is the experimental ladder. Tank reward-shaping sweep, then field n=5 trials with energy/time/cross-track metrics, then the 2 lb DR transfer. The CFD policy is visibly less brittle across the three reward recipes and is the only one that still completes waypoints once the weights go on. Related-work coverage is honest; they correctly note that prior CFD-RL stayed at 2–4 DOF or kept the solver in the loop.\n\nSoft spots are real but proportionate. The headline 31 % / 11 % / 19 % numbers come from one fixed reward set and five trials; the tank already shows the inertia-box model can look competitive under a different coefficient choice. So the field deltas partly reflect reward luck rather than pure hydrodynamic fidelity. The steady-state + independent lin/ang sampling is an acknowledged simplification; they never validate how much unsteady coupling is missing. Reward coefficients were iterated after tank looks, mild post-selection. No code or CFD data released. None of this overturns the central claim that the SDM route improves transfer robustness.\n\nMath is standard PPO + simple MLPs; data collection (coast-downs, CFD sampling) is transparent. Citations are appropriate, no padding.\n\nThis is for anyone doing sim-to-real marine control or high-DOF underwater RL. Worth a serious referee; the hardware results and the clean payload ablation justify the time. I would engage and cite the transfer experiment.","headline":"First real zero-shot 6-DOF AUV RL from CFD surrogates; payload transfer is clean evidence, field percentages partly confounded by reward choice.","tokens_in":12859,"tokens_out":512,"would_cite":true,"duration_ms":16697,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"grok-4.5","headline":"Surrogate CFD drag models let RL controllers for 6-DOF AUVs transfer zero-shot with lower energy, faster waypoints, and payload robustness that simpler physics models lack.","keywords":["sim-to-real transfer","reinforcement learning","autonomous underwater vehicles","computational fluid dynamics","surrogate drag models","domain randomization","6-DOF control"],"falsifier":"Repeat the identical tank and field waypoint trials with a new payload or vehicle geometry whose true hydrodynamics deviate strongly from the steady-state RANS assumption; if the SDM policy then fails while a full transient CFD or real-data model succeeds, the central claim is falsified.","tokens_in":12771,"feed_emoji":"🌊","tokens_out":626,"duration_ms":7881,"temperature":0.7,"pith_summary":"Fine control of autonomous underwater vehicles is hard because traditional controllers need retuning whenever payload or conditions change, and simplified drag models leave a large sim-to-real gap. This paper shows that lightweight neural networks trained on steady-state CFD data can stand in for full fluid-dynamics solvers inside a reinforcement-learning loop, producing a controller in minutes. The resulting policy is the first zero-shot RL controller for a full 6-DOF AUV trained this way. In field and tank tests it uses 31 percent less thruster energy, reaches waypoints 11 percent faster with 19 percent less path error, and is the only policy that still works after two pounds are added to the stern. The higher-fidelity drag model also makes training less sensitive to reward-shaping choices and more predictive of real-world transfer.","feed_headline":"CFD surrogates cut AUV energy 31% in zero-shot RL","feed_subtitle":"Only the high-fidelity drag model still works after a 2 lb stern payload change","key_machinery":"Surrogate drag models (SDMs): two small MLPs, one for linear velocity and one for angular velocity, each fitted to large RANS CFD wrench datasets and summed at every simulation step so that high-fidelity coupled drag can be queried in real time inside a massively parallel RL environment.","core_discovery":"Training an RL policy on surrogate drag models that approximate steady-state CFD data yields a zero-shot 6-DOF AUV controller that outperforms both an equivalent-inertia-box model and a system-identified diagonal model: lower energy, faster transit, smaller tracking error, greater robustness to reward design, and successful transfer under a 2 lb stern payload that defeats the other two policies.","pith_inferences":[],"forward_implications":[],"fun_headline_variants":["CFD surrogates cut AUV energy 31% in zero-shot RL","Zero-shot AUV RL via CFD drag surrogates uses 31% less energy","Surrogate CFD drag trains zero-shot 6-DOF AUV RL that beats box models","CFD-based RL AUV controller alone transfers after 2 lb payload shift","SDM-trained RL AUV policy: 11% faster, 19% less error, more robust"],"cache_read_input_tokens":128,"weakest_assumption_plain":"That sampling linear and angular velocities independently in steady-state CFD and then simply adding the two MLP outputs is faithful enough to the vehicle’s real coupled, possibly unsteady hydrodynamics for zero-shot transfer to succeed.","fun_headline_variants_meta":{"raw":{"variants":["CFD surrogates cut AUV energy 31% in zero-shot RL","Zero-shot AUV RL via CFD drag surrogates uses 31% less energy","Surrogate CFD drag trains zero-shot 6-DOF AUV RL that beats box models","CFD-based RL AUV controller alone transfers after 2 lb payload shift","SDM-trained RL AUV policy: 11% faster, 19% less error, more robust"]},"model":"grok-4.5","effort":"low","cost_usd":0.0052,"raw_usage":{"total_tokens":1488,"prompt_tokens":836,"num_sources_used":0,"completion_tokens":101,"cost_in_usd_ticks":52000000,"prompt_tokens_details":{"text_tokens":836,"audio_tokens":0,"image_tokens":0,"cached_tokens":256},"completion_tokens_details":{"audio_tokens":0,"reasoning_tokens":551,"accepted_prediction_tokens":0,"rejected_prediction_tokens":0}},"tokens_in":836,"tokens_out":101,"duration_ms":5548,"temperature":1.0,"reasoning_tokens":551,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-13T02:09:07.254099+00:00","model_set":{"reader":"grok-4.5"},"falsifier":"Repeat the identical tank and field waypoint trials with a new payload or vehicle geometry whose true hydrodynamics deviate strongly from the steady-state RANS assumption; if the SDM policy then fails while a full transient CFD or real-data model succeeds, the central claim is falsified.","supporting_citations":[],"review_version":1}