{"id":"37491099-d8f6-42d7-ba1f-4baad9582890","arxiv_id":"2507.04640","paper_version":1,"verdict":"REJECT","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"high","formal_verification":"none","parameter_count":1,"one_line_summary":"A feedback-augmented, risk-aware trajectory optimizer (RA-SAA+FB) is shown in simulation to reduce collision rate and energy use for a tethered USV-UUV litter collection system.","lead":"This paper adds a feedback controller to a risk-aware planning method so an underwater robot tethered to a surface boat can reach targets and avoid obstacles even when the weight and drag of collected litter are uncertain. Simulations on a 2D model suggest the combined method has fewer collisions and uses less energy than three comparison controllers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Collision-rate metric in §4.2 is sign-reversed, so Figure 4b does not support the central 'reduces collision risk' claim.","rationale":"I agree with the overall rejection, but the most load-bearing defect is not the tether-reaction or drag-sign modeling assumption, as identified in the reader's weakest_assumption. The more direct problem is that the evaluation metric for the headline empirical comparison is sign-reversed: as written in §4.2, it flags safe states as collisions and cannot produce the collision rates shown in Figure 4b. This undermines the paper's central claim even if the model were physically exact. The reader's tether-reaction concern affects transfer to the real system, whereas the collision-metric error invalidates the reported simulation evidence itself. The drag-sign error in Eqs. (12)-(13) is also real and material. Additionally, the closing paragraph of §3.2 asserts that 'the same asymptotic optimality results can be derived as in (Lew et al., 2023)' for the feedback-augmented scheme without supplying a proof; this omitted proof is a secondary but nontrivial gap. The verdict remains REJECT, so no adjustment is needed.","tokens_in":11229,"tokens_out":12272,"duration_ms":139729,"concrete_test":"Recompute Figure 4b from the stored Monte Carlo trajectories using the corrected collision indicator ρcollision_ij = 1 if max_{t∈[0,tf]} (aO^2 − ((x(t)−xO)^2 + (d(t)−dO)^2)) > 0, using only UUV coordinates and consistent with Eq. (23), then compare method rankings and significance markers. If the corrected plot changes the reported ordering or significance, the central claim is not established; if the plot is unchanged, the §4.2 formula is a typo and the collision conclusion can stand.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The collision metric in §4.2 is internally inconsistent with the safety definition used in the optimization. It defines ρcollision_ij = 1 if min_{t∈[0,tf]} (−||y(t)−yO||^2 + aO^2) < 0, where yO is the obstacle center and aO its radius. A trajectory is safe when ||y−yO|| > aO, so the expression is negative for safe states. Because the initial points are explicitly resampled to be safe, the minimum over every trajectory is negative, and the literal formula assigns collision rate 1 to all methods. An actual collision makes the expression positive during penetration, but the minimum is still taken while the trajectory is outside, so it does not detect collisions either. The optimization's own constraint in Eq. (23) uses C = max_k(aO − ||(x,d)−(xO,dO)||) ≤ 0, which corresponds to max_t(aO^2 − ||(x,d)−(xO,dO)||^2) > 0 as the collision indicator, using only the UUV coordinates. Figure 4b shows collision rates below 1 and differences between methods, contradicting the literal §4.2 definition. Either the formula, the code, or the plot is wrong. Until this is resolved, the strongest empirical claim that RA-SAA+FB reduces collision risk is unsupported.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper models a taut-tethered USV-UUV system in the vertical plane, derives the dynamics via Euler–Lagrange equations, and formulates a risk-aware stochastic optimal control problem with CVaR constraints and sample average approximation. It proposes an extension, RA-SAA+FB, that adds a fixed-gain PD feedback term around the nominal trajectory, and evaluates the method in simulation against RA-SAA, A*+PID+CBF, and MPPI. The stated contributions are the dynamical model, the closed-loop extension of RA-SAA, and simulation evidence that the proposed method reduces collision risk and energy consumption under parameter uncertainty.","tokens_in":11494,"tokens_out":7311,"duration_ms":82006,"significance":"The system being modeled is relevant and has not been treated in this taut-tether configuration before, so a validated planar model and a risk-aware control extension would be of interest to the marine robotics and stochastic optimal control communities. The idea of centering sampled trajectories around a nominal one with a local feedback controller is a sensible and potentially practical enhancement of RA-SAA, and the Monte Carlo comparison against three external baselines is a reasonable, non-circular evaluation strategy. However, the manuscript as written contains a sign-reversed collision metric and a sign error in the drag model, both of which affect the main empirical claims. A quantitative treatment of the tether-reaction assumption is also missing. With corrections to those points, the paper could make a solid contribution, but the current numerical evidence is not trustworthy.","major_comments":[{"comment":"The collision flag is defined as 1 whenever min_t(−‖y(t)−yO‖² + aO²) < 0. Since safe trajectories satisfy ‖y(t)−yO‖ > aO, the quantity −‖y(t)−yO‖² + aO² is negative at the start and during all safe motion, so the minimum is always negative and the flag is 1 for every trajectory that begins outside the obstacle. A penetration makes the expression positive only while the trajectory is inside the obstacle, but the minimum over the entire interval is taken at the starting point, so the stated formula cannot detect collisions either. Figure 4b reports collision rates well below 1 and differences among methods, which is inconsistent with this definition. The optimization constraint in Eq. (23) uses max_k(aO − ‖(x,d)−(xO,dO)‖) ≤ 0 as the safe condition, so the metric should be based on max_t(aO − dist) > 0 or equivalently min_t(dist − aO) < 0. Until the formula is corrected and the simulations are rerun, the headline claim that RA-SAA+FB reduces collision risk is unsupported.","section":"Section 4.2, collision rate definition"},{"comment":"The drag model has a sign error in the relative-velocity components. From the coordinate transformation (4)–(5), the UUV velocity relative to stationary water along the radial and tangential directions is \\dot r − \\dot X sin θ and r\\dot θ − \\dot X cos θ, respectively. The expressions \\dot X sin θ + \\dot r and \\dot X cos θ + r\\dot θ in Eqs. (12)–(13) therefore have the wrong sign on the \\dot X terms. Because this drag model is the one simulated for all methods, the error changes every quantitative result in Section 4, including the comparisons in Figure 4. The equations must be corrected and the simulations repeated before the empirical claims can be assessed.","section":"Section 2.1, Eqs. (12)–(13)"},{"comment":"The assumption that the USV propeller is strong enough to ignore the reaction force from the tether decouples the USV dynamics as M \\ddot X = −η_X + f_X. This is a load-bearing modeling premise for the heavy-lifting scenario, where tether tension is expected to be significant, but no quantitative justification or sensitivity analysis is provided. A concrete test would be to compare the proposed model with one that includes the horizontal component of the tether tension in the X dynamics, or to report typical tension magnitudes obtained from the optimized trajectories. Without such evidence, it is unclear whether the simulation results transfer to the physical system.","section":"Section 2.1, Eq. (1) and Eq. (9)"},{"comment":"The text states that \"the same asymptotic optimality results can be derived as in (Lew et al., 2023)\" for the modified problem SOCP+FB, which includes a fixed feedback gain, a saturation function, and a time-discretized implementation. This is a load-bearing claim about the method's guarantees, but no proof or precise adaptation of the assumptions (A1)–(A4) to this closed-loop problem is given. The authors should either provide the derivation or explicitly mark this as a conjecture/limitation rather than asserting it as a proven property.","section":"End of Section 3.2, convergence claim"}],"minor_comments":[{"comment":"The symbol x is used both for the horizontal coordinate and for the full state vector (q, \\dot q, f), which makes equations such as (23) difficult to read; a distinct symbol for the full state would improve clarity.","section":"Sections 2.1–2.2"},{"comment":"The metric uses ‖y(t)−yO‖ with y=(x,d,X), but the obstacles are defined only in (x,d) coordinates; the paper should specify how yO is embedded in the three-dimensional output space or restrict the metric to the UUV coordinates.","section":"Section 4.2"},{"comment":"Since H(x) is a squared distance, it is nonnegative and the lower inequality −δ_M ≤ (1/N)Σ_i H(x_i(t_f)) is redundant; if the intent is to bound the absolute terminal error, the constraint should be written as an absolute value or as an upper bound only.","section":"Eq. (20c)"},{"comment":"The acronym \"RAA-SAA\" in the organization paragraph appears to be a typo for \"RA-SAA\".","section":"Section 1, Paper Organization"},{"comment":"The use of ellipses inside the vector defining the uncertain parameter ξ is notationally unclear; using explicit indices for the obstacle parameters would make the mapping to Eq. (23) more transparent.","section":"Eq. (21)"}],"recommendation":"major_revision","confidential_remarks":"The paper is within the journal's scope, and the identified errors are serious but localized and fixable. A rejection would be premature if the numerical results survive correction of the drag model, the collision metric, and the model-fidelity analysis. I recommend major revision and ask the authors to rerun all simulations with the corrected equations and to either substantiate or soften the convergence claim for the closed-loop SAA problem."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The core idea here is a sensible extension of Lew et al.'s sample-average approximation: add a linear feedback term around the nominal trajectory to keep sample paths centered, then optimize the feedforward plan. That is a useful trick for risk-aware trajectory optimization under parametric uncertainty, and the SeaClear USV-UUV application is genuinely new. The Euler-Lagrange derivation is internally consistent, the uncertainty setup is clearly stated, and the comparison against RA-SAA, A*+PID+CBF, and MPPI is a fair benchmark. So the paper has a real kernel.\n\nThe problems are in the validation. First, the drag model in Eqs. (12)-(13) has a sign error: the UUV's radial and tangential velocities relative to the water are rdot - Xdot sin(theta) and r thetadot - Xdot cos(theta), not the plus-sign expressions given. As a result, the simulated physics do not match the stated coordinate transformation, which casts doubt on all simulation results.\n\nMore serious is the collision metric in Section 4.2. It defines a collision as min over t of (-||y - yO||^2 + aO^2) < 0. But for any safe trajectory, that expression is negative because the UUV is outside the obstacle, so the minimum is always negative. The literal formula assigns collision rate 1 to every trajectory. Figure 4b shows rates below 1, so either the formula, the code, or the plot is wrong. As written, the central empirical claim that RA-SAA+FB reduces collision risk is unsupported.\n\nThe convergence guarantee for the feedback-augmented SAA is asserted without proof. This is likely fixable, since the saturated feedback is Lipschitz and bounded, but the paper should at least outline the argument. The assumption that the USV's propeller can ignore tether reaction forces is a modeling simplification that limits results to low-tension regimes; that is acknowledged but worth flagging.\n\nOverall, the method is plausible and the application is worthwhile, but the current validation does not hold up. A serious referee should require a corrected drag model, a fixed collision metric, and a proof sketch for convergence. This deserves peer review, with major revision expected.","headline":"A plausible feedback-augmented SAA method for a fresh USV-UUV model, but the simulation evidence is undercut by a sign-reversed collision metric and an incorrect drag model.","tokens_in":12057,"tokens_out":5804,"would_cite":false,"duration_ms":56026,"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 adding a local feedback loop to risk-aware trajectory optimization makes an underwater litter-collection system safer and more energy-efficient under parameter uncertainty.","keywords":["unmanned underwater vehicles","unmanned surface vehicles","trajectory optimization","uncertainty","collision avoidance","risk-aware control","stochastic optimal control"],"falsifier":"Measure the USV's actual acceleration while the UUV drags a heavy load through a winch and tether, and compare it with the decoupled prediction $M \\ddot{X} = -\\eta_X(\\dot{X}) + f_X$; if the deviation exceeds the process-noise level, the model's central assumption is false and the reported collision and energy improvements should be re-computed with coupled dynamics.","tokens_in":11031,"feed_emoji":"🤖","tokens_out":10545,"duration_ms":97008,"temperature":0.7,"pith_summary":"This paper studies an uncrewed surface vessel (USV) winching an uncrewed underwater vehicle (UUV) that picks up litter, where the litter's mass and drag make the system's parameters uncertain. The authors build a planar three-degree-of-freedom model of the coupled USV-tether-UUV system and formulate the navigation task as a stochastic optimal control problem with a conditional value-at-risk (CVaR) constraint on collision avoidance. Their contribution is a planner, RA-SAA+FB, that wraps a local feedback controller around a risk-aware sample average approximation (SAA) optimizer, keeping sampled uncertain trajectories close to a nominal trajectory. In Monte Carlo simulations with up to 50% parameter uncertainty, the method yields lower collision rates, lower input energy, and better target convergence than the RA-SAA, A*+PID+CBF, and MPPI baselines. The paper argues this makes the safety guarantees of risk-aware planning compatible with the inevitable parameter uncertainty of picking up unknown litter.","feed_headline":"Feedback loop cuts collisions and energy in tethered ocean robot","feed_subtitle":"CVaR-constrained planning plus a local PD law keeps trajectories centered when litter weight is uncertain.","key_machinery":"The carrying mechanism is the feedback-augmented closed-loop controller inside the SAA: a control law of the form $\\mu(x_i,t) = \\mathrm{sat}(u(t) + K(x_{\\mathrm{nom}}^u(t) - x_i(t)))$, where the nominal trajectory $x_{\\mathrm{nom}}^u$ is generated with the expected parameter values, $K$ is a fixed PD gain, and the saturation keeps inputs in the feasible set. Inside the optimization, sampled trajectories are simulated with this controller, so the CVaR collision-avoidance constraint is evaluated on trajectories that stay tightly centered around the nominal path instead of fanning out under large parameter uncertainty. This tightness keeps the terminal target constraint feasible and reduces the conservativeness of the risk constraint, while the outer loop still optimizes the feedforward plan $u$, preserving the convergence properties of the CVaR-SAA method the paper builds on.","core_discovery":"The central claim is that planning over closed-loop trajectories rather than open-loop inputs is what makes risk-aware trajectory optimization practical for a tethered UUV-USV litter collector. The paper derives a dynamical model from Euler-Lagrange equations in polar coordinates, treating the tether as taut and massless, the USV as decoupled from tether forces, and the UUV as a point mass with uncertain weight and drag. It then formulates the planning problem as a CVaR-constrained stochastic optimal control problem and solves it with a feedback-augmented sample average approximation, where each sampled trajectory is controlled by a fixed-gain PD feedback toward the nominal trajectory while the optimized feedforward plan is common to all samples. The paper claims this preserves the asymptotic optimality guarantees of the original SAA formulation, and in simulation it reports statistically significant gains in collision avoidance and energy efficiency over the three baselines under parameter uncertainty.","pith_inferences":["Editorial inference: treating the feedback gain $K$ as a decision variable of the optimization, rather than a hand-set PD gain, could produce a principled trade-off between trajectory tightness and control effort, and would test how sensitive the reported gains are to the gain value.","Editorial inference: the same feedback-augmented CVaR-SAA scheme should transfer to other tethered marine tasks, such as subsea inspection or cable-guided intervention, where the same drag and weight uncertainty appears.","Editorial inference: since the simulations randomize system parameters uniformly while real estimation errors are often Gaussian, re-running the study with a Gaussian disturbance model would show whether the collision and energy advantages persist when large deviations are rarer but bounded."],"forward_implications":["As parameter uncertainty grows, the collision rate of RA-SAA+FB stays near zero, while the open-loop RA-SAA and MPPI baselines degrade.","Input energy for RA-SAA+FB is the same or lower than the named baselines across all tested uncertainty levels.","The final position error remains small even at 50% parameter uncertainty, because the feedback keeps sampled trajectories near the nominal path and the terminal constraint feasible.","The same feedforward plan can be run at lower frequency online, since the local feedback controller handles deviations between optimization updates at a higher rate."],"supporting_citations":[{"why":"Defines the risk-averse SAA formulation with CVaR constraints over whole trajectories that the paper extends with feedback.","marker":"(Lew et al., 2023)"},{"why":"Provides the sample average approximation theory for stochastic programs with equality constraints, justifying the relaxed terminal constraint.","marker":"(Lew et al., 2024)"},{"why":"Supplies the MPPI baseline that the proposed method is compared against in simulation.","marker":"(Williams et al., 2018)"},{"why":"Supplies the control barrier function method used in the A*+PID+CBF baseline.","marker":"(Ames et al., 2017)"},{"why":"Supplies the quadratic drag model used for the UUV's hydrodynamic forces in the derived dynamics.","marker":"(Mathai et al., 2019)"},{"why":"Motivates the use of CVaR as a tail-risk measure in stochastic optimal control.","marker":"(Chow et al., 2015)"},{"why":"Represents the prior tether-dynamics modeling approach that the paper contrasts with its taut-tether assumption.","marker":"(Hong et al., 2020)"}],"fun_headline_variants":["Closed-loop risk planning cuts collisions, energy for tether robot","CVaR feedback planning reduces tethered underwater robot risk","Underwater litter collector uses CVaR to cut collisions and power","Feedback-augmented SAA planning trims risk for tethered ocean robot","Risk-aware control keeps tethered litter robot centered under uncertainty"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the USV's propeller force is strong enough that the tether's pull on the USV can be ignored; if tether tension is significant, as it is when lifting heavy litter, the USV's motion is coupled to the UUV's, and the simulated benefits may not carry over to the real system.","fun_headline_variants_meta":{"raw":{"variants":["Closed-loop risk planning cuts collisions, energy for tether robot","CVaR feedback planning reduces tethered underwater robot risk","Underwater litter collector uses CVaR to cut collisions and power","Feedback-augmented SAA planning trims risk for tethered ocean robot","Risk-aware control keeps tethered litter robot centered under uncertainty"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000571,"raw_usage":{"total_tokens":2644,"prompt_tokens":830,"completion_tokens":1814,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":446,"completion_tokens_details":{"reasoning_tokens":1726}},"tokens_in":446,"tokens_out":1814,"duration_ms":13939,"temperature":1.0,"reasoning_tokens":1726,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T19:43:50.023928+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Measure the USV's actual acceleration while the UUV drags a heavy load through a winch and tether, and compare it with the decoupled prediction $M \\ddot{X} = -\\eta_X(\\dot{X}) + f_X$; if the deviation exceeds the process-noise level, the model's central assumption is false and the reported collision and energy improvements should be re-computed with coupled dynamics.","supporting_citations":[{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Defines the risk-averse SAA formulation with CVaR constraints over whole trajectories that the paper extends with feedback."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Provides the sample average approximation theory for stochastic programs with equality constraints, justifying the relaxed terminal constraint."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the MPPI baseline that the proposed method is compared against in simulation."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Supplies the quadratic drag model used for the UUV's hydrodynamic forces in the derived dynamics."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Motivates the use of CVaR as a tail-risk measure in stochastic optimal control."},{"cited_title":null,"cited_arxiv_id":null,"evidence_quote":"Represents the prior tether-dynamics modeling approach that the paper contrasts with its taut-tether assumption."}],"review_version":1}