{"id":"6f0a7843-9cbe-42d2-9d70-a6ea0bc382ba","arxiv_id":"2608.08025","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":8,"one_line_summary":"By tracking viewing directions per voxel and using a learned position-advantage field, DA-NBV lifts reconstruction coverage by about 3 percentage points and cuts Chamfer distance by 43% over strong baselines in simulated sea conditions.","lead":"DA-NBV is a drone-view planning method that remembers which directions each part of a ship has been seen from, not just whether it was seen. It produced more complete 3D ship reconstructions with shorter flight paths than prior planners in a maritime wave simulator.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The headline gain rests on an unvalidated simulator: if the equivalent-gravity/velocity scaling and four-point ship-response model are not faithful to real at-sea conditions, the 3 pp CR and 43% CD gains may not transfer to deployment.","rationale":"The reader's weakest assumption identifies the same load-bearing concern: the simplified maritime simulator may not be a faithful proxy for real at-sea scanning. My stress-test concurs, and I add a sharper mechanism: the equivalent-gravity and velocity scaling treat the UAV as if its dynamics scale with ship length, while real quadrotor mass, thrust, and drag do not. This could materially change path efficiency, wind sensitivity, and the amount of ship motion experienced between frames, all of which drive the reported As, Ap, and CD gaps. The paper's static cross-dataset results provide some support for the general direction-aware mechanism, but they do not test the sea-specific external validity. The absence of error bars and the lack of code/data release are additional reproducibility concerns, but the simulator fidelity issue is the most load-bearing because it threatens the central deployment claim. The reader's CONDITIONAL verdict remains appropriate: the internal design and ablations are solid, but the empirical support for at-sea transfer is conditional on validating the simulation model or providing real-world demonstrations.","tokens_in":18516,"tokens_out":8554,"duration_ms":100878,"concrete_test":"Run a computational sweep in the existing Isaac Gym environment: vary eta_L over {0.05, 0.15, 0.5} and fix the Crazyflie airspeed at its physical envelope (e.g., 8 m/s) instead of scaling it as 25 * eta_L, then compare DA-NBV versus Hestia on the same SeaShip-3D splits and metrics. If the CR/CD advantage shifts by more than the claimed 3 pp / 43%, the scale-normalization assumption is load-bearing; if the advantage persists across the sweep, the primary remaining risk is unmeasured sim-to-real transfer and a controlled field deployment would be needed.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central empirical claim rests entirely on a custom Isaac Gym simulator, and the simulator's scale-normalization assumptions are not validated against real at-sea scanning. Eqs. (28)-(31) scale all spatial quantities and velocities by eta_L while preserving time through g_eq = eta_L * g; this is internally consistent for the wave field, but the UAV is modeled after Crazyflie without scaling its mass, thrust, or drag. At eta_L around 0.15 (a 100 m vessel normalized to 15 m), the simulated airspeed V_a = 25 * eta_L and the wind speeds become roughly 3-4 m/s, so the wind-to-airspeed ratio and controller authority are not representative of a real Crazyflie scanning that vessel. The simplified four-point waterline ship response (Eqs. 32-34), height-comparison sea occlusion, and absence of GPS/IMU drift, gust transients, and ICP failure modes further mean that the measured 3 pp CR and 43% CD gains cannot be assumed to transfer to deployment. This is not an internal inconsistency: the planner's components are coherent and the ablations support their individual contributions. It is an external-validity gap that is load-bearing because the headline claim concerns reconstructing ships at sea, not just in the simulator.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"DA-NBV proposes a direction-aware next-best-view planner for UAV-based 3D reconstruction of ships at sea. It extends occupancy-grid states with voxel-level directional observation bins, learns a Position Advantage Field (PAF) that scores candidate viewpoints from missing-direction alignment, visibility, distance, and geometric complexity, and couples this with a locally constrained autoregressive action policy and a convex coverage-shaping reward. Training and evaluation are conducted in an Isaac Gym simulator built on the new SeaShip-3D dataset, with wave-induced heave, roll, pitch, simplified sea-surface occlusion, and a wind field. The paper reports that DA-NBV improves coverage rate by about 3 percentage points, reduces Chamfer distance by 43%, and improves path efficiency relative to GenNBV and Hestia, with ablations attributing the gains to the PAF, the local action space, and the reward shaping.","tokens_in":18833,"tokens_out":6963,"duration_ms":79149,"significance":"The core idea is well motivated: occupancy-based NBV policies discard information about the diversity of viewing directions, which is clearly relevant for self-occluded ship superstructures. The PAF formulation, the local autoregressive action decomposition, and the coverage-shaping reward are coherent, and the ablations in Table 2 are directionally consistent with the stated roles of each component. The problem selection and the construction of a ship-oriented dataset with an accompanying maritime simulator are useful contributions if validated. However, the headline quantitative claims currently rest on a single simplified simulator whose fidelity is not established, and the headline numbers are reported as point estimates without uncertainty quantification. The paper also provides no code or simulator release, which hinders independent verification. With simulation-only claims and proper statistical reporting, the proposed components would be a clear incremental contribution; as written, the 'ships at sea' claim is stronger than the evidence supports.","major_comments":[{"comment":"The evaluation environment is the sole evidence for the headline at-sea claim, and its scale-normalization assumptions are not validated. The paper scales all spatial quantities and airspeeds by eta_L while retaining the reference time scale through g_eq = eta_L g, but the UAV platform is modeled after the Crazyflie without scaling its mass, thrust, or drag. For a 100 m vessel normalized to 15 m, eta_L is about 0.15, so V_a becomes 3.75 m/s and the Sea State 9 reference winds of 20.8-24.4 m/s become about 3.1-3.7 m/s; this is a substantially different wind-to-airspeed and control-authority regime from a Crazyflie scanning a real 100 m ship. The four-point waterline response in Eqs. (32)-(34), the height-comparison sea-occlusion model, and the absence of GPS/IMU drift, gust transients, and registration failure modes mean that the measured 3 pp CR and 43% CD gains cannot be assumed to transfer to deployment. Please either validate the scaling and ship-response model against real ship/UAV data or explicitly restrict the central claim to simulation.","section":"Maritime Simulation Environment and Dataset (Eqs. 28-35)"},{"comment":"The headline results in Table 1 and Table 3 are point estimates. The supplementary protocol evaluates 10 randomized episodes per test object, but no standard deviations, confidence intervals, or significance tests are reported, so it is unclear whether the 2.96 pp CR gap and the 43% CD reduction are stable across episodes or dominated by a few favorable runs. Please report per-metric distributions across the SeaShip-3D episodes, include pairwise effect sizes or significance tests, and adjust the abstract and conclusion claims accordingly.","section":"Performance Comparison and Experimental Protocol"},{"comment":"The training reward uses ground-truth observability masks M(v,j) and ground-truth complexity weights c_v, as stated in the main text and detailed in the supplement, while the paper does not report whether the comparison baselines receive an equivalent privileged reward. Because the paper also states that all methods are trained under identical conditions, the reported gains may partly reflect an asymmetry in training signal rather than the proposed directional state and action space. Please either train the baselines with analogous ground-truth-derived rewards or report a DA-NBV ablation trained without the ground-truth reward quantities, so that the contribution of the proposed state/action/reward design can be isolated.","section":"Reward Design and Policy Optimization"}],"minor_comments":[{"comment":"The text says all methods are evaluated with the same fixed number of views, but DA-NBV includes a stop action and terminal reward; please clarify whether the stop action is exercised during evaluation and how the fixed view budget is reconciled with a learned termination decision.","section":"Experimental Setup"},{"comment":"The supplementary material states that 50 SeaShip-3D test ships and 10 episodes per ship are used for each sea state; please state the total number of episodes per metric to avoid ambiguity in the reader's interpretation of the reported values.","section":"Experimental Protocol"},{"comment":"The visibility approximation validation reports masking accuracy at a 0.10 threshold; it would be clearer to also report the effect of this approximation on final planning metrics, since the reported 13.4% false-mask rate is not directly connected to reconstruction quality.","section":"Visibility Approximation"},{"comment":"No code, trained models, or simulator source are indicated as available; providing them would substantially improve reproducibility and allow independent verification of the baseline implementations.","section":"General"},{"comment":"The 'All' columns average across dynamic maritime and static object datasets, which have different protocols; please justify this aggregate or present the datasets separately as the primary comparison.","section":"Table 1 and Table 3"}],"recommendation":"major_revision","confidential_remarks":"The central algorithm is coherent and the ablations support the component-level claims, but the empirical evidence is not yet sufficient for the abstract's at-sea claim. The main risk is external validity of the simulator, and the main missing technical element is uncertainty quantification for the headline numbers. I would not reject on the basis of simulator simplification alone, because simulation studies are common for NBV, but the paper should either strengthen the validation or explicitly narrow the claims."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: the core idea here is good and the paper is honestly written. Adding a 12-bin directional observation state and a learned Position Advantage Field on top of occupancy is a natural, useful extension of Hestia's face-level representation, and the ablations support each of the three components. The local action space and nonlinear coverage-shaping reward are sensible engineering choices, and the cross-dataset results on Houses3K and OmniObject3D give me some confidence the method isn't just memorizing ship geometry. I believe the internal logic holds up.\n\nThe soft spots are mostly about evidence, not method. The whole evaluation lives in a custom Isaac Gym simulator, and the simulator's scaling choices are questionable. They normalize every ship to 15 m and scale all spatial quantities and velocities by eta_L, then use an equivalent gravity g_eq = eta_L * g to keep the time scale. That is internally consistent for the wave field, but it also scales wind speed down to roughly 3-4 m/s for a 100 m vessel. So sea state 9 in the simulation corresponds to a mild breeze for a Crazyflie, not the wind conditions the reference sea-state table implies. That doesn't break the planner's internal validity, but it does mean the 'robust to severe maritime conditions' claim is only about this scaled-down proxy.\n\nTwo more concrete issues. First, there are no error bars or significance tests anywhere. Ten randomized episodes per ship is a reasonable amount of data, but Table 1 reports single point estimates, and the 43% Chamfer reduction could easily have overlapping confidence intervals with Hestia. Second, there is no code or dataset release. For an empirical systems paper, that is a real reproducibility gap—especially because they introduce SeaShip-3D as a contribution, but no link is provided.\n\nI would not call any of this fatal. The method is coherent, the ablation is clean, and the authors explicitly say real-world deployment is future work. The direction-aware state is worth taking seriously. But as it stands, the paper's central claim is 'we built a good simulator and our policy wins in it.' That is publishable with revisions, but not as-is. If I were the editor, I would ask for error bars, an availability statement, and either a higher-fidelity validation or a softer claim about real-at-sea performance.\n\nThis is a paper I'd send to peer review. The idea is novel enough and the design is careful enough to deserve referee time. I'd bring it to reading group if someone is interested in active perception; I wouldn't put it on my own citation list until the code/data situation is resolved.","headline":"Solid direction-aware NBV idea with clean ablations, but the headline gains are only demonstrated in an unvalidated simulator with no error bars or released artifacts.","tokens_in":19336,"tokens_out":3476,"would_cite":false,"duration_ms":37306,"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":"DA-NBV records each voxel's viewing-direction history and uses it to pick viewpoints, achieving 98.49% coverage and 43% lower Chamfer distance than the best baseline.","keywords":["next-best-view planning","3D reconstruction","directional coverage","maritime UAV scanning","reinforcement learning","Position Advantage Field","sea-state simulation","ship reconstruction"],"falsifier":"A field test on a real quadrotor over a known vessel under measured sea states, running DA-NBV with the paper's motion-compensation pipeline, would falsify the central claim if the resulting coverage rate and Chamfer distance do not reproduce the reported margin over Hestia.","tokens_in":18329,"feed_emoji":"🚢","tokens_out":9406,"duration_ms":93775,"temperature":0.7,"pith_summary":"The paper tries to establish that replacing an occupancy-only reconstruction state with one that also records, for each voxel, which directions it has been seen from makes an automated drone scan of a ship at sea both more complete and cheaper to execute—a concern for maritime supervision and damage assessment, where manual flight is costly. DA-NBV accumulates per-voxel viewing-direction vectors, encodes the missing directions into a learnable Position Advantage Field, and pairs this with a locally constrained autoregressive action policy and a convex coverage-shaping reward. On the paper's maritime simulator with 300 ship models, the policy reaches a 98.49% reconstruction coverage rate versus 95.53% for the strongest baseline, cuts Chamfer distance from 6.49 cm to 3.68 cm, and retains most of its margin when the sea state is raised from 0 to 9. If the simulator is a faithful proxy, the result is evidence that directional coverage is a missing state variable for efficient, high-quality scanning of self-occluding objects.","feed_headline":"Direction-aware drone scans lift ship reconstruction to 98.5%","feed_subtitle":"DA-NBV tracks which viewing angles each surface has been seen from, cutting reconstruction error by 43%.","key_machinery":"The mechanism that carries the argument is the Position Advantage Field (PAF), a learnable scalar field over a local lattice of candidate UAV positions. A Learnable Position Advantage Scorer evaluates each voxel–candidate pair from distance, approximate visibility transmittance, alignment with yet-unobserved direction bins, and PCA-based geometric complexity (linearity, scattering, and curvature), and sums the learned utilities over active voxels. The PAF tells the policy where missing directional coverage can be gained, while a two-stage action head first picks a local relative translation and then predicts yaw and pitch, and the reward uses a convex shaping $f(x)=x^3$ on directional-coverage increments so that late-game gains are not undervalued.","core_discovery":"DA-NBV's central claim is that a voxel marked as observed is an insufficient planning signal for ships: the planner must also know whether its surface has been seen from enough complementary directions. The method therefore discretizes the viewing sphere into twelve directional bins per voxel, records which bins have been activated by past observations, and derives a Position Advantage Field that scores candidate UAV positions by how well they align with the directions still missing, weighted by local geometric complexity and an approximate visibility transmittance. Trained with proximal policy optimization in the wave-driven simulator, this state representation yields reconstruction completeness of 98.49% and Chamfer distance of 3.68 cm on SeaShip-3D, together with higher coverage per step and per path length than occupancy-based and voxel-face-based baselines.","pith_inferences":["An untested step the paper leaves open is measuring how much of the gain depends on the approximate visibility model (axis-aligned bounding-box transmittance and height-comparison sea occlusion) by comparing with exact ray tracing in a subset of scenes.","The directional state could be combined with appearance-based quality signals, so a natural next reward would penalize reprojection error or texture fidelity, since geometric coverage alone does not guarantee surfaces usable for photogrammetric measurement.","The readiest deployment test is to treat the learned termination action as the stopping rule under a fixed battery budget and measure reconstruction quality per unit flight time, which is closer to the real operating constraint than a fixed view count."],"forward_implications":["An occupancy-only or voxel-face-only state systematically underestimates the reconstruction needs of self-occluding structures, so recording per-voxel direction history should improve completeness for any such object.","With ICP motion compensation, wave-induced heave, roll, and pitch are manageable: DA-NBV's coverage drops only 0.89 percentage points and Chamfer distance rises only 0.88 cm from Sea State 0 to Sea State 9.","The locally constrained action space and the nonlinear reward act on different components—path efficiency versus directional completeness—and the complete system obtains both, implying the design choices are complementary rather than redundant.","Reported cross-dataset results on Houses3K and OmniObject3D show the directional state transfers to static objects, so the method is not overfit to ship geometry."],"supporting_citations":[{"why":"Scan-RL, the learning-based NBV baseline that first derives next viewpoints from image history; DA-NBV is compared against it.","marker":"Peralta et al. 2020"},{"why":"GenNBV, the occupancy-grid generalizable NBV policy whose probabilistic 3D geometry and action history state DA-NBV extends with directional statistics.","marker":"Chen et al. 2024"},{"why":"Hestia, the strongest baseline; it models six voxel-face observation states and is the method DA-NBV must beat on coverage and Chamfer distance.","marker":"Lu et al. 2026"},{"why":"ActiveRMAP, the radiance-field active-mapping baseline compared in the main experiments.","marker":"Zhan et al. 2022"},{"why":"ICP, used to register each new depth observation to the ship-centered reconstruction map and compensate for wave-induced motion.","marker":"Besl and McKay 1992"},{"why":"Wind-sea spectrum underlying the simulated irregular wave field and sea states.","marker":"Pierson and Moskowitz 1964"},{"why":"PPO, the reinforcement-learning algorithm used to train the DA-NBV policy and value network.","marker":"Schulman et al. 2017"},{"why":"PCA-based linearity, scattering, and curvature descriptors used as online geometric complexity for the Position Advantage Field.","marker":"Weinmann et al. 2015"}],"fun_headline_variants":["Direction-aware planner cuts ship reconstruction error by 43%","View-history tracking yields 98.5% complete ship scans","Sea-state aware NBV boosts 3D ship scanning efficiency","Tracking viewing angles makes ship scans 43% more accurate"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the simplified maritime simulator—an eight-wave ocean with waterline-derived heave, roll, and pitch, height-comparison sea occlusion, and scaled ship and wind parameters—faithfully predicts how a real UAV scan would perform at sea.","fun_headline_variants_meta":{"raw":{"variants":["Direction-aware planner cuts ship reconstruction error by 43%","View-history tracking yields 98.5% complete ship scans","Sea-state aware NBV boosts 3D ship scanning efficiency","Tracking viewing angles makes ship scans 43% more accurate"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.0002,"raw_usage":{"total_tokens":1395,"prompt_tokens":985,"completion_tokens":410,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":601,"completion_tokens_details":{"reasoning_tokens":341}},"tokens_in":601,"tokens_out":410,"duration_ms":5278,"temperature":1.0,"reasoning_tokens":341,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T00:32:13.907345+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"A field test on a real quadrotor over a known vessel under measured sea states, running DA-NBV with the paper's motion-compensation pipeline, would falsify the central claim if the resulting coverage rate and Chamfer distance do not reproduce the reported margin over Hestia.","supporting_citations":[],"review_version":1}