{"id":"0db82e82-d4ad-416a-88b4-eedcdd985c6e","arxiv_id":"2505.08382","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A deep RL agent with a self-adaptive curriculum learns continuous, curvature-constrained coverage paths for fixed-wing UAVs, solving procedurally generated and hand-crafted maps.","lead":"This paper trains a drone to plan smooth curved flight paths that fully photograph a target area while avoiding no-fly zones, using deep reinforcement learning on continuous maps instead of grids. The agent learns energy-saving routes in simulation and can handle new maps after training.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"No measured energy or baseline comparison appears in the results; the paper's central claim to 'energy-efficient' coverage is supported only by success rates and trajectory figures, so the main quantitative claim is unsubstantiated as reported.","rationale":"The reader's weakest assumption is the unvalidated power model. That is a real concern, but it is downstream: even accepting Eq. 17, the paper never reports the resulting power integral or any baseline that would show the learned policy is energy-efficient. I therefore stress the missing quantitative energy evidence and the absence of any comparison, because these are directly testable and are needed to support the central claim as worded. If the proposed energy-and-baseline evaluation shows the policy beats or matches a boustrophedon heuristic on the modeled power integral, the central claim would be substantially supported; if not, the claim should be weakened. This does not change the reader's CONDITIONAL verdict, but it sharpens the condition: the revision should add energy metrics and a baseline, not merely validate the power model.","tokens_in":10653,"tokens_out":6444,"duration_ms":71268,"concrete_test":"Run all three trained checkpoints on a fixed set of, say, 100 procedurally generated difficulty-1 maps from the same generator, and for each rollout compute: complete-coverage success, total ∫_0^T P(t)dt using Eq. 17, total path length, mean absolute curvature, and number of constraint resets. Compare these against a deterministic boustrophedon/lawnmower baseline with the same FoV, obstacle avoidance, and turn-radius constraints, and also against a random feasible-action policy, using paired mean±std. If AM-SAC is not statistically better than the boustrophedon baseline on total power integral, or is worse on coverage completeness, the 'energy-efficient' claim should be replaced by a coverage-only claim.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim (abstract, contributions) is that the agent learns energy-efficient coverage strategies. The experiments, however, never report the optimized quantity. Fig. 5 reports training success and curriculum difficulty; Fig. 6 reports success rate versus difficulty; Figs. 7–8 show trajectories. None of these report total power integral ∫_0^T P(t)dt from Eq. 17, path length, turn/curvature statistics, or any comparison against a standard CPP heuristic (e.g., boustrophedon/lawnmower) or a simpler learned baseline. Without such evidence, a high completion rate only establishes coverage; a policy can finish maps with arbitrary loops, and the authors themselves note that loops and trajectory intersections occur (Sec. V-B.3). Since the reward is only a power penalty plus terminal completion bonus, the existence of a learned power-efficient strategy is exactly the claim that needs quantitative support. This concern is distinct from and prior to the reader's point about unvalidated coefficients A and B: even if the power model were perfect, the paper would still lack evidence that the trained agent minimizes it. The absence of any energy metric or baseline is therefore the load-bearing gap in the paper's central claim.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper formulates continuous-world coverage path planning (CPP) for a fixed-wing UAV as a power-minimization problem with complete coverage, no-fly-zone avoidance, and roll-angle/curvature constraints. The environment is modeled with axis-aligned rectangles, UAV motion with quartic Bézier curves, and the problem is cast as an SCMDP solved by an action-mapping-based Soft Actor-Critic (AM-SAC) agent with a self-adaptive curriculum and an attention-based neural network. The authors train three agents on procedurally generated scenarios, report training curves and success rates versus difficulty, and show trajectory examples on procedurally generated and hand-crafted maps.","tokens_in":10934,"tokens_out":4768,"duration_ms":49058,"significance":"If the central claim is established, the paper would be a useful step toward continuous, energy-aware CPP for fixed-wing UAVs: the Bézier-curve action space with continuity constraints is carefully derived, the curriculum design is transparent and adaptive, and the attention-based set representation is a reasonable way to handle variable-size environment descriptions. However, the current evidence does not substantiate the abstract's claim of 'energy-efficient coverage strategies.' The reported experiments contain no measurement of the optimized quantity, no comparison to any baseline planner, and no independent validation of the power model. The kinematic formulation and architecture are interesting, but the empirical support for the paper's main claim is the load-bearing weak point.","major_comments":[{"comment":"The central claim, as stated in the abstract and contributions, is that the agent learns energy-efficient coverage strategies, but no result reports the optimized objective ∫_0^T P(t)dt from Eq. (18), nor path length, curvature/turn statistics, coverage overlap, or any other energy-related quantity. Success rate alone establishes coverage completion, not energy efficiency, and the paper itself notes in Section V-B.3 that loops and trajectory intersections occur. This is a load-bearing gap: without reporting the power integral or a proxy such as path length and curvature, the paper does not support its central claim.","section":"V-B (Figs. 5-8)"},{"comment":"There is no comparison to any baseline planner, such as a boustrophedon/lawnmower heuristic, the related fixed-wing planner of [6], or a simpler learned policy. Without such a comparison, the claimed effectiveness of the method is not quantitatively established relative to existing approaches, and the success rates around 0.8 at maximum difficulty—which equals the curriculum success threshold s_th=0.8—are hard to interpret as strong performance.","section":"V-B"},{"comment":"The power model P(t) = A/(v cos^2 φ(t)) + B v^3 with A=1130.97 and B=0.01353 is taken from the authors' earlier work [22] and is used both in the optimization objective and in the reward. The paper provides no validation of this model against real flight data, no sensitivity analysis, and no measurement of actual energy in the experiments. If the model coefficients are inaccurate, the learned 'energy-efficient' paths may not be energy-efficient in practice; this concern is distinct from the missing evaluation metric in the first major comment.","section":"III-D and IV-A.4"}],"minor_comments":[{"comment":"The caption of Fig. 5 says 'Training curve of one agent' but the figure shows three agents and the text discusses agents 1-3; please correct the caption.","section":"V-B.1 and Fig. 5"},{"comment":"The phrase 'the performs drops rapidly' should be 'the performance drops rapidly.'","section":"V-B.1"},{"comment":"The feasibility model checks spline points for curvature, map bounds, NFZ intersection, and length, but the number of sampled points is not specified. Since the check's soundness depends on adequate sampling density, please state the sample count and any sensitivity analysis.","section":"IV-A.3"},{"comment":"The notation x(t),v(t),κ(t)∈C0([0,T]) only encodes continuity of those functions, while the text describes 'continuity of position, velocity, and curvature' (i.e., C2 continuity of x). Please align the notation with the intended smoothness requirement.","section":"Eq. (22) and III-A"},{"comment":"Several references are incomplete: for example, [2] and [7] lack journal names and years in the formatted bibliography. Please complete all references.","section":"References"},{"comment":"The hand-crafted maps are named 'Manhattan32' and 'TUM50' but no description of their layout is given; please clarify what features make them out-of-distribution relative to the procedurally generated training scenarios.","section":"Fig. 8"}],"recommendation":"major_revision","confidential_remarks":"The manuscript applies the authors' previously introduced AM-SAC, feasibility-policy, and power-model machinery to a new CPP setting. This is not in itself a problem, but the revision should clearly delineate which components are new to this paper. The main issue is the evaluation: the absence of energy/path metrics and baselines must be addressed before the paper can support its central claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague—\n\nThe new thing here is a complete DRL pipeline for continuous, curvature-constrained fixed-wing coverage. The authors formulate the problem carefully as an SCMDP with a Bézier-curve action space enforcing position, velocity, and curvature continuity, and they train a feasibility policy plus an AM-SAC objective policy with a self-adaptive curriculum that deliberately omits coverage rewards. The training story is plausible: success rates reach around 0.8 at maximum difficulty, and the agent generalizes to out-of-distribution hand-crafted maps with visibly smooth paths. That is a genuine engineering contribution.\n\nWhat the paper does not do is measure the quantity it claims to minimize. The abstract and contributions say the method learns energy-efficient strategies, but no experiment reports integrated power consumption, path length, or turn/curvature statistics, and there is no comparison against a boustrophedon/lawnmower heuristic or a simpler learned baseline. A high success rate only establishes coverage; the agent can finish maps with inefficient loops, and the authors acknowledge loops occur. The power coefficients A and B are taken from the authors' own earlier work and are not validated here. Even if the coefficients were perfect, the paper would still lack evidence that the trained policy minimizes that model. This is the load-bearing gap.\n\nI also want to note the heavy self-citation. It is not a problem in itself because the prior AM-SAC, feasibility-policy, and power-model work is genuinely theirs and directly relevant. But the paper leans on those results without making the dependency explicit enough; a reader should check [15], [19], and [22] before trusting the training pipeline.\n\nThe soft spots are fixable. Add a lawnmower baseline and a simple RL baseline, report the actual power integral and path length for both the learned policy and the baselines, and either validate the power model with real flight data or clearly label it as a limitation. That would make the energy-efficiency claim supportable.\n\nWho is this for: people actively working on RL for fixed-wing path planning and curriculum design will find the formulation and the no-coverage-reward curriculum worth studying. It is not yet a paper from which you can deploy an energy-efficient planner. It deserves a serious referee: the novelty is real and the engineering is non-trivial, but the revision should be major rather than minor. I would recommend sending it to peer review with a request for those missing experiments.","headline":"Useful RL-based CPP pipeline for fixed-wing, but the energy-efficiency claim is unmeasured and unbashed; deserves peer review with required revisions.","tokens_in":11448,"tokens_out":2312,"would_cite":false,"duration_ms":22739,"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 deep reinforcement learning agent can plan complete coverage paths for fixed-wing UAVs directly in continuous space, minimizing power while avoiding no-fly zones.","keywords":["coverage path planning","deep reinforcement learning","fixed-wing UAV","action mapping","soft actor-critic","curriculum learning","Bézier curves","power minimization"],"falsifier":"Instrument a fixed-wing UAV with a power/current sensor, execute one of the learned trajectories, and compare the integrated measured energy against $\\int P(t)\\,dt$ predicted by the model; a large systematic mismatch would show that the learned paths are not demonstrably energy-efficient simply because the objective was wrong.","tokens_in":10433,"feed_emoji":"✈️","tokens_out":5524,"duration_ms":49471,"temperature":0.7,"pith_summary":"This paper argues that continuous-world coverage path planning for fixed-wing UAVs—traditionally done on grids—can be handled directly in the continuous plane by a reinforcement learning agent. The agent represents the map as variable-size axis-aligned rectangles, plans motion as curvature-constrained Bézier curves, and minimizes a power-consumption objective while covering target zones. To make learning tractable, the authors pair an action-mapping feasibility policy with a Soft Actor-Critic objective policy and a self-adaptive curriculum that grows the map area as performance improves. In experiments the trained agent solves procedurally generated scenarios with success rates around 0.8 at maximum difficulty and generalizes to hand-crafted maps, supporting the claim that learning-based continuous planning is viable.","feed_headline":"Deep RL plans continuous coverage paths for fixed-wing UAVs","feed_subtitle":"Action-mapped SAC with a self-adaptive curriculum reaches ~80% success on hard generated maps.","key_machinery":"The central mechanism is the action-mapping-based Soft Actor-Critic (AM-SAC) pipeline: an approximate feasibility model labels Bézier-control-point actions as feasible or not, a feasibility policy is pretrained to mimic that model, and an objective policy is then trained with SAC to maximize return among feasible actions. The action space itself does the geometric work: two control points are fixed by position and velocity continuity, the third by the current curvature, and the remaining parameters map to offsets along the velocity and normal directions, so every action produces a spline that is automatically continuous and curvature-limited. A progress variable $p_{k+1} = \\tau s_k d_k + (1-\\tau)p_k$ with success threshold $s_{th}$ governs the self-adaptive curriculum that scales map area during training.","core_discovery":"The paper's central claim is that complete coverage of continuous environments by a fixed-wing UAV can be learned directly as a reinforcement-learning task without grid discretization, optimizing power. Every $T_b$ seconds the agent chooses the control points of the next quartic Bézier curve, which enforces continuity of position, velocity, and curvature at the join. A feasibility model checks each candidate curve against map bounds, no-fly zones, maximum curvature, and a length range; the feasibility policy is pretrained on this model, and then the objective policy is trained with SAC to minimize power plus a single finish reward. A curriculum increases difficulty adaptively from a filtered success-progress variable, making it possible to learn without any coverage-based reward shaping. Results show success rates above $0.8$ at maximum difficulty and smooth, mostly overlap-free trajectories on unseen and hand-crafted maps, indicating the agent has learned a general coverage strategy.","pith_inferences":["The axis-aligned rectangle restriction is likely the next bottleneck; a polygon-based observation and feasibility model would be a natural extension, and the attention architecture already supports variable-sized sets.","Because the feasibility policy is approximate, the method inherits occasional constraint violations; replacing it with a differentiable feasibility layer could make the guarantees exact without a separate pretraining phase.","The curriculum's success-progress filter could be transferred to other sparse-reward long-horizon RL problems (e.g., search and rescue or exploration) where task difficulty is continuously adjustable.","The reported success rate around 0.8 at maximum difficulty suggests that pushing the same architecture with more computation or with equivariance-based augmentations could close the remaining safety gap."],"forward_implications":["Continuous-world CPP can be solved end-to-end by deep RL without grid representations, preserving smooth, autopilot-feasible trajectories.","The self-adaptive curriculum allows the agent to learn long-horizon coverage tasks from a reward that contains no coverage signal, avoiding premature greedy behavior.","The trained agent generalizes from procedurally generated maps to hand-crafted Manhattan-style layouts, suggesting the training distribution covers realistic no-fly-zone structures.","Because power depends on roll angle, the learned trajectories actively trade coverage efficiency against turn sharpness, pointing to energy-aware planning as a concrete objective."],"supporting_citations":[{"why":"Supplies the action-mapping technique that decouples constraint satisfaction from objective-driven action selection.","marker":"[15]"},{"why":"Provides the Soft Actor-Critic algorithm used to train the objective policy.","marker":"[18]"},{"why":"Derives the base power model that the simplified two-term energy objective is built from.","marker":"[21]"},{"why":"Provides the power-model coefficients A and B used in the simulations.","marker":"[22]"},{"why":"Defines the State-wise Constrained MDP formalism in which the problem is cast.","marker":"[16]"},{"why":"Supplies the grid-based CPP approach and the hand-crafted Manhattan32 and TUM50 map scenarios used for out-of-distribution evaluation.","marker":"[8]"},{"why":"The autopilot framework that motivates the assumption that Bézier spline paths are trackable.","marker":"[14]"}],"fun_headline_variants":["Deep RL masters continuous fixed-wing coverage without grids","Deep RL ditches grids for continuous fixed-wing coverage","Action-mapped SAC clears 80% on hard continuous fixed-wing maps","Self-adaptive curriculum unlocks continuous fixed-wing coverage","Fixed-wing UAVs learn energy-efficient coverage via deep RL"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the two-term power model $P(t) = A/(v\\cos^2\\phi(t)) + Bv^3$ with coefficients $A=1130.97$ and $B=0.01353$ accurately reflects real energy consumption; the paper gives no validation of this model against measured flight data, and experiments never measure actual power.","fun_headline_variants_meta":{"raw":{"variants":["Deep RL masters continuous fixed-wing coverage without grids","Deep RL ditches grids for continuous fixed-wing coverage","Action-mapped SAC clears 80% on hard continuous fixed-wing maps","Self-adaptive curriculum unlocks continuous fixed-wing coverage","Fixed-wing UAVs learn energy-efficient coverage via deep RL"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001491,"raw_usage":{"total_tokens":5934,"prompt_tokens":844,"completion_tokens":5090,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":460,"completion_tokens_details":{"reasoning_tokens":5011}},"tokens_in":460,"tokens_out":5090,"duration_ms":33508,"temperature":1.0,"reasoning_tokens":5011,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-15T21:55:50.799383+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Instrument a fixed-wing UAV with a power/current sensor, execute one of the learned trajectories, and compare the integrated measured energy against $\\int P(t)\\,dt$ predicted by the model; a large systematic mismatch would show that the learned paths are not demonstrably energy-efficient simply because the objective was wrong.","supporting_citations":[{"cited_title":"uavEE: A modular, power-aware emulation environment for rapid prototyping and testing of UA Vs,","cited_arxiv_id":null,"evidence_quote":"Provides the power-model coefficients A and B used in the simulations."},{"cited_title":"A high-fidelity, low- order propulsion power model for fixed-wing electric unmanned air- craft,","cited_arxiv_id":null,"evidence_quote":"Derives the base power model that the simplified two-term energy objective is built from."},{"cited_title":"Action Mapping for Reinforcement Learning in Continuous Environments with Constraints","cited_arxiv_id":"2412.04327","evidence_quote":"Supplies the action-mapping technique that decouples constraint satisfaction from objective-driven action selection."},{"cited_title":"Soft actor-critic: Off- policy maximum entropy deep reinforcement learning with a stochastic actor,","cited_arxiv_id":null,"evidence_quote":"Provides the Soft Actor-Critic algorithm used to train the objective policy."},{"cited_title":"Learning to Recharge: UAV Coverage Path Planning through Deep Reinforcement Learning","cited_arxiv_id":"2309.03157","evidence_quote":"Supplies the grid-based CPP approach and the hand-crafted Manhattan32 and TUM50 map scenarios used for out-of-distribution evaluation."},{"cited_title":"uavAP: A modular autopilot framework for UA Vs,","cited_arxiv_id":null,"evidence_quote":"The autopilot framework that motivates the assumption that Bézier spline paths are trackable."}],"review_version":1}