{"id":"8ae2d269-b26b-4144-b70d-24e2028792d2","arxiv_id":"2607.15690","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A GCS-OCP hybrid planner produces wind- and obstacle-aware 4D eVTOL trajectories across ascent, cruise, and descent, handling up to 250 buildings in NYC simulations.","lead":"This paper combines a graph-based path search with an optimal-control optimizer to plan four-dimensional eVTOL flight trajectories around New York City buildings, including wind and battery constraints. If it holds up, urban-air-mobility operators could use it offline to find fast, feasible routes through dense cityscapes instead of relying on straight-line estimates.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Candidate-set incompleteness may undermine reported flight-time comparisons.","rationale":"The reader's weakest assumption is exactly the candidate-set completeness, and I agree. The paper's hybrid approach separates global planning from optimal control, but the global planner is heuristic: it uses a convex relaxation, randomized rounding, and top-few pruning. The final OCP is only as good as the corridors provided. Since the objective is travel time, missing a better corridor cannot be compensated later. The paper acknowledges no optimality guarantee (Sec. V-D) but does not bound the suboptimality. This is load-bearing because the headline quantitative claims ('20% underestimate', Table II comparisons) are all computed relative to the GCS-OCP solution; if that solution is far from optimal, the comparisons lose their meaning. The D-OCP circular-approximation issue is also real, but it is a baseline fairness problem that could be fixed by using polygons in D-OCP; the candidate issue is intrinsic to the proposed method. Therefore the central claim needs a sensitivity analysis or a completeness argument. I would keep the CONDITIONAL verdict: the paper is a useful framework but the optimality/completeness of the candidate set is not established. The concrete test of increasing the candidate pool and observing whether the output changes would settle the concern.","tokens_in":13191,"tokens_out":7902,"duration_ms":62107,"concrete_test":"For Origin 2 at 136 m with 112 obstacles, rerun the candidate generation with 100 random sequences instead of the reported “top few” (e.g., 5). Run the OCP on the top 20 additional sequences not in the original C. If the minimum t_F improves by >2%, the original pruning loses meaningful corridors and the reported times are not representative. If no improvement, the heuristic is likely sufficient in this case.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim that GCS-OCP yields low-travel-time trajectories hinges on the candidate set C (Sec. III-B) containing a polygon sequence close to the true optimal. GETCANDIDATEPATHS solves a convex relaxation of the mixed-integer GCS problem (Eqs. 17–24), rounds via random sampling, and retains only “top few” sequences. No completeness or suboptimality bound is provided. Since the OCP in Sec. III-C is restricted to corridors in C, any missed corridor is unrecoverable (Eqs. 26–27). The authors explicitly admit (Sec. V-D) that the final solution “is not guaranteed to be truly optimal,” but they do not quantify the gap. This affects the headline 20% underestimate: if GCS-OCP overestimates the feasible optimum, the true underestimate by BL1 could be smaller, and the D-OCP comparison (Table II) could be biased. Thus the evidence presented does not rule out that the candidate heuristic is the bottleneck.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript proposes GCS-OCP, a hybrid framework for 4D (time, latitude, longitude, altitude) eVTOL trajectory design in dense urban environments. The method first decomposes the obstacle-free cruise plane into intersecting convex polygons using IRIS, then generates a small set of candidate polygon sequences via a Graph of Convex Sets (GCS) planner with randomized rounding of a convex relaxation, and finally solves an optimal control program (OCP) for each candidate to obtain a time-minimizing trajectory that respects wind, eVTOL flight dynamics, actuator/battery limits, and phase-specific constraints. The framework is evaluated on routes to JFK airport in New York City with up to roughly 250 building obstacles. The reported results show that a back-of-the-envelope baseline (BL1) underestimates end-to-end flight time by as much as 20% and that GCS-OCP often yields lower flight times than a direct OCP baseline (D-OCP) with circular obstacle approximations, while D-OCP fails to converge in some high-obstacle cases. Limitations (offline-only, local optimality, static uniform wind, stationary obstacles) are explicitly acknowledged in Sec. V-D.","tokens_in":1579,"tokens_out":1395,"duration_ms":74598,"significance":"If the claims are validated, the framework is a useful offline tool for UAM trajectory planning, integrating real building footprints, weather data, a reasonably detailed multirotor dynamics model, and multi-phase operation. The paper's strengths are its use of real NYC data, the clear separation of global discrete planning and continuous optimization, and the explicit enumeration of limitations. However, the two headline claims—the 20% underestimation by simplified baselines and the superiority over D-OCP—rest on evidence that needs strengthening. In particular, the GCS candidate set has no completeness or suboptimality guarantee, and the D-OCP comparison is biased by unequal obstacle representations. These issues are fixable and should be addressed in a revision. The paper does not oversell its contribution; it is candid that the OCP solution is not globally optimal and that the method is offline.","major_comments":[{"comment":"Candidate-set completeness: The GCS planner solves a convex relaxation, rounds via random sampling, and retains only the 'top few' candidate sequences. No suboptimality or completeness bound is given, and the final OCP is restricted to these candidates (Eqs. 26-27). While Sec. V-D acknowledges that the result is not guaranteed to be truly optimal, the effect on the headline 20% underestimate and on the D-OCP comparison is not quantified. Please report the gap between the convex relaxation lower bound and the selected candidate cost, or compare against exhaustive enumeration for small obstacle counts. At minimum, discuss how candidate restriction could change the reported flight-time differences.","section":"Sec. III-B, Eqs. (17)-(24); Sec. V-D"},{"comment":"Unequal obstacle representations bias the comparison: D-OCP approximates each building as a circle with radius equal to the maximum vertex distance from the centroid (a conservative enlargement), while GCS-OCP uses exact polygon planforms. This makes D-OCP appear to fly longer routes and may inflate the reported advantage of GCS-OCP. Re-run D-OCP with the same polygonal obstacles (e.g., using distance-to-polygon constraints) or provide a sensitivity analysis with inscribed/circumscribed circles to establish the claimed travel-time benefit of GCS-OCP.","section":"Sec. II-D vs Sec. III-A; Table II"},{"comment":"The 'did not converge' (DNC) entries for D-OCP are attributed to nonconvexity, but they could be artefacts of IPOPT's local optimization and poor initialization. The paper does not describe how D-OCP is initialized. Provide initial guesses (e.g., from BL1 or a simple collision-free path), try multiple random restarts, and report the best result. Without this, the claim that D-OCP is less scalable than GCS-OCP is not conclusively supported.","section":"Table II, D-OCP rows"},{"comment":"The GCS-OCP obstacle constraint is stated for discretized trajectory points, but the mapping of cruise segments to polygons is ambiguous: Eq. (29) appears to apply to all cruise indices rather than the subset belonging to each polygon-constrained segment. Since the continuous trajectory between grid points is assumed straight-line, clarify how segment-to-polygon mapping is implemented and verify (e.g., by dense sampling) that the entire continuous trajectory remains inside the safe polygons, not just the discrete collocation points.","section":"Sec. IV-C, Eq. (29)"}],"minor_comments":[{"comment":"Typesetting issue: the induced velocity equation should read v_i = v_h^2 / sqrt((V cos α)^2 + (V sin α + v_i)^2).","section":"Eq. (2)"},{"comment":"Typo in bullet: 'topassthrough' should be 'to pass through'.","section":"Sec. I-B"},{"comment":"The table is wide and the subcolumns (Flight time, Run time, IPOPT time) for D-OCP and GCS-OCP are not visually separated; add clear column groupings or split into two tables for readability.","section":"Table II"},{"comment":"Specify the color correspondences in the obstacle panel (obstacles, convex polygons, trajectory) in the caption itself, as the caption currently references colors only implicitly.","section":"Fig. 2"}],"recommendation":"major_revision","confidential_remarks":"The paper is a competent engineering contribution with a useful application, but the central comparisons need rework. The candidate-set optimality gap and the D-OCP circular-vs-polygonal obstacle mismatch are the main technical obstacles. The authors are transparent about limitations, which is commendable. If the comparison is made fair and the optimality gap is quantified, the paper could be acceptable for an applied trajectory-planning venue."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"This paper is a well-executed integration of off-the-shelf components: IRIS convex decomposition, GCS-based shortest-path over convex sets, and a nonlinear OCP with eVTOL dynamics, wind, battery limits, and multi-phase constraints. That integration is genuinely new for the UAM trajectory-design problem, and the NYC case study with real building footprints and wind is a concrete step beyond toy examples. The authors are also honest about their limitations — offline only, static wind, stationary obstacles — which is more than many papers in this space manage.\n\nThe strongest part is the pipeline itself: candidate corridors from GCS are refined by OCP, and the resulting trajectories respect the hard constraints and look plausible. The 20% underestimate from the naive BL1 baseline is directionally convincing, because any realistic modeling will add time relative to a straight-line constant-speed assumption. The scaling claim to 250 obstacles also holds up as a demonstration that the convex-containment reformulation is more tractable than direct nonconvex obstacle constraints.\n\nThe soft spots are real but not fatal. The GCS candidate set is generated by a randomized rounding of a convex relaxation, retaining only the top few sequences. There is no completeness or suboptimality bound, so the OCP is restricted to corridors that may not include the true time-optimal route. The authors acknowledge this in Section V-D, but they do not quantify the gap. That means the reported GCS-OCP times are potentially overestimates, and the 20% number could be smaller if a better corridor existed. The D-OCP comparison also has a fairness issue: D-OCP uses circular approximations of buildings while GCS-OCP uses exact polygons, and the D-OCP non-convergence might be solver-specific — IPOPT with one settings file is not a definitive verdict.\n\nThese issues are addressable. A sensitivity analysis on the number of candidates and seed spacing, a fairer D-OCP baseline with polygonal obstacles or better solver tuning, and a comparison against an exhaustive corridor search on small instances would tighten the claims considerably. None of this undermines the core architecture; it just means the paper claims a bit more than the evidence fully supports.\n\nFor a reader working in UAM planning, this is worth engaging with. It is a useful integration paper and the results are reproducible enough to build on. My recommendation: send it to peer review with a request for the sensitivity and baseline fairness work. It is not a desk reject, but it needs revision before the quantitative claims can be trusted.","headline":"A solid integration of known planning pieces that makes a plausible case for GCS-OCP in dense urban eVTOL planning, but the headline 20% figure leans on a candidate-corridor search with no completeness guarantee.","tokens_in":13883,"tokens_out":1621,"would_cite":true,"duration_ms":16709,"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 hybrid GCS-OCP framework designs obstacle-aware, time-minimizing four-dimensional eVTOL trajectories in dense urban environments, and shows that ignoring such constraints underestimates flight time by as much as 20%.","keywords":["Urban Air Mobility","eVTOL trajectory design","four-dimensional trajectory","Graph of Convex Sets","optimal control","obstacle avoidance","wind-aware planning","New York City"],"falsifier":"Take a small obstacle field with a few dozen polygons, exhaustively enumerate all valid polygon sequences from start to end, solve the OCP for each, and compare the minimum over all sequences with the minimum over the top-few candidates returned by the randomized selection; if any omitted sequence yields a lower flight time, the candidate-set assumption fails.","tokens_in":13085,"feed_emoji":"🚁","tokens_out":5593,"duration_ms":51079,"temperature":0.7,"pith_summary":"This paper claims that a two-stage hybrid framework can produce time-minimizing four-dimensional trajectories for eVTOL aircraft in dense urban environments, respecting building obstacles, a constant wind field, nonlinear flight dynamics, and operational limits across ascent, cruise, and descent. The first stage fills obstacle-free space with intersecting convex polygons and searches a graph of those polygons for promising low-time corridors; the second stage solves an optimal control problem that constrains the cruise trajectory to pass through the selected polygons. Evaluated on New York City routes, the framework designs trajectories in scenes with up to 250 building obstacles, where a directly formulated nonconvex optimal control approach fails to converge. The paper argues that ignoring wind, obstacles, and accurate dynamics understates end-to-end flight time by as much as 20%, and that passing through convex safe regions yields lower flight times than actively avoiding obstacles.","feed_headline":"Ignoring obstacles understates eVTOL flight time by 20%","feed_subtitle":"A two-stage planner designs 4D routes respecting buildings, wind, and dynamics in up to 250-obstacle scenes.","key_machinery":"The load-bearing mechanism is the pair (GCS planner, OCP): the Graph of Convex Sets (GCS) planner selects a short candidate list of convex polygon sequences through obstacle-free airspace (using IRIS-generated polygons as vertices), and the Optimal Control Program (OCP) then finds the minimum-time trajectory that satisfies full eVTOL dynamics and operational constraints while forcing each cruise segment to lie inside the assigned polygon. The convex containment constraint is what converts a nonconvex, poorly scaling obstacle-avoidance problem into a convex per-polygon constraint set, enabling the framework to scale to 250 obstacles.","core_discovery":"The central claim is that the time-optimal eVTOL trajectory problem in cluttered urban airspace becomes tractable when obstacle avoidance is reformulated as convex containment: instead of requiring the trajectory to stay outside each building, the framework requires it to stay inside a sequence of overlapping obstacle-free convex polygons. This reformulation, combined with a hybrid of a graph-of-convex-sets path planner and a full optimal control program, yields dynamically feasible trajectories that estimate not only path coordinates but also heading, airspeed, acceleration, thrust, pitch, roll, power, and energy consumption in ascent, cruise, and descent. The authors report that this frame","pith_inferences":["If the 20% underestimation holds across other cities and altitude bands, prior system-level UAM analyses that assume straight-line constant-speed trajectories may need recalibration, since their time and energy estimates embed the same omission.","The convex-containment idea transfers to other nonconvex trajectory planning problems (fixed-wing urban delivery, autonomous ground vehicles) wherever a graph of convex safe sets can be constructed.","Because the candidate set is formed by randomized sampling from a relaxed GCS solution, the reported optimum depends on the random draws; an exhaustive or completeness-guaranteed corridor enumeration would be needed to certify that the returned trajectory is globally optimal.","The framework's sensitivity to wind is likely understated by the constant-field experiments; the GCS edge weights already depend on ground speed along each edge, so a spatially varying wind field could change corridor selection and final time in ways the current static-wind results do not capture."],"forward_implications":["UAM operators can generate offline, dynamically feasible, time-minimizing trajectories in dense urban environments with up to 250 building obstacles, and can extract full state and control profiles (including power and energy) for each phase.","Simplified trajectory estimates that ignore wind, obstacles, and detailed dynamics may be optimistic by up to 20% in flight time, which propagates to vertiport placement and air-corridor planning studies that rely on such estimates.","Enforcing passage through convex obstacle-free regions produces lower travel times than directly enforcing obstacle avoidance in an OCP, by up to 22% in the tested New York City scenarios.","The framework scales better with obstacle count than a directly formulated nonconvex OCP: at low obstacle counts the direct approach is cheaper, but as obstacles grow the direct approach's variables and constraints grow faster and it can fail to converge (e.g., at 250 obstacles)."],"fun_headline_variants":["Building-aware 4D routes cut eVTOL time errors by 20%","Convex polygons make eVTOL routes 20% more accurate","Two-stage planner handles 250 obstacles for eVTOL","eVTOL trajectory design gets obstacle-proof in 4D","Hybrid planner beats 20% time error in urban air mobility"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The final optimal control solve can only be as good as the candidate polygon sequences the GCS planner selects, and the paper does not prove that the true time-optimal, dynamically feasible trajectory is among the top few randomized candidates.","fun_headline_variants_meta":{"raw":{"variants":["Building-aware 4D routes cut eVTOL time errors by 20%","Convex polygons make eVTOL routes 20% more accurate","Two-stage planner handles 250 obstacles for eVTOL","eVTOL trajectory design gets obstacle-proof in 4D","Hybrid planner beats 20% time error in urban air mobility"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000231,"raw_usage":{"total_tokens":1328,"prompt_tokens":755,"completion_tokens":573,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":499,"completion_tokens_details":{"reasoning_tokens":492}},"tokens_in":499,"tokens_out":573,"duration_ms":5474,"temperature":1.0,"reasoning_tokens":492,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-01T22:33:35.015255+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Take a small obstacle field with a few dozen polygons, exhaustively enumerate all valid polygon sequences from start to end, solve the OCP for each, and compare the minimum over all sequences with the minimum over the top-few candidates returned by the randomized selection; if any omitted sequence yields a lower flight time, the candidate-set assumption fails.","supporting_citations":[],"review_version":1}