{"id":"bf25573e-a41e-45ce-9a67-0d767c4460db","arxiv_id":"2501.06694","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"Warm-starting robust low-thrust trajectory optimizers with solutions to earlier non-robust problems improves feasibility and solution quality, but cumulative gains are mixed when seed-generation costs are included.","lead":"This paper compares two ways to create starting guesses for optimizing low-thrust space trajectories that must stay workable if a thruster fails for a while. In a simulated Lunar Gateway mission, the method that borrows starting guesses from easier no-failure problems found more feasible solutions faster, although the advantage shrinks once the cost of creating those easier solutions is counted.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Cumulative time-to-solve for conditional search is undercounted: seed cost is not divided by the conditional feasibility ratio, so the claimed efficiency gain at P2/P3 may be an artifact.","rationale":"I read the paper as an empirical comparison of two initial-guess strategies for robust low-thrust trajectory design. The direct-metric results—feasibility ratio, per-solution time, and Δv distribution—are informative, and the warm-starting idea is plausible. However, the paper's own cumulative-metric discussion is internally inconsistent: feasibility is accumulated multiplicatively (Fig. 16), which correctly reflects that extra seeds are needed when conditional success rates are below 100%, but time is accumulated additively (Fig. 17), which omits the same multiplier. This is not a disagreement about modeling assumptions; it is an arithmetic inconsistency in the precise metric used to support the computational-efficiency part of the central claim. The proposed check is a straightforward recomputation from the authors' own run logs, so it can settle the issue without external data. If the corrected figure preserves the conditional advantage, the central claim survives; if not, the abstract's 'significantly improves convergence rate' overclaim should be softened to a direct-metric result only. I therefore keep the reader's CONDITIONAL verdict but for a different, more load-bearing reason than the MTE discretization concern.","tokens_in":26750,"tokens_out":9978,"duration_ms":100083,"concrete_test":"Recompute Figure 17 using the corrected cumulative time per final solution: for S(k|0), use T_0 / F_{k|0} + T_{k|0}, and analogously for S(k|k') use T_{k'} / F_{k|k'} + T_{k|k'}, where F denotes the direct conditional feasibility ratio and T denotes the reported time per feasible solution. If, after this correction, the cumulative time for S(k|0) exceeds S(k) at any k/Δτ combination where the paper claims an advantage, the headline efficiency claim fails and the figure must be revised.","verdict_should_be":"CONDITIONAL","load_bearing_attack":"The central efficiency claim depends on the cumulative time metric in §VI (Figs. 16–17). For S(k|0), the paper defines cumulative solving time as the sum of the S(0) time-per-solution and the S(k|0) time-per-solution. This ignores that each conditional initial guess consumes one P0 seed, and only a fraction F_{k|0} of conditional guesses succeed. If T_0 is the S(0) time per feasible P0 solution and T_{k|0} is the conditional time per feasible solution, the correct per-final-solution seed cost is T_0 / F_{k|0}, not T_0. The paper's own feasibility data (Fig. 14) show F_{k|0} is often at or below single-digit percentages, so the omitted factor is large (e.g., F_{k|0}=0.1 makes the seed term ten times larger). Since Fig. 16 already shows non-conditional cumulative feasibility is higher at P1 and P2, correcting this term is likely to erode or reverse the claimed conditional advantage in cumulative solving time at higher k. The conclusion that 'conditional methods achieve better cumulative average solving times at higher robustness depths' is therefore not currently supported; it may be an artifact of inconsistent accounting between the feasibility and time metrics.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper considers a finite-realization robust optimal control formulation for low-thrust transfers subject to missed-thrust events (MTEs), with at most one MTE and a small set of initiation times and durations. It compares two global-search initialization strategies: a non-conditional strategy S(k), which uses uniform sampling with monotonic basin hopping, and a conditional strategy S(k|k'), in which solutions to a less robust problem P_{k'} are mapped into initial guesses for P_k. The comparison uses feasibility ratio, average solving time, and Delta-v, together with cumulative versions of the feasibility and time metrics that account for seed-generation cost. In a Lunar Gateway Power and Propulsion Element case study, the authors report that S(k|0) improves feasibility and solution quality over S(k) and discuss how these advantages evolve with depth of robustness k.","tokens_in":26968,"tokens_out":7739,"duration_ms":76563,"significance":"The question is practical and timely: if the empirical advantage of conditional seeding is robust, this is a useful, low-cost heuristic for preliminary robust low-thrust trajectory design. The paper's strengths include a realistic high-dimensional cislunar case study, a clear problem transcription with analytic derivatives drawn from prior work, and a genuine attempt to define cumulative metrics so that seed-generation overhead is not ignored. The main limitations are inferential: the headline claim of statistical significance is not backed by any test or confidence interval, and the cumulative time accounting appears to undercount seed cost. These issues affect the central claim but are addressable within the manuscript's scope.","major_comments":[{"comment":"The abstract and §I state that the conditional approach 'significantly improves' convergence rate and solution quality, but no statistical test, confidence interval, or bootstrap analysis appears anywhere in §VI. The sample sizes at higher depths are very small: Table 8 reports only 3–8 feasible S(3) solutions with feasibility ratios of 0.08–0.22%, so the point-estimate comparisons in Figs. 11–13 and 14–15 may lie within sampling noise. Please provide two-sample bootstrap intervals for the Δv distributions and binomial confidence intervals for the feasibility ratios, or restrict the claims to the observed samples.","section":"§VI, Table 8 and Figs. 10–15"},{"comment":"The cumulative solving time for S(k|0) is defined as the sum of the S(0) time per solution and the S(k|0) time per solution. This undercounts the seed-generation cost: if each conditional solve consumes one feasible P0 seed and only a fraction F_{k|0} of conditional solves succeed, the expected number of P0 seeds per final conditional solution is 1/F_{k|0}, so the seed contribution is T_0/F_{k|0}, not T_0. Since Fig. 14 shows F_{k|0} can be near or below 0.1, this is at least a tenfold correction to the seed term. The §VII conclusion that conditional methods achieve better cumulative average solving times at higher robustness depths is not supported by the metric as defined; recompute the cumulative time with the expected seed usage or report total wall-clock time per final feasible solution for each complete pipeline.","section":"§VI, Fig. 17"},{"comment":"The validated scenario set is narrow: Assumption A1 excludes multiple MTEs, and the text states that the tested durations up to 1.5 days cover roughly 40% of observed outage durations. Because the relative difficulty of P_k versus P_0 depends on which MTE scenarios are included, the ranking of conditional versus non-conditional search has been demonstrated only for this particular subset of the risk distribution. Please add a sensitivity study (for example, longer outages, additional initiation locations, or a second MTE) or explicitly delimit the paper's claim to the tested scenario distribution.","section":"§II.B and Tables 6–7"}],"minor_comments":[{"comment":"The conditional strategy is written as S(k|k') ≡ π∘M_k^{k'}, but the definition two paragraphs earlier includes the seed-generation map π∘X_{k'}; either define M_k^{k'} to include that composition or fix the equation for consistency.","section":"§III.A, Eq. (9)"},{"comment":"The text says the analysis is restricted to a maximum of three MTE initiation points, but Table 6 lists ten segment indices for P1 and Table 7 lists pairs and triples of indices; reconcile the stated assumption with the actually tested sets.","section":"§II.B and Table 6"},{"comment":"The symbol N is defined twice in the nomenclature, once as 'number of segments' and once as 'number of decision variables'; please use distinct symbols.","section":"Nomenclature"},{"comment":"Please state explicitly whether 'Time/Solution' is wall-clock time per feasible solution or per initial guess; this distinction is necessary for interpreting the cumulative time metric in Fig. 17.","section":"§VI, Table 8"},{"comment":"Minor language errors remain, for example 'has been been explored' in §I and 'with with number' in §II.C; a careful copyedit is needed.","section":"Throughout"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Two things to know. First, the core empirical comparison is real: conditional warm-starting from non-robust solutions (S(k|0)) beats uniform sampling on direct metrics — feasibility ratio, time per solution, and Δv — across k=1,2,3 in a realistic PPE cislunar transfer. Second, the paper's headline efficiency claim for cumulative time at higher k does not survive the paper's own accounting. Figure 17 adds S(0) time per solution to S(k|0) time per solution, but each conditional solution consumes one P0 seed and only a fraction F_{k|0} of seeds lead to feasible solutions. The seed term should be T_0 / F_{k|0}, which can be an order of magnitude larger. With F_{k|0} often in single digits, the claimed cumulative time advantage at P2/P3 is likely an artifact.\n\nWhat is new: the formal distinction between conditional and non-conditional initial guess generation for robust MTE design, the mapping operator count (though Eq. (10) has a typo — should be (k')^k + 1, not k' k + 1), and the adaptive segmentation strategy that keeps realization control authority consistent with the reference. The paper gives a systematic comparison, not an anecdote, and is honest about the limited MTE coverage (one MTE, three start locations, durations up to 1.5 days covering ~40% of observed outages) and the tiny sample sizes at k=3 (3–8 solutions).\n\nSoft spots in order of severity: (1) The cumulative time metric is wrong as defined. This is the load-bearing flaw for the 'conditional wins at higher k' conclusion. The authors need to recompute with seed cost divided by conditional feasibility, or drop the cumulative time claim. (2) The abstract and contribution list promise 'statistically significant improvements,' but no significance test or confidence interval appears anywhere. With P3 sample sizes in single digits, that language is unsupportable. (3) No code or data release; DyLAN is proprietary, so the numbers cannot be independently checked. (4) Eq. (10) is a concrete error.\n\nThe central idea — warm-start hard robust problems with solutions to easier ones — is plausible and the case study is well chosen. The direct metrics consistently favor S(k|0), so I don't think the paper is wrong in its qualitative finding; it's the quantitative efficiency claim that is currently unsupported.\n\nWho it's for: astrodynamics and mission-design practitioners. A serious referee should see it, but it needs major revision before acceptance: fix the cumulative accounting, add statistical tests or temper the language, and release data/code. Recommendation: send to peer review with a request for major revision.","headline":"Warm-starting robust MTE design from non-robust solutions looks genuinely better on direct metrics, but the paper's cumulative time accounting is off by a factor of 1/F and the 'significant' claim has no significance test behind it.","tokens_in":27515,"tokens_out":4667,"would_cite":false,"duration_ms":41122,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49K15","90C30","90C26","70M20"],"pacs":[],"model":"deepseek-v4-flash","headline":"Warm-starting robust low-thrust searches with non-robust solutions improves feasibility, speed, and fuel across missed-thrust depths.","keywords":["low-thrust trajectory design","missed-thrust events","robust optimal control","initial guess generation","conditional warm start","global search","cislunar transfer","nonlinear programming"],"falsifier":"Re-run the $P_1$, $P_2$, and $P_3$ comparisons with missed-thrust initiation points drawn from all 50 reference segments and durations sampled from the full historical Weibull distribution, including outages beyond 1.5 days, keeping cumulative seed-generation cost in the metrics; if $S(k|0)$ no longer dominates $S(k)$ in feasibility or cumulative solving time at $k \\ge 2$, the paper's central claim fails.","tokens_in":26501,"feed_emoji":"🛰️","tokens_out":12304,"duration_ms":183793,"temperature":0.7,"pith_summary":"Low-thrust spacecraft missions are vulnerable to missed-thrust events, but designing trajectories that survive them is computationally difficult because the robust problem is a high-dimensional nonlinear program with few feasible regions. This paper proposes generating initial guesses for a robust problem $P_k$ by projecting previously computed solutions to a simpler problem $P_{k'}$ — in particular the non-robust problem $P_0$ — into the robust decision space, instead of sampling uniformly from the global space. In a Lunar Gateway Power and Propulsion Element transfer to a near-rectilinear halo orbit, the conditional strategy $S(k|0)$ attains higher feasibility ratios, lower average solving times, and lower propellant costs than non-conditional uniform sampling, with the advantage persisting through robustness depths $k=1,2,3$ even after accounting for seed-generation cost in cumulative metrics. The paper concludes that feasible-region access, not partial robustness of the seed, governs initial-guess quality.","feed_headline":"Seeding robust low-thrust searches with simpler solutions wins","feed_subtitle":"Non-robust seeds raise feasibility, cut solve time, and lower fuel on a Gateway transfer.","key_machinery":"The central object is the conditional initial-guess generator $S(k|k') = \\pi \\circ M^{k}_{k'}$, where $M^{k}_{k'}$ maps an optimal solution of a lower-robustness problem $P_{k'}$ into the decision space of $P_k$ and $\\pi$ is the solver's projection onto a local optimum. The mapping is not unique: realization-to-realization assignments are counted by $C^{k}_{k'} = k'^k + 1$, and the paper uses the simplest reference-to-realization mapping after excluding control segments before the missed-thrust event, with zero initialized realization coast time. This projection carries the argument by placing initial guesses in basins of attraction that survive increasing robustness depth, while adaptive segmentation of realization trajectories keeps control authority comparable and keeps the Jacobian sparse. The alternative, $S(k)$, samples uniform global distributions and is the exploration baseline whose decay in feasibility with $k$ is the contrast that makes the conditional advantage visible.","core_discovery":"The central claim is that the performance of global search for missed-thrust-robust low-thrust trajectories is controlled by how initial guesses are generated, and that conditioning those guesses on solutions of a less robust problem is systematically better than unconditional sampling. Concretely, the paper defines $P_k$ as the robust optimal control problem with $k$ realization trajectories, one per missed-thrust scenario, for $k=0,\\dots,3$, and compares $S(k)$, which samples a fixed global distribution, with $S(k|k')$, which maps optimal solutions of $P_{k'}$ into the higher-dimensional decision space of $P_k$. Across all tested outage durations, $S(k|0)$ yields the highest feasibility ratios, the lowest mean solving times, and the best (lowest) $\\Delta v$ distributions; $S(3|0)$ even beats seeds from partially robust problems $S(3|1)$ and $S(3|2)$, because seeds from more constrained problems often map into infeasible regions. The authors state that $S(k|0)$ appears to be the optimal initial guess generation strategy and interpret the result as evidence that feasible-region accessibility dominates partial-robustness information.","pith_inferences":["If the feasible-region-access explanation is general, then any method that broadens the basin coverage of low-robustness solutions, such as denser basin hopping around multiple non-robust optima, should amplify the conditional gain at high $k$.","The same conditional seeding idea may transfer to other high-dimensional robust trajectory problems with multiple realization scenarios, whenever the feasible set shrinks monotonically with robustness depth.","Because the paper tests only three missed-thrust locations and durations up to 1.5 days, an immediate testable extension is to condition on non-robust seeds under missed-thrust initiation at all 50 reference segments and durations drawn from the full historical Weibull distribution."],"forward_implications":["For robust missed-thrust design at depths $k=1,2,3$, initializing from non-robust solutions yields higher feasibility ratios and lower mean $\\Delta v$ than sampling from a fixed global distribution.","The advantage persists in cumulative metrics that include the cost of producing the seed solutions: conditional search has comparable or better cumulative feasibility at depth $k=3$ and better cumulative solving time at higher robustness depths.","Seeding from non-robust solutions outperforms seeding from partially robust solutions at depths 2 and 3, so robustness of the seed is secondary to feasible-region accessibility.","Conditional search reduces solution diversity, so a mission design phase wanting a broad family of options may need to seed from a diverse set of non-robust solutions or mix strategies."],"supporting_citations":[{"why":"supplies the historical missed-thrust frequency and duration data used to justify the single-event, finite-duration uncertainty model.","marker":"[1]"},{"why":"introduces the reference-plus-realization lifted formulation and the three-initiation-point approximation adopted here.","marker":"[18]"},{"why":"extends missed-thrust design to robust low-thrust problems and motivates the realization-based constraint structure.","marker":"[19]"},{"why":"provides the monotonic basin-hopping algorithm used to define the non-conditional search baseline.","marker":"[37]"},{"why":"specializes basin hopping to funneling landscapes, giving the local-search structure both strategies rely on.","marker":"[38]"},{"why":"demonstrates that non-uniform sampling can outperform uniform sampling, the baseline distribution choice for the non-conditional strategy.","marker":"[39]"},{"why":"supplies the astrodynamics trajectory-optimization software performing transcription, global search, and local solve for the case study.","marker":"[42]"},{"why":"supplies the gradient-based nonlinear-program solver that refines every initial guess in both strategies.","marker":"[43]"}],"fun_headline_variants":["Non-robust seeds win: faster, cheaper, more feasible Gateway transfer","Simpler seeds beat complex ones for robust low-thrust design","Conditional seeding improves robust trajectory search for Gateway","Non-robust seeds: best start for robust low-thrust design"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the discretized missed-thrust model — at most one outage per mission, starting at one of three fixed points on the transfer, and lasting 0.5, 1.0, or 1.5 days — faithfully represents the real missed-thrust risk; the paper itself notes these durations cover roughly 40% of observed outage durations, so outages elsewhere or longer than 1.5 days lie outside the tested design space.","fun_headline_variants_meta":{"raw":{"variants":["Non-robust seeds win: faster, cheaper, more feasible Gateway transfer","Simpler seeds beat complex ones for robust low-thrust design","Conditional seeding improves robust trajectory search for Gateway","Non-robust seeds: best start for robust low-thrust design"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001414,"raw_usage":{"total_tokens":5738,"prompt_tokens":1001,"completion_tokens":4737,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":617,"completion_tokens_details":{"reasoning_tokens":4664}},"tokens_in":617,"tokens_out":4737,"duration_ms":31941,"temperature":1.0,"reasoning_tokens":4664,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T20:55:04.928629+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the $P_1$, $P_2$, and $P_3$ comparisons with missed-thrust initiation points drawn from all 50 reference segments and durations sampled from the full historical Weibull distribution, including outages beyond 1.5 days, keeping cumulative seed-generation cost in the metrics; if $S(k|0)$ no longer dominates $S(k)$ in feasibility or cumulative solving time at $k \\ge 2$, the paper's central claim fails.","supporting_citations":[{"cited_title":"Modeling spacecraft safe mode events,","cited_arxiv_id":null,"evidence_quote":"supplies the historical missed-thrust frequency and duration data used to justify the single-event, finite-duration uncertainty model."},{"cited_title":"Missed Thrust Analysis and Design For Low Thrust Cislunar Transfers,","cited_arxiv_id":null,"evidence_quote":"introduces the reference-plus-realization lifted formulation and the three-initiation-point approximation adopted here."},{"cited_title":"Multi-Objective Low-Thrust Trajectory Optimization with Robustness to Missed Thrust Events,","cited_arxiv_id":null,"evidence_quote":"extends missed-thrust design to robust low-thrust problems and motivates the realization-based constraint structure."},{"cited_title":"Tuning monotonic basin hopping: improving the efficiency of stochastic search as applied to low-thrust trajectory optimization,","cited_arxiv_id":null,"evidence_quote":"demonstrates that non-uniform sampling can outperform uniform sampling, the baseline distribution choice for the non-conditional strategy."},{"cited_title":"DynamicallyLeveragedAutomated(N)MultibodyTrajectory Optimization (DyLAN),","cited_arxiv_id":null,"evidence_quote":"supplies the astrodynamics trajectory-optimization software performing transcription, global search, and local solve for the case study."}],"review_version":1}