{"id":"3c059b83-69d1-438f-80e4-331db22c7e98","arxiv_id":"2502.01907","paper_version":2,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":2,"one_line_summary":"This paper presents a sequential convex programming method that jointly optimizes a cislunar trajectory and an affine feedback correction policy while probabilistically bounding state and control uncertainty.","lead":"This paper develops an optimization algorithm that designs a spacecraft trajectory and its correction maneuvers together, treating navigation and maneuver errors as random and enforcing probabilistic safety constraints. It demonstrates the method on Earth-Moon transfers, showing that simulated spacecraft stay near the intended path and that the computation is faster than prior robust-design approaches.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The method's probabilistic safety claims depend on linear covariance propagation Eq. (38c) being faithful in the strongly nonlinear flyby regime; the paper's own Remark 3 concedes this is inexact, and 200-sample Monte Carlo cannot certify the 1% tail.","rationale":"The reader's weakest-assumption analysis identifies the same load-bearing concern: the linearized covariance propagation, Eq. (38c), must accurately approximate the true distribution for the chance constraints, covariance constraints, and Delta V99 estimates to retain their probabilistic meaning. This is the most direct threat to the central claim because the safety guarantee is expressed as a probability under the assumed error models, and if the propagation model is wrong, the probability is computed against the wrong distribution. The paper deserves credit for explicitly acknowledging the inexact linearization in Remark 3 and for providing Monte Carlo evidence, local Lyapunov exponent analysis, and two nontrivial cislunar examples. These give partial support to the claim that the method works in practice. However, 200 Monte Carlo samples cannot meaningfully validate 1% tail probabilities, and the acknowledgment in Section V.B that without the covariance bound the samples escape the region of validity shows that the method's success is tied to the linearization assumption holding. A stress test with larger sample counts and increased uncertainty would settle whether the linear covariance model is faithful in the active-constraint regime near the flyby. Since the concern is real but the paper already frames the method as contingent on this assumption and gives encouraging numerical evidence, the appropriate verdict remains conditional acceptance rather than rejection.","tokens_in":26179,"tokens_out":17910,"duration_ms":204736,"concrete_test":"Re-run the NRHO-Halo Monte Carlo validation with 10^4 samples, and also with the unmodeled-acceleration intensity sigma_a increased by 3x and the initial covariance scaled by 3x. At every node, compute the empirical covariance matrix and the empirical 99th-percentile Mahalanobis radius of the samples, and compare them with the linear prediction P_k = Phat_k + Ptilde_k. Treat Eq. (38c) as validated only if the largest eigenvalue of the empirical covariance is within roughly 20% of the predicted value and the empirical exceedance rate of the predicted 99% ellipsoid is within Monte Carlo error (about 0.7%-1.3% for 10^4 samples). If either condition fails near the lunar flyby, the linearized covariance model is not faithful and the probabilistic guarantees are not substantiated.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the SCP-based method probabilistically guarantees safety constraints and accurately controls state uncertainty in the cislunar regime. The load-bearing assumption is that Eq. (38c), the linearized filtered covariance propagation, faithfully represents the true stochastic closed-loop covariance. The authors explicitly concede in Remark 3 that Eq. (38c) is an inexact linearization because A_k, B_k, c_k, and Qhat_k are frozen at the reference mean and the effect of changes in the mean trajectory on these matrices is ignored. In Section V.B they further state that without the maximum covariance constraint, \"the state dispersion can grow outside the region where linear covariance propagation is a valid approximation.\" The maximum covariance constraint Eq. (15) only bounds the predicted covariance; it does not certify that the true nonlinear covariance remains small or that first-order sensitivities dominate. Near the lunar flyby, where the local Lyapunov exponent is large (Figs. 11-12), higher-order terms in the state transition tensor can make the linear covariance, and hence the chance constraints (38d) and the Delta V99 bound, arbitrarily inaccurate. The Monte Carlo validation uses only 200 samples, which is far too few to validate a 1% violation probability. Thus, the probabilistic guarantees are conditional on an unverified small-uncertainty and weak-nonlinearity assumption in exactly the regime the paper targets.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper formulates a chance-constrained stochastic optimal control problem for cislunar low-thrust trajectory design, combining linearized covariance propagation with an affine state-estimate feedback policy (Eq. 20), and solves it via a sequential convex programming (SCvx*) scheme. The main methodological contributions are a full-covariance formulation with a trace-regularized quantile objective (Eq. 45), a lossless-convexification LMI (Eq. 51d) adapted to the quantile cost, a difference-of-convex relaxation of the control chance constraint (Eqs. 47–49), and a scalar scaling strategy (Eq. 58). Numerical demonstrations on DRO-to-DRO and NRHO-to-Halo transfers in the CR3BP include Monte Carlo verification of the covariance prediction, the Delta V99 bound, and the final covariance constraint, as well as a local Lyapunov exponent comparison of robust versus deterministic trajectories.","tokens_in":26450,"tokens_out":3299,"duration_ms":27222,"significance":"If the claims hold, the paper provides a practically relevant contribution: a computationally efficient SCP-based framework that simultaneously optimizes a nominal trajectory and an affine TCM policy with explicit probabilistic safety constraints in a strongly nonlinear, locally chaotic regime. The paper is careful in several respects: it explicitly acknowledges the inexact linearization of the covariance propagation (Remark 3) and the loss of the SCvx* convergence guarantee, it validates the Delta V99 predictions against independent Monte Carlo simulations rather than fitting to them, and it makes concrete benchmark claims about runtime versus block-Cholesky formulations. The comparison table of stochastic optimal control approaches and the LLE analysis add useful context. The main caveat is that the probabilistic guarantees are conditional on the validity of linear covariance propagation, a point the authors themselves flag.","major_comments":[{"comment":"The central probabilistic safety claim rests on the fidelity of the linearized covariance propagation Eq. (38c) in the presence of a lunar flyby, where the local Lyapunov exponent is large (Figs. 11–12). The authors correctly concede in Remark 3 that Eq. (38c) is an inexact linearization, and in Section V.B that without the maximum covariance constraint \"the state dispersion can grow outside the region where linear covariance propagation is a valid approximation.\" The Monte Carlo validation uses only 200 samples, which is insufficient to certify the 1% violation probability asserted by the chance constraints (Eq. 14a) or the 99% quantile of the cost (Eq. 18). The paper should therefore either soften the language of \"probabilistically guarantee\" to \"probabilistically guarantee under the linearized model,\" or provide additional evidence, such as a larger Monte Carlo campaign or a second-order (state transition tensor) check, that the linear covariance prediction remains accurate in the flyby regime.","section":"Section V.B and Remark 3; Eq. (38c)"},{"comment":"The acceptance criterion (Eq. 56) accepts a step whenever rho is in [1-eta_0, 1+eta_0], but the paper provides no analysis of whether this criterion preserves the descent or convergence properties of the original SCvx* algorithm. The authors explicitly state that Algorithm 1 \"does not fully inherit the convergence guarantee provided by the SCvx* algorithm\" due to inexact linearization. Since the modified acceptance criterion is a key algorithmic departure, and the numerical results rely on it, the paper should either provide an analysis under a bounded-inexactness assumption, or clearly state that the convergence property is empirical. As written, the numerical examples support the effectiveness claim, but the convergence claim is weaker than implied.","section":"Section IV.B, Eq. (56)"},{"comment":"The lossless convexification guarantee for the quantile-2-norm objective relies on the strict positive definiteness of the gradient of the objective with respect to Y_k, obtained by adding the trace regularization term epsilon_Y tr(Y_k). The authors acknowledge that the trace term modifies the objective, but the manuscript does not provide a quantitative check of how the optimal Delta V99 and the recovered policy change with epsilon_Y, despite the text noting that \"Theoretically, epsilon_Y can be an arbitrarily small positive number, but in practice, we should not use a value that is too small.\" A sensitivity study over epsilon_Y (at least a few values spanning the stated range 1e-4 to 1e-6) would make the claim of negligible impact concrete. As written, the claim rests on a single value.","section":"Section IV.A.2 and Eq. (45)"}],"minor_comments":[{"comment":"The manuscript contains several typographical errors in the equation display of Section IV.C (e.g., \"ill-scaled reformulation\" should read \"well-scaled reformulation\" given the surrounding discussion, and \"maxmimum covariance constraint\" in Section V.B). These should be corrected in a final revision.","section":"Throughout"},{"comment":"The Monte Carlo verification uses 200 samples for both transfers. While this is adequate to illustrate the qualitative behavior, the paper should state the standard error on the empirical Delta V99 and the empirical violation rate for the 1% chance constraint, or at least report the maximum observed violation rate. This would allow a reader to gauge the statistical precision of the claimed upper bound.","section":"Section V.A"},{"comment":"The Gates model is described with a sign error in Eq. (7a): the tilde on the left-hand side is undefined if the right-hand side is the additive error. The notation should be clarified, e.g., by defining u_k = u_bar_k + tilde u_k.","section":"Section II.B.3"},{"comment":"The discretization uses the reference control u*_k in the integration of Eq. (24a), but the linearized dynamics in Eq. (23) use c_k defined via Eq. (25). The consistency of these definitions with the affine term c_k should be stated explicitly, as the reader may otherwise confuse the reference control used for trajectory propagation with the variable u_k in the subproblem.","section":"Section III.B, Eq. (24)"},{"comment":"The finite-time local Lyapunov exponent is defined as Lambda = (1/Delta t) ln ||Phi(t+Delta t, t)||_2, but the text describes it as a local Lyapunov exponent along the trajectory. It would help to state that this is a finite-time quantity and that its interpretation depends on the choice of Delta t, which the authors do mention. A reference for this definition would also be useful.","section":"Section V.C, Eq. (60)"}],"recommendation":"major_revision","confidential_remarks":"The paper is a solid contribution to an active area, and the authors are unusually transparent about the limitations of the linearization-based approach. My main concern is the gap between the claimed probabilistic guarantee and what is actually delivered: the chance constraints and Delta V99 bound are computed under a linearized covariance model whose fidelity in the flyby regime is only checked with 200 Monte Carlo samples. This is a load-bearing issue for the central claim, but it is fixable within the scope of the paper by reframing the claims as conditional on the linearization and by adding a more quantitative validation. I would not reject on the basis of the convergence issue, since the authors explicitly disclose it and the empirical evidence is supportive. A major revision seems appropriate."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: this is a competent and honest extension of SCP-based covariance steering to cislunar transfers, and the main weakness is the one the authors themselves flag: the probabilistic guarantees rest on linear covariance propagation near a lunar flyby, where nonlinearity is strongest. I think it deserves review.\n\nWhat's new: previous SCP covariance work in cislunar was mostly stationkeeping. Here they do full transfers (DRO–DRO, NRHO–halo) with a full-covariance SDP formulation, plus a trace-regularization tweak to make the quantile objective convexifiable, a scalar scaling scheme that reportedly avoids numerical issues, and an LLE comparison showing the robust trajectory is locally more stable. The computational speed-up over block-Cholesky formulations is plausible and consistent with the linear covariance steering literature. Credit where due: the paper is unusually candid. Remark 3 admits the linearized covariance dynamics are inexact linearization and that the SCvx* convergence guarantee no longer applies, and they propose a heuristic acceptance criterion rather than pretending otherwise. The Monte Carlo validation, 200 samples, matches the predicted covariance for the closed-loop cases and shows the open-loop contrast, which is real evidence the machinery is doing something.\n\nSoft spots, in proportion. The stress-test note is right that 200 samples cannot certify a 1% violation probability, and the linear covariance model is least trustworthy exactly where the max covariance constraint is active. But the paper already says this; it does not oversell. The bigger issue is that the chance constraints and DeltaV99 are only as good as Eq. (38c), and no second-order tensor or nonlinear UQ check is used to bound the error. The trace regularization also means the solved objective is not exactly the DeltaV99 quantile, though they show the bound is tight in examples. That is a real but moderate gap, not a load-bearing flaw, because the paper's claims are framed as engineering demonstrations.\n\nWho this is for: astrodynamicists and spacecraft GNC people, less so pure control theorists. It is a useful benchmark for robust cislunar design. Deserves a serious referee; I would recommend majors rather than reject. Ask for a statement on the MC sample size and tail validation, and ideally release code, but the core result is reproducible in principle.","headline":"Honest, practical extension of covariance steering to cislunar transfers; the acknowledged linearization gap is the main thing to watch.","tokens_in":26968,"tokens_out":1519,"would_cite":true,"duration_ms":17116,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["49M37","90C26","93E20","70F15"],"pacs":[],"model":"deepseek-v4-flash","headline":"Covariance steering makes cislunar transfers robust in minutes.","keywords":["cislunar trajectory optimization","chance-constrained covariance steering","sequential convex programming","trajectory correction maneuvers","Circular Restricted Three-Body Problem","uncertainty quantification","lossless convexification","low-thrust propulsion"],"falsifier":"Run a Monte Carlo simulation for the NRHO-to-Halo transfer with the maximum covariance constraint removed, as the paper reports that the SCP can converge while the samples escape the cislunar region; alternatively, increase the initial position uncertainty from 50 km to several hundred kilometers and check whether the predicted final 3-sigma ellipsoid still contains at least 99 percent of the Monte Carlo samples.","tokens_in":25919,"feed_emoji":"🌙","tokens_out":3904,"duration_ms":41982,"temperature":0.7,"pith_summary":"The paper claims that robust cislunar trajectory design can be solved as a single optimization that simultaneously outputs a nominal trajectory and a feedback-based trajectory correction policy, with explicit probabilistic guarantees on safety. It propagates state uncertainty through linearized covariance dynamics and uses sequential convex programming with a lossless convexification step, making each subproblem a semidefinite program. Demonstrated on Earth–Moon transfers, the method keeps Monte Carlo samples within the predicted bounds and is orders of magnitude faster than prior block-Cholesky approaches. The paper also shows that the robust trajectories have better local stability, measured by lower local Lyapunov exponents at key trajectory peaks.","feed_headline":"Covariance steering makes cislunar transfers robust in minutes","feed_subtitle":"An SCP-based method designs a trajectory plus correction policy with explicit 99 percent safety bounds, verified by Monte Carlo.","key_machinery":"The central object is the belief-state dynamics: the mean state is propagated through the nonlinear CR3BP dynamics, while the state covariance is propagated through linearized Kalman-filter equations, including orbit determination updates and maneuver execution errors. The lossless convexification step replaces the bilinear product $K_k \\hat{P}_k$ with new decision variables $U_k$ and $Y_k$, together with an LMI constraint, making the covariance steering subproblem convex; a small trace penalty $\\epsilon_Y \\operatorname{tr}(Y_k)$ guarantees the strict monotonicity condition required for losslessness. The chance constraint on the control 2-norm is handled by a chi-squared quantile bound and a difference-of-convex linearization with slack variables, all embedded in the SCvx* sequential convex programming framework.","core_discovery":"The central claim is that chance-constrained covariance steering, formulated as a sequential convex program, can simultaneously design an optimal nominal trajectory and a trajectory correction policy that probabilistically guarantees safety constraints in the highly nonlinear cislunar environment. The discovery is that a full-covariance formulation, with an exact lossless convexification of the covariance propagation, yields a convex subproblem whose complexity scales linearly with the number of discretization nodes rather than quadratically. This enables solving a 200-node NRHO-to-Halo transfer in about ten minutes, with the resulting closed-loop policy keeping Monte Carlo sample dispersions within the predicted 3-sigma ellipsoids and satisfying the control chance constraint.","pith_inferences":["The method's reliance on linearized covariance propagation implies that for large initial uncertainties or long flight times, the probabilistic guarantees could degrade; a hybrid approach that switches to Gaussian mixture or unscented transform uncertainty propagation when the covariance grows beyond a threshold would be a natural extension the paper leaves implicit.","The authors openly note that the inexact linearization breaks the convergence guarantee of the base SCvx* algorithm; rigorously analyzing convergence under this inexactness, or comparing against a second-order sensitivity (state transition tensor) formulation, would clarify when the method can be trusted in safety-critical scenarios.","The $\\epsilon_Y \\operatorname{tr}(Y_k)$ regularization term, while needed for losslessness, changes the objective from pure $\\Delta V_{99}$; a testable question is how the choice of $\\epsilon_Y$ affects the gap between the predicted and actual $\\Delta V_{99}$ across a range of uncertainty levels."],"forward_implications":["If the central claim is correct, robust cislunar transfers can be designed with explicit probabilistic margins (e.g., $\\Delta V_{99}$) without iterating between trajectory design and uncertainty analysis.","The linear computational scaling in the number of nodes allows rapid comparison of $\\Delta V_{99}$ across different uncertainty models, making it practical to conduct uncertainty-parameter trade studies.","The framework naturally extends to other constraints, such as keep-out zones and state chance constraints, as the authors note, enabling future work on collision-safe cislunar operations.","The local Lyapunov exponent analysis suggests that uncertainty-aware optimization automatically reshapes the nominal trajectory to improve local stability, independent of the chosen trajectory correction policy."],"supporting_citations":[{"why":"Provides the foundational SCP-based chance-constrained covariance control formulation for low-thrust trajectories, including the block-Cholesky approach that this paper compares against and the proof that the total covariance is the sum of estimate and error covariances.","marker":"[14]"},{"why":"Proposes the full-covariance convex formulation that this paper builds on; its lossless convexification proof is later shown incorrect, motivating the new convexification presented here.","marker":"[15]"},{"why":"Earlier stochastic sequential convex programming work that models orbit determination and mass uncertainty; supplies the block-Cholesky implementation whose computation times are used as a baseline.","marker":"[16]"},{"why":"Provides the deterministic upper bound on the p-quantile of the norm of a Gaussian random vector, which is used for the objective and the control chance constraint.","marker":"[17]"},{"why":"A critical comment showing that the lossless convexification in [15] is not actually lossless, directly motivating the new lossless formulation in this paper.","marker":"[18]"},{"why":"Establishes the lossless convexification result for covariance steering of discrete-time linear stochastic systems, which is the theoretical basis for the convex subproblem used here.","marker":"[19]"},{"why":"Demonstrates the computational efficiency of the full-covariance formulation for linear systems, which this paper extends to nonlinear cislunar transfers.","marker":"[20]"},{"why":"Supplies the Gates model for maneuver execution errors, the control-dependent noise model used in the uncertainty formulation.","marker":"[35]"},{"why":"Supplies the Kalman filter covariance propagation equations and the discretization method used to derive the linearized filtered covariance dynamics.","marker":"[37]"},{"why":"Provides the SCvx* sequential convex programming algorithm with augmented Lagrangian and trust-region updates, which the paper adapts for the covariance steering problem.","marker":"[43]"}],"fun_headline_variants":["Cislunar trajectory design that tames uncertainty in minutes","Chance-constrained steering for robust cislunar missions","Fast covariance steering for safe cislunar transfers","Linear-time convex steering beats heuristics for cislunar"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The method's statistical guarantees rely on the linearized covariance propagation accurately describing the true distribution of the nonlinear stochastic system, meaning the state uncertainty must stay small enough that a first-order Taylor expansion around the mean trajectory is valid.","fun_headline_variants_meta":{"raw":{"variants":["Cislunar trajectory design that tames uncertainty in minutes","Chance-constrained steering for robust cislunar missions","Fast covariance steering for safe cislunar transfers","Linear-time convex steering beats heuristics for cislunar"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000233,"raw_usage":{"total_tokens":1440,"prompt_tokens":842,"completion_tokens":598,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":458,"completion_tokens_details":{"reasoning_tokens":545}},"tokens_in":458,"tokens_out":598,"duration_ms":7105,"temperature":1.0,"reasoning_tokens":545,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-09T14:01:16.403875+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run a Monte Carlo simulation for the NRHO-to-Halo transfer with the maximum covariance constraint removed, as the paper reports that the SCP can converge while the samples escape the cislunar region; alternatively, increase the initial position uncertainty from 50 km to several hundred kilometers and check whether the predicted final 3-sigma ellipsoid still contains at least 99 percent of the Monte Carlo samples.","supporting_citations":[{"cited_title":"no linearization & discretization; leverages convex solver solves multiple small LMIs (seconds to minutes)","cited_arxiv_id":null,"evidence_quote":"Proposes the full-covariance convex formulation that this paper builds on; its lossless convexification proof is later shown incorrect, motivating the new convexification presented here."},{"cited_title":"A Sweeping Gradient Method for Ordinary Differential Equations with Events,","cited_arxiv_id":null,"evidence_quote":"Earlier stochastic sequential convex programming work that models orbit determination and mass uncertainty; supplies the block-Cholesky implementation whose computation times are used as a baseline."},{"cited_title":"Nonlinear Programming Approach to Trajectory Optimization under Uncertainty: Direct Forward-Backward Shooting Formulation,","cited_arxiv_id":null,"evidence_quote":"Provides the deterministic upper bound on the p-quantile of the norm of a Gaussian random vector, which is used for the objective and the control chance constraint."},{"cited_title":"Robust Low-Thrust Trajectory Correction Planning Under Uncertainty: Primer Vector Theory Approach,","cited_arxiv_id":null,"evidence_quote":"Establishes the lossless convexification result for covariance steering of discrete-time linear stochastic systems, which is the theoretical basis for the convex subproblem used here."},{"cited_title":"Chance-Constrained Covariance Control for Low-Thrust Minimum-Fuel Trajectory Optimization,","cited_arxiv_id":null,"evidence_quote":"Demonstrates the computational efficiency of the full-covariance formulation for linear systems, which this paper extends to nonlinear cislunar transfers."},{"cited_title":"Minimum-Fuel Closed-Loop Powered Descent Guidance with Stochastically Derived Throttle Margins,","cited_arxiv_id":null,"evidence_quote":"Supplies the Gates model for maneuver execution errors, the control-dependent noise model used in the uncertainty formulation."},{"cited_title":"Chance-Constrained Gaussian Mixture Steering to a Terminal Gaussian Distribution,","cited_arxiv_id":null,"evidence_quote":"Supplies the Kalman filter covariance propagation equations and the discretization method used to derive the linearized filtered covariance dynamics."},{"cited_title":"D., Schutz, B","cited_arxiv_id":null,"evidence_quote":"Provides the SCvx* sequential convex programming algorithm with augmented Lagrangian and trust-region updates, which the paper adapts for the covariance steering problem."}],"review_version":1}