{"id":"c0882ac1-52ae-491d-aedd-0958f529487f","arxiv_id":"2501.00930","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"T-SCvx predicts tight constraints and warm-start solutions with transformers, cutting 6-DoF powered descent guidance solve time by 66% on a Mars landing benchmark.","lead":"This paper trains transformer networks to predict which constraints will be active and to propose a warm-start trajectory for a 6-degree-of-freedom Mars powered descent guidance problem, cutting average solve time by 66 percent versus a standard successive convexification solver. A generalist reader should care because it shows a learning-based shortcut that could make fuel-optimal landing guidance fast enough for onboard computers.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reduced-problem solution is returned without a full-problem feasibility check, so the claimed feasibility guarantee is not supported by Algorithm 3 as written.","rationale":"The reader's weakest assumption is exactly the surrogate validity of the reduced problem defined by predicted tight constraints. My stress-test pass found no additional concern that supersedes it; in fact the pseudocode makes the issue more direct. Algorithm 3 returns the result of Reduced-Solve with no explicit verification against the full nonconvex problem. The prose claims that feasibility is ensured by keeping the full penalty cost, and Algorithm 2 does compute ΔJ with the full penalty, but the subproblem solved at each iteration is built only from the predicted tight constraints, and the predicted change ΔL is based on that reduced subproblem. There is therefore a mismatch between the model used to generate candidate steps and the full penalty used to accept or reject them. A classifier that is only 0.5% better than a constant-zero baseline cannot by itself support a hard feasibility guarantee; the missing piece is an explicit check that the returned solution satisfies every constraint of the original problem. This is fixable in practice by running the full SCvx problem initialized with the reduced solution and verifying the final residual, but the paper does not show that step in Algorithm 3, and the theoretical Claim in Section III.C does not supply the missing proof. Because the concern is about a missing verification and an incomplete proof rather than a demonstrated contradiction in the empirical results, the appropriate disposition remains conditional on the authors either releasing code that includes the full-problem check or adding an explicit verification step and residual reporting to the algorithm. This matches the reader's CONDITIONAL verdict, so no change to the verdict is needed.","tokens_in":19013,"tokens_out":4418,"duration_ms":45887,"concrete_test":"Take the 545 test samples used in Figure 2. For each, run Algorithm 3 exactly as written, returning only the Reduced-Solve output, and compute the maximum violation of the full constraints (Eqs. 3-6) and the dynamics residual at the returned trajectory. Then repeat the experiment with the tight set replaced by the oracle active set from full SCvx. If any returned trajectory violates the full feasibility tolerance (feas tol = 0.5) or differs materially in cost from the oracle solution, the feasibility claim fails. Additionally, rerun the constraint NN against the 'predict only zeros' baseline on the subset where binary prediction is wrong; if T-SCvx still reports 0% violations, the omitted-constraint concern is settled.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing step is the assertion that solving the reduced problem defined by the predicted tight set (Algorithm 3's Reduced-Solve) returns a feasible, locally optimal trajectory of the full 6-DoF problem. Algorithm 3 literally returns that reduced solve; no full SCvx continuation, constraint-residual check, or penalty-zero verification appears. Algorithm 2 computes ΔJ with the full penalty J, but the subproblem and the predicted change ΔL are built only from the predicted tight constraints. If the predictor omits a constraint that is active (or violated) at the candidate, the subproblem contains no term to repair it; the trust-region logic can only shrink or reject, and the stopping test ΔJ=0 can be met at an infeasible point if the full penalty is positive and no improving step is found. The Claim in Section III.C does not bridge this gap: the proof assumes a descent direction d with ∇f(x*)ᵀd<0 and then concludes that the active set solves the problem; it never proves that a point satisfying the active constraints satisfies the omitted inequalities, nor that the reduced problem's solution is feasible for the original nonconvex constraints. Table 3 makes this concrete: the constraint classifier is 96.45% accurate versus 95.95% for predict-only-zeros, so roughly 4% of tight-constraint entries are wrong on the test set, and nothing in Algorithm 3 checks whether those errors are benign. Hence the 66% mean runtime reduction is meaningful only if every returned reduced solve is later verified against the full problem constraints, which is not shown.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This paper proposes T-SCvx, a learning-based warm-start and problem-reduction method for the 6-DoF powered descent guidance problem. A transformer predicts the set of tight/active constraints at each SCvx iteration and also predicts a full state/control/final-time initial guess. The reduced subproblem defined by the predicted tight constraints is then solved, and the paper claims this preserves feasibility and local optimality while reducing mean solve time by 66% (from 14.61 s to 4.98 s) on 545 held-out test cases. A rotation-invariant data augmentation scheme is used to expand a base dataset of roughly 1,592 samples to 11,634 samples. The method is benchmarked against SCvx and against linear-interpolation and kd-tree lookup-table approaches.","tokens_in":19350,"tokens_out":7612,"duration_ms":69443,"significance":"Fast onboard 6-DoF powered descent guidance is an important and active problem, and the idea of learning active sets to reduce successive convexification subproblems is attractive. The rotation-augmentation contribution is concrete and potentially useful beyond this specific application, and the paper provides a substantial empirical evaluation against SCvx and lookup-table baselines. However, the central feasibility guarantee is not supported by the algorithms as written, and the data-splitting procedure may compromise the reported generalization results. These issues are significant but appear fixable within the scope of a revision.","major_comments":[{"comment":"The proof of the Claim is incomplete and internally unclear. It assumes a feasible sequence with a descent direction d satisfying d^T grad f(x*) < 0 and then concludes that the active set suffices, but it never shows that a solution to the reduced problem satisfies the omitted inequality constraints of the original nonconvex problem. The displayed inequality 'f(z_k) < f(x*) + (1/2)||z_k - x*|| d^T grad f(x*)' is also problematic: since d^T grad f(x*) < 0, the right-hand side is less than f(x*), so the inequality would contradict local optimality rather than establish it. As written, the proof does not support the load-bearing assertion that solving the reduced problem recovers the full optimal solution.","section":"Section III.C, Claim and Proof"},{"comment":"Algorithm 3 returns the output of Reduced-Solve directly and contains no full-problem feasibility check, no penalty-residual verification, and no warm-started full SCvx continuation. The statement in Section III.E that 'the full problem is used for the evaluation cost function' is not reflected in the algorithm, because the cost returned by Reduced-Solve is computed on the reduced subproblem. This is a load-bearing gap: Table 3 shows the constraint classifier achieves only 96.45% binary accuracy versus 95.95% for a predict-only-zeros baseline, so roughly 4% of tight-constraint predictions are wrong on the test set, and nothing in Algorithm 3 checks whether those errors are benign. Even in Algorithm 2, the stopping test Delta J = 0 does not by itself certify feasibility, since the exact penalty J can be positive at a stationary point of the penalty; the cited SCvx theory only guarantees a KKT point when the limit point is feasible. The claimed feasibility guarantee therefore requires either an explicit full-problem verification step in Algorithm 3, a warm-started full SCvx solve, or a proof that wrong predictions cannot cause constraint violation.","section":"Section III.E, Algorithm 3"},{"comment":"The data split is performed after rotation augmentation: the text states 'All sampled data, including the rotated samples, were split into 80% training and validation data and 20% test data.' This allows rotated copies of the same base sample to appear in both the training and test sets, because the rotation is applied before the split and no grouping by base sample is described. As a result, the reported 96.45% binary accuracy and the 66% solve-time reduction may reflect memorization of augmented variants rather than generalization to unseen initial conditions. The split should be performed on the 1,592 base samples before augmentation, or the authors should verify that no test sample shares a base sample with any training sample.","section":"Section III.E.3, Training, Validation, and Testing"},{"comment":"The reported constraint-prediction accuracy is only 0.5 percentage points above the predict-only-zeros baseline (96.45% versus 95.95%). Because tight-constraint labels are presumably sparse, binary accuracy is dominated by the zero class, and a predictor that never predicts a tight constraint can already achieve 95.95%. The paper's runtime and feasibility arguments depend on the quality of the positive (tight) predictions, so precision, recall, and F1 for the tight class should be reported, together with the effect of false negatives and false positives on the returned trajectories.","section":"Table 3, Constraint NN Accuracy"},{"comment":"Algorithm 2 modifies the SCvx update by scaling the trust region with the predicted constraint-change fraction tau_r and by solving subproblems built only from predicted tight constraints, but no convergence or optimality analysis is provided for this modified algorithm. The SCvx guarantees cited from [19] apply to the full convex subproblem; it is not immediate that they carry over when constraints are dropped at each iteration and the trust-region radius is scaled by a learned quantity. The paper should state explicitly which convergence and feasibility guarantees T-SCvx inherits, or should reframe the relevant claims as empirical rather than as consequences of the SCvx theory.","section":"Section III.D, Algorithm 2, Trust-Region Modification"}],"minor_comments":[{"comment":"The sentence 'the standard deviation for T-SCvx is slightly higher than SCvx, at 5.24 seconds' should specify which distribution the 5.24 s standard deviation refers to, and the following sentence 'one standard deviation remains below the SCvx mean of 14.61' is ambiguous because a standard deviation is not a bound on the mean.","section":"Section IV.B"},{"comment":"The conclusion states that the networks achieve 'test accuracies of over 96% and under 1 MSE,' but Table 3 reports a solution NN test MSE of 1.040, which is not under 1. Please correct the inconsistency.","section":"Section VI, Conclusion"},{"comment":"The text refers to 'Equation (21)' and 'Eq. (22)' for the trust-region update, but the relevant equations in the manuscript appear to be (14) and (15); the equation numbering in the surrounding text is inconsistent and should be harmonized.","section":"Section III.B.1"},{"comment":"The nomenclature lists 'omega = penalty coefficients,' but the penalty weights in the formulation are denoted lambda and tau; please correct this entry.","section":"Nomenclature"},{"comment":"The text says 'less than 1,600 samples' and later gives '1,592 samples'; please use one consistent number throughout.","section":"Section IV.A"},{"comment":"The abstract and conclusion state that T-SCvx 'enables onboard computation,' but Section V notes that verification on flight-grade radiation-hardened hardware remains future work. Please soften the wording to avoid overclaiming.","section":"Abstract and Section V"}],"recommendation":"major_revision","confidential_remarks":"The core idea is promising and the empirical speedup is potentially important, but the split-after-augmentation issue and the missing full-problem feasibility verification in Algorithm 3 are serious. I would be willing to reconsider after a revision that addresses these points. I would also encourage the authors to release the dataset, the base-sample grouping for the train/test split, and the code implementing Algorithm 3, as this would make the evaluation auditable."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"My take: this is a solid incremental paper with real experimental value, but the headline guarantee is load-bearing and it does not hold up. The authors extend their T-PDG idea to 6-DoF SCvx, predicting both tight constraints and a full warm-start solution, and add a trust-region scaling by constraint changes. The rotation-invariant augmentation is the cleanest part—under 1,600 base samples become 11,634, and that is a genuinely useful trick for non-convex trajectory datasets. The held-out evaluation is on 545 test cases, and the comparison to linear interpolation and KD-tree lookup is honest, including the fact that KD-tree is faster in-distribution while T-SCvx degrades less out-of-distribution. I also appreciate that the authors explicitly flag flight-hardware verification as future work rather than overselling the desktop ECOS numbers.\n\nThe soft spot is exactly what the stress-test note says: Algorithm 3 returns the reduced-problem solution without any full-problem feasibility check. The text claims that using the full penalty cost in the trust-region acceptance rule guarantees feasibility, but that is not correct. SCvx can converge to a stationary point of the penalty function that is infeasible for the original problem, and if the predictor omits a constraint that actually becomes active, the reduced subproblem has no mechanism to repair that violation. The proof in Section III.C is also incomplete: it assumes a descent direction exists and then asserts that the active set suffices, but it never shows that a point satisfying the active constraints satisfies the omitted inequalities. The constraint-prediction accuracy numbers reinforce the concern: 96.45% versus a predict-only-zeros baseline at 95.95%—the reported accuracy is dominated by the zero class, and we do not see recall or precision on the actual active constraints, which is what matters for feasibility. These are fixable: add an explicit full-problem residual check before returning, or prove that the reduced solution keeps the omitted constraints inactive. As written, the 66% runtime reduction is meaningful only if every returned reduced solve is subsequently verified, which the algorithm does not do.\n\nMinor point: the benchmark uses ECOS on a desktop, not a custom sub-second solver, so the runtime numbers are indicative rather than flight-representative, and the 20-iteration cap could bias the comparison. But that is a scope limitation, not a flaw.\n\nWho should read this: people working on learning-based warm-starts for nonconvex optimal control, and anyone doing powered descent guidance with SCvx. It is a useful data point and the augmentation scheme is worth stealing. With a verification step or a more modest feasibility claim, I would be comfortable with it. I would send it to peer review, expecting a major revision on the feasibility argument.\n\nFor the record: the work is clearly serious and the authors engage honestly with the literature; my skepticism is about a specific algorithmic gap, not the overall approach.","headline":"A genuinely useful empirical extension of T-PDG to 6-DoF SCvx with a clever rotation augmentation and honest benchmarks, but the feasibility guarantee is not supported by the algorithm as written.","tokens_in":19870,"tokens_out":2769,"would_cite":true,"duration_ms":28303,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["90C30"],"pacs":[],"model":"deepseek-v4-flash","headline":"Transformer-based Successive Convexification predicts which constraints bind at the optimum, then solves only those to cut 6-DoF Mars powered descent guidance solve time from 14.61 s to 4.98 s.","keywords":["T-SCvx","tight constraint prediction","transformer neural network","successive convexification","six-degree-of-freedom powered descent guidance","rotation-invariant data augmentation","Mars powered landing","real-time trajectory optimization"],"falsifier":"Run T-SCvx and plain SCvx on a batch of initial conditions deliberately drawn outside the training ranges; if any T-SCvx trajectory violates the glideslope, thrust-magnitude, gimbal, or attitude constraints, or has measurably higher fuel use, while the SCvx trajectory is feasible, the reduced-problem surrogate claim is falsified. A sharper check is to compare the network's predicted tight-constraint set with the true active set at the converged solution for those out-of-distribution cases and show the mismatch rate grows with Mahalanobis distance from the training data.","tokens_in":18838,"feed_emoji":"🚀","tokens_out":5920,"duration_ms":51320,"temperature":0.7,"pith_summary":"This paper claims that a transformer network can learn which constraints are active at the optimum of a 6-DoF powered descent guidance problem, and that solving only those constraints plus a learned warm start cuts solve time dramatically. On a 50-timestep free-final-time Mars landing case, the resulting algorithm, T-SCvx, reduces mean solve time from 14.61 s to 4.98 s, a 66% reduction, while keeping the full-problem penalty check that guarantees feasibility of returned trajectories. The paper also shows that rotating the training data about the vertical axis preserves the active-constraint pattern, so fewer than 2,000 raw samples suffice for over 96% binary accuracy in constraint prediction. The significance is that fast, reliable onboard trajectory generation for high-fidelity 6-DoF descent becomes plausible on flight-class hardware.","feed_headline":"T-SCvx cuts 6-DoF Mars descent guidance solve time by 66%","feed_subtitle":"Mean runtime falls from 14.6 to 5.0 seconds on a 50-node free-final-time powered landing problem.","key_machinery":"The load-bearing machinery is a transformer neural network that maps a 17-dimensional problem-parameter vector (initial position, velocity, attitude, angular velocity, mass, glideslope angle, pitch angle, and SCvx iteration number) to a binary vector marking which discretized inequality constraints are tight, plus a second transformer that predicts the full state, control, and final time. Rotation-invariant data augmentation around the Up axis inflates 1,592 raw samples to 11,634, exploiting the symmetry that the active-constraint pattern is unchanged under such rotations. The reduced subproblem is solved with only the predicted tight constraints, and its solution warm-starts the full problem so that feasibility is checked against the original constraints.","core_discovery":"The central claim is that the optimal solution of the full nonconvex problem is recovered by the reduced problem formed from equality constraints plus the predicted tight inequality constraints, so a neural network can replace most of the repeated convexification work. T-SCvx predicts both the tight-constraint set and a full state, control, and final-time warm start at each successive-convexification iteration, scales the trust region by the fraction of constraints that changed, and accepts a solution only when the full-problem penalty cost shows no improvement. In the 6-DoF Mars powered landing test with 50 timesteps and free final time, this reduces mean solve time from 14.61 s to 4.98 s, a 66% mean reduction and a 70% median reduction, with feasibility enforced by warm-starting the full problem.","pith_inferences":["If the tight-constraint surrogate holds in general, the active-set-plus-warm-start recipe could apply to other sequential convex programming problems beyond powered descent, since the reduction is formulation-agnostic.","The rotation-invariant augmentation suggests a broader principle: any symmetry of the dynamics that leaves the active set invariant can be encoded cheaply in the training data rather than in the network architecture, potentially lowering sample counts in other guidance problems.","The reported speedups are on a particular convex solver; flight hardware may show different gains, and the paper itself lists verification on radiation-hardened processors as future work.","A testable extension is to use the predicted tight-constraint set to sparsify the linear algebra inside the convex solver, which would compound the wall-clock savings by reducing per-iteration factorization cost."],"forward_implications":["T-SCvx reaches a mean 4.98 s solve time on the 50-timestep free-final-time 6-DoF Mars PDG benchmark, versus 14.61 s for SCvx, a 66% mean and 70% median reduction.","Training requires fewer than 2,000 raw samples: rotation augmentation gives 11,634 samples, and the constraint-prediction network reaches 96.45% binary accuracy on held-out test cases.","Because the full-problem penalty is retained in the convergence check, returned trajectories are asserted feasible and locally optimal under SCvx's convergence conditions.","T-SCvx beats linear-interpolation lookup tables by more than 99% in inference time and memory usage, and generalizes better to out-of-distribution inputs than k-d-tree lookup.","The same active-set prediction and warm-starting recipe can be carried into active-set-based nonlinear programming solvers, as the paper identifies as future work."],"supporting_citations":[{"why":"Establishes the T-PDG tight-constraint prediction method that T-SCvx extends from 3-DoF to 6-DoF.","marker":"[5]"},{"why":"Supplies the SCvx algorithm and its global and superlinear convergence guarantees that T-SCvx augments.","marker":"[19]"},{"why":"Provides the successive-convex-programming tutorial and solver framework used to implement SCvx and generate training samples.","marker":"[21]"},{"why":"Defines the 6-DoF powered descent guidance problem with free final time that is the paper's test case.","marker":"[2]"},{"why":"ECOS, the second-order cone solver used to solve the convex subproblems in SCvx and T-SCvx.","marker":"[53]"},{"why":"Supports the claim that rotation-based data augmentation achieves accuracy comparable to rotationally invariant network architectures.","marker":"[54]"},{"why":"Provides the flight computer's one-second runtime and 60 MB memory constraints used as the benchmarking target.","marker":"[61]"}],"fun_headline_variants":["T-SCvx cuts 6-DoF Mars guidance time by 66% – now real-time","AI predicts constraints: 66% faster Mars powered descent solve","T-SCvx: neural net tight-constraint prediction speeds Mars landing 66%","Real-time Mars guidance: T-SCvx slashes solve time by two-thirds","T-SCvx: from 14.6 to 5.0 seconds for 6-DoF Mars landing"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the predicted tight-constraint set is a faithful surrogate for the full nonconvex problem, so solving the reduced subproblem and warm-starting the full problem returns a feasible, locally optimal trajectory; if the prediction is wrong on out-of-distribution inputs, the returned trajectory can violate constraints or be suboptimal.","fun_headline_variants_meta":{"raw":{"variants":["T-SCvx cuts 6-DoF Mars guidance time by 66% – now real-time","AI predicts constraints: 66% faster Mars powered descent solve","T-SCvx: neural net tight-constraint prediction speeds Mars landing 66%","Real-time Mars guidance: T-SCvx slashes solve time by two-thirds","T-SCvx: from 14.6 to 5.0 seconds for 6-DoF Mars landing"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000779,"raw_usage":{"total_tokens":3486,"prompt_tokens":1029,"completion_tokens":2457,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":645,"completion_tokens_details":{"reasoning_tokens":2336}},"tokens_in":645,"tokens_out":2457,"duration_ms":15238,"temperature":1.0,"reasoning_tokens":2336,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T22:39:02.405792+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run T-SCvx and plain SCvx on a batch of initial conditions deliberately drawn outside the training ranges; if any T-SCvx trajectory violates the glideslope, thrust-magnitude, gimbal, or attitude constraints, or has measurably higher fuel use, while the SCvx trajectory is feasible, the reduced-problem surrogate claim is falsified. A sharper check is to compare the network's predicted tight-constraint set with the true active set at the converged solution for those out-of-distribution cases and show the mismatch rate grows with Mahalanobis distance from the training data.","supporting_citations":[{"cited_title":"RAD750™SpaceWire-Enabled Flight Computer for Lunar Reconnaissance Orbiter,","cited_arxiv_id":null,"evidence_quote":"Provides the flight computer's one-second runtime and 60 MB memory constraints used as the benchmarking target."},{"cited_title":"Improving Computational Efficiency for Powered Descent Guidance via Transformer-based Tight Constraint Prediction,","cited_arxiv_id":null,"evidence_quote":"Establishes the T-PDG tight-constraint prediction method that T-SCvx extends from 3-DoF to 6-DoF."},{"cited_title":"Successive Convexification of Non-Convex Optimal Control Problems and Its Convergence Properties,","cited_arxiv_id":null,"evidence_quote":"Supplies the SCvx algorithm and its global and superlinear convergence guarantees that T-SCvx augments."},{"cited_title":"Convex Optimization for Trajectory Generation: A Tutorial on Generating Dynamically Feasible Trajectories Reliably and Efficiently,","cited_arxiv_id":null,"evidence_quote":"Provides the successive-convex-programming tutorial and solver framework used to implement SCvx and generate training samples."},{"cited_title":"Successive Convexification for Real-Time Six-Degree-of-Freedom Powered Descent Guidance with State-Triggered Constraints,","cited_arxiv_id":null,"evidence_quote":"Defines the 6-DoF powered descent guidance problem with free final time that is the paper's test case."},{"cited_title":"ECOS: An SOCP solver for embedded systems,","cited_arxiv_id":null,"evidence_quote":"ECOS, the second-order cone solver used to solve the convex subproblems in SCvx and T-SCvx."},{"cited_title":"Revisiting Data Augmentation for Rotational Invariance in Convolutional Neural Networks,","cited_arxiv_id":null,"evidence_quote":"Supports the claim that rotation-based data augmentation achieves accuracy comparable to rotationally invariant network architectures."}],"review_version":1}