{"id":"7d2275c7-41f6-4296-b98f-dfbc459931ff","arxiv_id":"2501.14526","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"A robustified two-stage time-optimal controller co-optimizes feedback gains and state covariances in a fixed-grid first stage and minimizes travel time in a variable-grid second stage, with online replanning.","lead":"This paper presents a two-stage optimal control method that plans a fast, safe path for robots under noise by optimizing the path, feedback corrections, and uncertainty growth together, then re-plans online. It matters because time-optimal robot motions usually sit at constraint boundaries, where unmodeled noise can cause collisions or actuator saturation.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Stage-2 robustification freezes stage-1 covariance and never propagates uncertainty through the variable-time stage, so the claimed safety margins in (6j)-(6l) are not tied to the actual closed-loop covariance.","rationale":"The reader's weakest_assumption identifies exactly the same issue: stage-2 constraints are robustified by freezing stage-1 terminal covariance and feedback gain, with no covariance recursion through the variable-time stage 2. I reviewed the full formulation and the numerical example. The concern is real and load-bearing because it attacks the core claim of robustified constraints: if the covariance is not propagated, the ellipsoidal safety margins in (6j) and (6l) do not correspond to the actual closed-loop uncertainty, and the probability guarantee in (4) is not established. The inconsistency between using Σ1,N1−1 for intermediate constraints and Σ1,N1 for the terminal constraint further weakens the formulation. The paper does not offer a counterargument or a bound on the growth of stage-2 uncertainty; it simply freezes the stage-1 values. This is an omitted proof/step rather than a mere lack of tuning. However, the rest of the paper—two-stage formulation, Riccati-based decomposition, ASAP-MPC embedding—remains potentially valuable, and the issue is addressable by adding covariance propagation through stage 2. Therefore the reader's CONDITIONAL verdict remains appropriate; my stress-test does not move it. I agree with the reader's prioritization of this concern over the missing convergence proof and the KKT-tolerance comparison, since the safety guarantee is the central claimed novelty.","tokens_in":11518,"tokens_out":5752,"duration_ms":53861,"concrete_test":"Propagate the stage-2 covariance analytically using the linearized dynamics of the scaled system (6h) with the constant feedback gain K1,N1−1, starting from Σ1,N1, over the N2 variable-time steps. For each step, compute the predicted covariance Σ2,n and then the corresponding β2,n and β2,N2 via (5). Compare these with the frozen values used in (6j) and (6l). If the propagated terminal covariance exceeds the frozen one by more than 10%, the safety margins are underestimated. Additionally, run Monte Carlo closed-loop simulations (e.g., 10^4 samples) of the full unicycle model with Gaussian noise, executing the planned stage-2 control law with gain K1,N1−1, and estimate the empirical stage-2 constraint-violation probability; if it exceeds the intended p=0.00135 (for σ=3), the robustification claim is refuted.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that OCP (6) yields a robustified time-optimal motion with stage and terminal chance constraints. The weakest link is the handling of stage-2 uncertainty. In (6j), the intermediate stage-2 constraint variances β2,n are computed from (K1,N1−1, Σ1,N1−1), and in (6l) the terminal variance β2,N2 is computed from Σ1,N1. No Riccati recursion or covariance propagation is performed along the stage-2 variable-time grid, even though the closed-loop control during stage 2 is implicitly governed by the frozen gain K1,N1−1. With N2=3 and a stage-2 duration of roughly 4.6 s in the numerical example, each stage-2 step covers about 1.5 s. Linearized process noise with covariance Σw can grow substantially over such an interval, and the terminal covariance Σ1,N1 gives no information about the covariance at later points. Using Σ1,N1−1 for intermediate constraints further introduces an inconsistency: it is the covariance before the last stage-1 control, not after the stitch. Thus the constraints labeled robust are not calibrated to the actual uncertainty distribution. If the true stage-2 covariance grows, the probability of constraint violation exceeds the intended σ-level, so the proposed method does not deliver the advertised robustness guarantee. This concern is structurally distinct from algorithmic convergence or implementation aspects and directly affects the validity of the central contribution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a robustified two-stage optimal control problem (OCP) for time-optimal point-to-point motion planning and control under Gaussian process noise. Stage 1 uses a fixed time grid and optimizes the nominal trajectory together with feedback gains and state covariances; stage 2 uses a variable time grid to minimize the remaining motion time. The stage-1 covariance and gain information is used to robustify stage and terminal constraints in both stages. The authors embed this OCP in an asynchronous MPC scheme with timely replanning and present an alternating solver that iterates between a Riccati recursion for the feedback gains and a nominal time-optimal OCP with frozen safety margins. A numerical example with a unicycle robot moving around an ellipsoidal obstacle reports a 5.22 s motion time under timely replanning versus 5.2 s for single planning, with each replanning solved within the 0.6 s stage-1 horizon.","tokens_in":11874,"tokens_out":4271,"duration_ms":38736,"significance":"If the safety argument were fully valid, the paper would offer a practically attractive framework for combining time optimality with constraint robustification at a computational cost suitable for online replanning. The decoupling into a Riccati subproblem and a nominal OCP is elegant, and the numerical example demonstrates plausible computation times. However, the stage-2 robustification as formulated does not propagate uncertainty through the variable-time stage, so the advertised chance-constraint guarantees are not established. Since this issue directly affects the central robustness claim, the paper in its current form cannot be accepted without substantial revision.","major_comments":[{"comment":"The stage-2 constraint variances are computed from stage-1 terminal data only: β2,n = H(¯s2,n, ¯u2,n, K1,N1−1, Σ1,N1−1) in (6j) and β2,N2 = Htf(¯s2,N2, Σ1,N1) in (6l). No covariance recursion is performed along the variable-time stage-2 grid, even though each stage-2 step covers roughly 1.5 s in the numerical example (N2 = 3, T2 ≈ 4.6 s). Over such an interval, process noise with the reported covariance can accumulate substantially, and Σ1,N1 does not by itself provide information about the covariance at later times. Moreover, using Σ1,N1−1 for the intermediate stage-2 constraints is not the covariance at the stitch point after the last stage-1 control. Consequently, the constraints in (6i) and (6k) are not calibrated to the actual closed-loop distribution, and the claimed σ-level chance constraints are not guaranteed. The authors should either propagate covariance through stage 2 (including the time-scaling map) or provide a rigorous argument, such as a dominant-tube bound, showing that the frozen stage-1 covariance is conservative for the entire stage-2 tail.","section":"Sec. II-C, Eqs. (6j)-(6l)"},{"comment":"The alternating tailored solver is presented without a convergence proof. The stopping criterion \"KKT conditions (10) meet threshold\" assumes that the iterates approach a KKT point, but no theorem is given to show that the alternation between subproblem (14) and subproblem (16) converges, nor that a fixed point of the iteration satisfies the KKT conditions of the original problem (9). The numerical example shows convergence for one instance and one tolerance, which is not sufficient to support the paper's claim of a tailored algorithm for reliable real-time OCP execution. A local convergence analysis, or at least an explicit statement that convergence is heuristic and must be verified online, is needed.","section":"Sec. III, Algorithm 2"},{"comment":"The objective (6a) minimizes T2 together with weighted covariance traces, so the resulting \"time-optimal\" trajectory is only optimal for a particular choice of the regularization weights Rregu and Rregu_tf. These weights are free parameters, and they are chosen differently for single planning (Rregu = diag(80, 80, 80, 500, 500), Rregu_tf = 1000 I3) and for timely replanning (Rregu = I5, Rregu_tf = 50 I3) in the numerical section. Therefore the reported comparison of 5.2 s versus 5.22 s does not isolate the effect of the replanning strategy; it also reflects different objective tunings. The authors should either justify a canonical weight selection or present a Pareto-style discussion of the trade-off between motion time and state/control uncertainty.","section":"Sec. II-C, Eq. (6a) and Sec. IV"}],"minor_comments":[{"comment":"Equation (6h) reads ¯s2,n+1 = fT(¯s2,n, ¯s2,n, T2/N2), where the second argument appears to be a typo; based on the surrounding text and the discretized dynamics, it should depend on the control ¯u2,n, e.g., fT(¯s2,n, ¯u2,n, T2/N2).","section":"Sec. II-C, Eq. (6h)"},{"comment":"The KKT stationarity condition for β is printed as µiσ 2√βi + ϵ − ηi = 0, which appears to have missing parentheses and a division; it should presumably be µi σ/(2 sqrt(βi + ϵ)) − ηi = 0.","section":"Sec. III, Eq. (10)"},{"comment":"The caption states that blue ellipses and red ellipses denote closed-loop and open-loop uncertainty, but the figure and text do not explain how the ellipse sizes relate to the covariances or to the probability level σ; adding a short explanation would improve interpretability.","section":"Fig. 1 caption"}],"recommendation":"major_revision","confidential_remarks":"The paper builds heavily on the authors' previous work [8], [17], and [18]; the relationship is acknowledged, but the novelty relative to [17] in particular could be clarified for the editor. The main technical risk is the stage-2 covariance issue, which is likely fixable by adding a stage-2 covariance propagation or a conservative bound, but as it stands it prevents the paper from delivering its advertised safety guarantee."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a plausible engineering integration with a real hole in the safety argument. The new part is the two-stage OCP: stage 1 on a fixed grid optimizes nominal trajectory, feedback gains, and covariances; stage 2 on a variable grid minimizes time; and the two are stitched by feeding stage-1 terminal covariance into stage-2 robust constraints. That co-optimization is sensible and, as far as I can tell from the cited literature, genuinely new in this combination. The numerical unicycle result is plausible: timely replanning gives 5.22 s vs 5.2 s for single planning, with each replan solved within the 0.6 s stage-1 horizon.\n\nWhat the paper does well: the formulation is clean, the Riccati subproblem is grounded in Messerer and Diehl's work, and the gradient-correction term in Algorithm 2 is a reasonable way to approximate KKT conditions without solving the full robust problem. The idea of using stage-1 terminal uncertainty to set stage-2 margins is cheap and practical.\n\nThe main soft spot is exactly what the stress-test flags. Constraints (6j) and (6l) freeze K1,N1-1 and Sigma_1,N1-1 / Sigma_1,N1; there is no covariance recursion through the variable-time stage 2. With N2=3 and a stage-2 duration near 4.6 s, each stage-2 step is about 1.5 s, long enough for linearized noise to grow substantially. The terminal covariance of stage 1 does not bound the covariance at later stage-2 points. So the robustified stage-2 constraints are not calibrated to the actual closed-loop uncertainty, and the chance-constraint claim is not a guarantee. The rest of the method can still work as a heuristic, but the safety part of the contribution is weaker than the paper's wording suggests.\n\nTwo smaller issues. Algorithm 2 has no convergence proof; saying 'if KKT conditions meet a threshold' is not an argument that the alternation reaches a fixed point. And the timing comparison is not apples-to-apples: single planning uses KKT tolerance 5e-3 while replanning uses 5e-5, which biases the reported computation times. The evaluation is also a single simulation with no code or data, so I would not over-read the numbers.\n\nWho is this for? Control and robotics people working on real-time time-optimal motion planning under uncertainty. They will get a usable framework and a clear formulation; they should treat stage-2 margins as heuristic until the covariance is propagated or a bound is proven.\n\nRecommendation: send it to peer review, but push the authors to either propagate covariance through stage 2 or explicitly weaken the safety claim to a heuristic margin. That is fixable, and the integration itself deserves a serious look.","headline":"A credible two-stage robust time-optimal motion planning integration with a real gap: stage-2 safety margins are not backed by covariance propagation.","tokens_in":12367,"tokens_out":3804,"would_cite":false,"duration_ms":33760,"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":"Time-optimal point-to-point motion under process noise can be planned and controlled in one robustified two-stage optimal-control problem, solved online by alternating Riccati recursion with a nominal time-optimal OCP.","keywords":["time-optimal control","motion planning under uncertainty","two-stage optimal control","Riccati recursion","model predictive control","chance constraints","feedback control","mobile robot"],"falsifier":"Run the planned stage-2 motion on the real nonlinear system, or in Monte Carlo simulation, with the assumed process noise and count how often the obstacle constraint or control limits are violated; if the violation rate exceeds the chosen probability level $p = 1 - C(\\sigma)$ during the variable-time stage, the frozen-margin premise is refuted. A targeted version would lengthen $T_2$ (for instance, by a more distant target) and check whether violation frequency grows with stage-2 duration.","tokens_in":11299,"feed_emoji":"🤖","tokens_out":8083,"duration_ms":64683,"temperature":0.7,"pith_summary":"This paper argues that time-optimal point-to-point motion under process noise can be planned and controlled in one robustified two-stage optimal-control problem, solved fast enough to be replanned online. Stage 1 uses a fixed time grid and optimizes the nominal trajectory together with feedback gains and state covariances, shrinking uncertainty before a variable-time stage 2 that minimizes the remaining motion time. Constraints in both stages are robustified with safety margins derived from those covariances, replacing heuristic margins with a chance-constraint rationale. A tailored algorithm alternates a Riccati recursion for the feedback gains with a nominal time-optimal OCP, and an asynchronous replanning loop keeps the plan feasible. In the paper's numerical example the replanned motion takes 5.22 s versus 5.2 s for single planning, with every replanning solve finishing within the 0.6 s stage-1 horizon.","feed_headline":"Two-stage solver replans safe time-optimal motion within 0.6 s","feed_subtitle":"Fast replanning plus robustified constraints yields 5.22-second motion versus 5.2 seconds for single planning.","key_machinery":"The load-bearing object is the two-stage OCP (6), whose first stage has a fixed time grid and optimizes nominal states and controls, feedback gains $K_1$, covariances $\\Sigma_1$, and constraint variances $\\beta_1$, while the second stage has a variable time grid and minimizes total time $T_2$. The robustification mechanism is the covariance recursion $\\Sigma_{n+1} = \\Phi(\\bar{s}_n, \\bar{u}_n, K_n, \\Sigma_n)$ from linearization around the nominal trajectory, together with the constraint backoff $h + \\sigma\\sqrt{\\beta} + \\epsilon \\le 0$ where $\\beta$ is computed from the linearized constraint and the current covariance. The algorithm that carries the tractability claim is the alternation of a Riccati recursion for feedback gains and a nominal two-stage OCP with frozen safety margins and gradient corrections, iterating until the KKT conditions of the coupled problem are met.","core_discovery":"On its own terms, the paper establishes that the robustified two-stage OCP (6)—a fixed-grid stage 1 with optimized gains and covariances stitched to a variable-time stage 2 whose total time is minimized—is a tractable formulation for time-optimal motion planning and control under Gaussian process noise, and that a decoupled solver can handle it online. The solver alternates between a Riccati recursion that computes optimal feedback gains for the current nominal trajectory and a nominal two-stage OCP armed with frozen safety margins, iterating until the KKT conditions of the full problem are met. Embedded in an ASAP-MPC replanning loop, the approach keeps constraints robustified without a separate tracking layer. In the reported unicycle example, timely replanning reaches the target in 5.22 s while single planning takes 5.2 s, and each replanning solve completes within the 0.6 s stage-1 horizon.","pith_inferences":["The frozen stage-2 covariance is the main extrapolation risk: if a longer variable-time stage 2 lets uncertainty grow substantially, the fixed safety margins could under-cover; one way to test this is to scale the process-noise covariance or lengthen the second-stage duration.","The same alternation of Riccati recursion and nominal OCP could be applied to other robustification schemes, such as tube-based or zero-order robust NMPC, potentially carrying the online-computability property to those settings.","The reported computation times depend on a specific laptop and solver settings, so the architecture's real-time claim should be re-verified on embedded hardware, a step the paper itself lists as future work."],"forward_implications":["Time-optimal motion planning and feedback control are solved in one OCP, so the planned trajectory, feedback gains, and covariance are consistent instead of being designed separately.","Safety margins in the constraints come from propagated state covariances, giving a formal chance-constraint interpretation rather than heuristic tuning, and the margins shrink near activated obstacle constraints.","The tailored alternating solver makes the robustified OCP solvable in real time: each replanning in the example finishes within the 0.6 s stage-1 horizon.","Closing the loop via ASAP-MPC keeps feasibility under replanning delays; the final approach uses an exponentially weighted OCP to reach exactly the target.","The cost of robustness in total motion time is small in the example: 5.22 s with timely replanning versus 5.2 s for single planning."],"supporting_citations":[{"why":"Supplies the previous two-stage time-optimal OCP that this paper robustifies with feedback gains and covariances.","marker":"[8]"},{"why":"Supplies the alternating Riccati/nominal-OCP decomposition and the lemma used to rewrite the feedback-gain subproblem.","marker":"[15]"},{"why":"Predecessor robustified time-optimal formulation with time scaling that motivates the two-stage split.","marker":"[17]"},{"why":"Provides the ASAP-MPC asynchronous update scheme underlying the timely replanning loop.","marker":"[18]"},{"why":"Provides the ellipsoidal-tube covariance propagation idea used in the covariance recursion.","marker":"[10]"},{"why":"Supplies the chance-constraint robustification with the probability factor sigma and the backoff term.","marker":"[19]"},{"why":"Interior-point NLP solver used to solve the nominal OCP subproblem in Algorithm 2.","marker":"[20]"}],"fun_headline_variants":["Robustified two-stage OCP replans motion in 0.6 s","Safe replanning under uncertainty: 0.6 s solves, no tracking layer","Gains and covariances robustify time-optimal replanning","Uncertainty-aware motion: robust two-stage replanning within horizon","Replan with robustified constraints: time-optimal in 0.6 s"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The paper assumes that the uncertainty at the end of stage 1 remains a valid measure of the uncertainty during the entire variable-time stage 2: the stage-2 safety margins freeze the stage-1 terminal covariance $\\Sigma_{1,N_1}$ and feedback gain $K_{1,N_1-1}$, so if process noise accumulates more as the second stage runs longer, the constraints could be under-protected.","fun_headline_variants_meta":{"raw":{"variants":["Robustified two-stage OCP replans motion in 0.6 s","Safe replanning under uncertainty: 0.6 s solves, no tracking layer","Gains and covariances robustify time-optimal replanning","Uncertainty-aware motion: robust two-stage replanning within horizon","Replan with robustified constraints: time-optimal in 0.6 s"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000299,"raw_usage":{"total_tokens":1693,"prompt_tokens":877,"completion_tokens":816,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":493,"completion_tokens_details":{"reasoning_tokens":713}},"tokens_in":493,"tokens_out":816,"duration_ms":7191,"temperature":1.0,"reasoning_tokens":713,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-10T15:03:55.354311+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the planned stage-2 motion on the real nonlinear system, or in Monte Carlo simulation, with the assumed process noise and count how often the obstacle constraint or control limits are violated; if the violation rate exceeds the chosen probability level $p = 1 - C(\\sigma)$ during the variable-time stage, the frozen-margin premise is refuted. A targeted version would lengthen $T_2$ (for instance, by a more distant target) and check whether violation frequency grows with stage-2 duration.","supporting_citations":[{"cited_title":"Time-optimal point-to-point motion plan- ning: A two-stage approach,","cited_arxiv_id":null,"evidence_quote":"Supplies the previous two-stage time-optimal OCP that this paper robustifies with feedback gains and covariances."},{"cited_title":"An efficient algorithm for tube-based robust nonlinear optimal control with optimal linear feedback,","cited_arxiv_id":null,"evidence_quote":"Supplies the alternating Riccati/nominal-OCP decomposition and the lemma used to rewrite the feedback-gain subproblem."},{"cited_title":"Robustified time-optimal collision-free motion planning for autonomous mobile robots under disturbance conditions,","cited_arxiv_id":null,"evidence_quote":"Predecessor robustified time-optimal formulation with time scaling that motivates the two-stage split."},{"cited_title":"ASAP-MPC: An Asynchronous Update Scheme for Online Motion Planning with Nonlinear Model Predictive Control","cited_arxiv_id":"2402.06263","evidence_quote":"Provides the ASAP-MPC asynchronous update scheme underlying the timely replanning loop."},{"cited_title":"Robustness and stability optimization of power generating kite systems in a periodic pumping mode,","cited_arxiv_id":null,"evidence_quote":"Provides the ellipsoidal-tube covariance propagation idea used in the covariance recursion."},{"cited_title":"Practical methods for approximate robust periodic optimal control of nonlinear mechanical systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the chance-constraint robustification with the probability factor sigma and the backoff term."},{"cited_title":"On the implementation of an interior- point filter line-search algorithm for large-scale nonlinear program- ming,","cited_arxiv_id":null,"evidence_quote":"Interior-point NLP solver used to solve the nominal OCP subproblem in Algorithm 2."}],"review_version":1}