{"id":"dd5e7bb1-8b34-44b3-ba88-4a9006dbdc01","arxiv_id":"2507.13613","paper_version":1,"verdict":"CONDITIONAL","confidence":"HIGH","novelty_score":6.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":1,"one_line_summary":"A conformal-prediction-based contraction controller guarantees, with probability 1-alpha, that the tracking error stays below an explicit exponential bound despite unknown nonlinear uncertainty.","lead":"This paper merges contraction theory with conformal prediction to produce distribution-free, finite-time probabilistic bounds on trajectory tracking error for continuous-time nonlinear systems with unstructured, state- and control-dependent uncertainty. It also constructs probabilistically robust control invariant tubes that can shrink the state constraints in motion planning so that a perturbed system stays safe with high probability.","discovery_kind":"new_method","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Theorem 2's motion-planning guarantee is unsupported: tightened constraints break the exchangeability assumption of Lemma 3, and Remark 1's two-step fix is unproved and potentially circular.","rationale":"The central Theorem 1 is internally coherent: the contraction inequality, the conformal residual bound, and the comparison lemma are standard; the proof's metric factorization and geodesic constant-speed argument are acceptable, and the sqrt(m) factor is consistent with the upper metric bound used in the tube-radius optimization of Eq. (21). The genuinely fragile load-bearing premise is the transfer of that guarantee to the motion-planning setting, where the tightening feedback loop breaks exchangeability. This is not a disagreement with consensus or an ad hominem; it is a gap between a stated theorem and its proof, and the manuscript itself flags the issue in Remark 1. The reader's weakest_assumption pointed to the same item, and I agree. Because Theorem 1 remains valid under its explicit exchangeability hypothesis, the appropriate verdict is unchanged: conditional acceptance pending a formal treatment of the two-step calibration and an experiment that actually tightens constraints.","tokens_in":13694,"tokens_out":13145,"duration_ms":166503,"concrete_test":"Numerically test the exact pipeline of Theorem 2 on the Section V-A example. Split calibration into halves. Step 1: from half 1 compute s1 and R1 = sqrt(m) s1 / lambda. Step 2: plan a set of reference trajectories with the tightened sets (19) using R1, then simulate the closed-loop system (1),(6) from initial conditions drawn from the same distribution; compute the empirical fraction of trials with d_RM(x(t), xbar(t)) <= R1 and x(t) in S(t) for all t. Step 3: compute the half-2 residual scores under this new distribution, obtain s2, and compare R2 with R1. If coverage is below 1-alpha or R2 > R1 substantially, the Theorem 2 guarantee fails as stated and the Remark 1 fix needs a formal nested-calibration argument.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The load-bearing concern is the calibration/exchangeability step in Theorem 2. Theorem 1's guarantee is a standard split-conformal statement: if the new closed-loop trajectory is exchangeable with D_cal in (12), the residual quantile in Lemma 3 bounds sup_T ||R(t)||, and the comparison lemma gives (15). That part of the argument is coherent. The problem is that Theorem 2 applies this guarantee after the motion planner (18)-(19) tightens constraints using the PRCI tube of Corollary 1. The tightening changes the reference-trajectory distribution, hence the distribution of the closed-loop state-control trajectories and of the scores in (13). The test sample is therefore not exchangeable with D_cal, which was generated from untightened reference trajectories in Section III-B. Remark 1 acknowledges this and sketches a two-step calibration (first half sets the PRCI radius, second half recalibrates), but no theorem or proof is provided, and as written the fix is circular: the radius from the first half determines the tightening, which determines the distribution for the second half; if the second half produces a different radius, the planner should be re-run, changing the distribution again. Thus the claim Pr[x(t) in S(t) and u(x(t),t) in A(t) for all t] >= 1-alpha is not established. The Section V motion-planning tests use unconstrained random reference trajectories, so they do not exercise Theorem 2's tightened-constraint regime.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript proposes a data-driven robust control framework for continuous-time perturbed nonlinear systems. The key idea is to combine contraction-based control with conformal prediction: an arbitrary predictor \\hat{\\zeta} is trained to approximate the state- and control-dependent uncertainty \\zeta, and a conformal calibration dataset is used to bound the closed-loop residual R(t) in (8) by a quantile s_zeta^(j_alpha) with probability at least 1-alpha. Theorem 1 then uses a comparison lemma to derive a finite-time incremental exponential boundedness bound for the Riemannian distance between the perturbed and reference trajectories. Corollary 1 converts this bound into a probabilistically robust control invariant (PRCI) tube, and Theorem 2 applies a tightened motion-planning problem to state that perturbed trajectories and inputs remain in the original constraint sets with probability 1-alpha. The paper also gives an SOS formulation for computing the contraction metric and validates the approach in two numerical examples, one with parametric uncertainty and one planar VTOL.","tokens_in":13941,"tokens_out":9341,"duration_ms":110206,"significance":"The main conceptual contribution is to carry conformal uncertainty quantification through a contraction-theoretic comparison argument in continuous time, yielding a distribution-free bound that is agnostic to the uncertainty model and to the predictor. This is a useful step beyond prior conformal control results, which are mostly discrete-time. The proof of Theorem 1 is coherent given the exchangeability assumption of Lemma 3, and the empirical coverage (96.7% in the first experiment) is consistent with the nominal 95% level. However, the motion-planning guarantee in Theorem 2 is not established as stated because the PRCI tightening changes the data-generating process and breaks the exchangeability on which Lemma 3 depends. Since the PRCI tube is a headline contribution, this gap materially reduces the significance of the paper in its current form.","major_comments":[{"comment":"The proof of Theorem 2 applies Corollary 1, which inherits the exchangeability assumption of Lemma 3. However, the constraint tightening in (18)-(19) changes the distribution of the reference trajectory (xbar, ubar) relative to the distribution used to build the calibration set D_cal in (12), because D_cal is generated in Section III-B from the reference policies in D_ref before any tightening is applied. Thus the new test data point is not exchangeable with D_cal, and the conformal bound (14) need not hold for the planned trajectory. Remark 1 acknowledges this issue and sketches a two-step calibration, but it does not state or prove a theorem for the two-step procedure; as written the procedure is circular, since the first-step quantile determines the PRCI radius, which determines the tightening, which determines the distribution of the second-step calibration, and if the second step produces a different radius the planner must be re-run, changing the distribution again. The claimed guarantee Pr[x(t) in S(t) and u(x(t),t) in A(t) for all t in T] >= 1-alpha is therefore not established. To repair this, the authors should either prove a formal nested-calibration result using a dataset generated under the tightened distribution, or state Theorem 2 under an explicit assumption that the planned reference trajectory is exchangeable with the reference trajectories used to build D_cal.","section":"Section IV-B, Theorem 2 and Remark 1"},{"comment":"The construction of D_cal does not specify the distribution of the reference policies \\bar u^(k) used to generate the calibration trajectories, but Lemma 3 requires the new test trajectory to be exchangeable with these data. For Theorem 2, the test reference is produced by the motion planner (18), so exchangeability requires that the calibration reference policies be drawn from the planner's output distribution; this is not stated or verified. Moreover, the first numerical experiment applies the controller to random reference trajectories and does not test Theorem 2's tightened-constraint regime, so the simulations do not demonstrate the exchangeability needed for the PRCI tube claim.","section":"Section III-B, Eq. (12), and Section V-A"}],"minor_comments":[{"comment":"The geodesic boundary condition is stated as gamma(0,t)=x(t) and gamma(1,t)=x(t); the second endpoint should be \\bar x(t).","section":"Section III-A, Lemma 1 and Eq. (5)"},{"comment":"The two metric bounds are typeset with the same symbol m, which also collides with the control input dimension m; use \\underline{m} and \\overline{m} to avoid ambiguity with the tube radius \\bar d_RM.","section":"Section II-A, Eq. (3a)"},{"comment":"The displayed bound uses (d_RM(x(0),\\bar x(0)) - sqrt(m) s_zeta^(j_alpha)/lambda) without an absolute value, while the definition of c1 in the next sentence uses an absolute value; the notation should be harmonized.","section":"Section IV-A, Eq. (15) and following text"},{"comment":"The description of the 245 test reference trajectories should state whether they solve the nominal motion planning problem (18) or are independent random references, because this determines whether the experiment tests Theorem 1 or Theorem 2.","section":"Section V-A"},{"comment":"Ten test trajectories with zero violations is too small to confirm a 95% coverage claim; report an exact binomial confidence interval or run more trials.","section":"Section V-B"}],"recommendation":"major_revision","confidential_remarks":"The main issue is confined to Theorem 2 and its exchangeability gap; the rest of the technical machinery is sound. If the authors can supply a rigorous nested-calibration argument or explicitly restrict Theorem 2 to the exchangeable case, the paper would be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Here's the short version. The core tracking guarantee (Theorem 1) is a genuine, non-circular combination of two established tools: contraction metrics from [7], [8] and split conformal prediction from [12], [13]. The proof is a valid comparison-lemma argument given the conformal bound on the residual, and the numerical experiments are consistent with the claimed 95% coverage. That part deserves to be taken seriously.\n\nWhat's new is the continuous-time setting with uncertainty that is nonlinear in both state and control, and the fact that the predictor model is arbitrary—no structural assumptions on ζ. The PRCI tube is a natural extension, and the SOS formulation for the metric is a nice practical touch.\n\nThe soft spot is Theorem 2. The constraint tightening in (18)-(19) changes the reference-trajectory distribution, so the test trajectories are not exchangeable with the calibration set built in Section III. The paper acknowledges this in Remark 1 but only sketches a two-step calibration, with no theorem or proof. As written, the fix has a circularity: the first half determines the tube radius, which determines the tightening, which determines the distribution for the second half; if the second half yields a different radius, you'd need to re-run the planner. So the claim Pr[x(t) in S(t) and u(x(t),t) in A(t)] >= 1-α is not established. The Section V motion-planning tests use unconstrained references, so they don't exercise the tightened regime. That's a real gap, not a nitpick.\n\nMinor issues: the notation for the metric bounds (m vs \\bar{m}) is confusing, the local Lipschitz assumption on d_RM is stated without proof, and the VTOL experiment has only ten trajectories—statistically weak, though not damning.\n\nBottom line: the tracking theorem is a solid contribution and the paper is worth reading for anyone working at the intersection of conformal prediction and contraction-based control. But the motion-planning guarantee currently overclaims. A serious referee should insist on a rigorous treatment of the distribution shift—either a proven two-step calibration or a more modest claim about Theorem 2. Send it to review, but expect the authors to need to fix this before I'd trust the planning guarantee.","headline":"Solid conformal-contraction tracking theorem; the motion-planning guarantee has an unproven exchangeability fix and should not be taken at face value.","tokens_in":14493,"tokens_out":2571,"would_cite":true,"duration_ms":27940,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":["93C10","93D09","93D30"],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper proves a distribution-free, finite-time probabilistic bound on the tracking error of a controlled nonlinear system under unstructured uncertainty, and builds certified invariant tubes for motion planning from the same bound.","keywords":["contraction theory","conformal prediction","distribution-free uncertainty quantification","incremental exponential boundedness","robust control","motion planning","control invariant tube","nonlinear systems"],"falsifier":"Simulate many test trajectories from the same distribution as the calibration set, compute the score $s_\\zeta$ for each, and record the empirical fraction for which $d_{\\mathrm{RM}}(x(t),\\bar{x}(t))$ exceeds the right-hand side of (15) at any $t \\in [0,T]$; if that fraction exceeds $\\alpha$ by more than sampling error, the claimed $1-\\alpha$ guarantee fails. A second check is to repeat the experiment after applying the Theorem 2 constraint tightening without the two-step calibration of Remark 1: coverage should drop if exchangeability is genuinely broken.","tokens_in":13465,"feed_emoji":"🎯","tokens_out":8406,"duration_ms":90598,"temperature":0.7,"pith_summary":"The paper sets out to give a distribution-free, finite-time statistical guarantee for robust tracking in continuous-time nonlinear systems whose uncertainty depends nonlinearly on both state and control, with no structural assumption on the uncertainty or on the learned model that predicts it. The key move is to feed a conformal-prediction quantile of the closed-loop residual into a contraction-theoretic bound on Riemannian separation, yielding an exponential bound on the tracking error that holds with probability at least 1-α over a given horizon. The same quantile defines a probabilistically robust control invariant (PRCI) tube, so that motion plans generated with tightened constraints keep both state and input inside their admissible sets with the same probability. If correct, the framework turns black-box predictors into certified controllers for uncertain nonlinear systems.","feed_headline":"Probabilistic cap on tracking error with no disturbance model","feed_subtitle":"Contraction theory plus conformal prediction yields finite-time exponential error bounds and invariant tubes for safe planning.","key_machinery":"The load-bearing object is a contraction metric $M(x)$ satisfying the generalized conditions (3): it defines a Riemannian distance whose geodesic energy contracts at rate $\\lambda$ under the nominal dynamics. The paper couples this with a conformal score $s_\\zeta$ defined in (13) as the supremum over the horizon of the residual norm $\\|\\zeta(x,u)-B(x)B(x)^\\dagger\\hat{\\zeta}(x,u_-;\\theta)\\|$, whose calibration order statistics provide the quantile $s_\\zeta^{(j_\\alpha)}$. The proof differentiates the geodesic energy along the controlled trajectory, uses the parallel-transport identity to bound the residual term by $\\sqrt{m}\\|R(t)\\|\\sqrt{E(\\gamma)}$, and applies the comparison lemma to obtain the exponential bound.","core_discovery":"The paper's central claim is Theorem 1: for the closed-loop system driven by the uncertainty-compensating policy (6), the Riemannian distance between the perturbed trajectory $x(t)$ and the reference $\\bar{x}(t)$ satisfies $$d_{\\mathrm{RM}}(x(t),\\bar{x}(t)) \\le \\left(d_{\\mathrm{RM}}(x(0),\\bar{x}(0)) - \\frac{\\sqrt{m}\\,s_\\$zeta^{{(j_\\alpha)}}$}{\\$\\lambda$}\\right)$e^{{-\\lambda t}}$ + \\frac{\\sqrt{m}\\,s_\\$zeta^{{(j_\\alpha)}}$}{\\$\\lambda$}$$ for all $t \\in [0,T]$, with probability at least $1-\\alpha$, where $s_\\zeta^{(j_\\alpha)}$ is the conformal quantile of the residual score (13). Consequently the closed-loop system is finite-time incrementally exponentially bounded without knowing the uncertainty model or its distribution. Corollary 1 turns the steady-state radius into a PRCI tube, and Theorem 2 certifies that robustly tightened motion planning keeps the perturbed trajectory and its control inputs inside the original constraints with probability $1-\\alpha$.","pith_inferences":["If the two-step calibration sketched in Remark 1 were integrated into the main theorems, the motion-planning guarantee would remain valid under the distribution shift caused by constraint tightening; the paper leaves that fix as a remark rather than a theorem.","Because the conformal quantile enters only through the product $\\sqrt{m}s_\\zeta^{(j_\\alpha)}/\\lambda$, users can tune the contraction rate $\\lambda$ to trade transient decay against tube radius, as the paper's line-search suggestion indicates.","The residual score bundles prediction, discretization, and projection errors, so the bound can be tight even when the predictor is a poor model of the uncertainty; designing that score is where application-specific knowledge still matters.","A natural extension is to replace the exchangeability assumption with online or adaptive conformal prediction, which would open the framework to non-stationary environments; the paper lists this only as future work."],"forward_implications":["The closed-loop tracking error is guaranteed to shrink at rate $\\lambda$ down to a steady-state ball of radius $\\sqrt{m}s_\\zeta^{(j_\\alpha)}/\\lambda$, with the guarantee holding simultaneously over the whole horizon $[0,T]$.","Any predictor, including a neural network trained by arbitrary means, can be plugged into the controller and still carry the same probabilistic bound, as long as the residual scores are calibrated.","The PRCI tube in (17) is a certified safety envelope: if the initial state starts inside it, the whole trajectory remains inside with probability at least $1-\\alpha$.","Motion planning under the tightened sets (19) keeps both the state and the control input inside their original admissible sets with probability $1-\\alpha$, enabling safe trajectory generation under unstructured uncertainty.","The framework directly covers three error sources at once — prediction error, the one-step discretization error of $u_-$, and the projection error of $BB^\\dagger$ — without modeling any of them separately."],"supporting_citations":[{"why":"Supplies the foundational contraction analysis and the definition of incremental stability on which the paper's boundedness notion rests.","marker":"[6]"},{"why":"Provides the control contraction metric conditions and the convex controller synthesis that the paper augments with uncertainty prediction.","marker":"[7]"},{"why":"Gives the Riemannian contraction bound for perturbed systems and the robust motion planning formulation the paper extends to probabilistic settings.","marker":"[8]"},{"why":"Provides the comparison lemma and exponential boundedness definitions used in the proof of Theorem 1.","marker":"[2]"},{"why":"Introduces the conformal prediction quantile result (Lemma 2) that yields the distribution-free coverage guarantee.","marker":"[12]"},{"why":"Supplies the book-level treatment of conformal prediction, including exchangeability and quantile bounds, that underpins Lemma 3.","marker":"[13]"}],"fun_headline_variants":["Conformal contraction: error bounds, no disturbance model","Distribution-free finite-time error bounds for nonlinear control","Contraction meets conformal prediction for robust control","Probabilistic caps on tracking error without model knowledge","Safe control with unknown disturbances: conformal guarantee"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"Everything rests on the new test trajectory being exchangeable with the calibration trajectories — in practice, drawn from the same distribution of initial states, references, and uncertainties; the paper's own constraint tightening in Theorem 2 can change that distribution, and the fix for this is only sketched in Remark 1.","fun_headline_variants_meta":{"raw":{"variants":["Conformal contraction: error bounds, no disturbance model","Distribution-free finite-time error bounds for nonlinear control","Contraction meets conformal prediction for robust control","Probabilistic caps on tracking error without model knowledge","Safe control with unknown disturbances: conformal guarantee"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000226,"raw_usage":{"total_tokens":1455,"prompt_tokens":918,"completion_tokens":537,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":534,"completion_tokens_details":{"reasoning_tokens":464}},"tokens_in":534,"tokens_out":537,"duration_ms":5915,"temperature":1.0,"reasoning_tokens":464,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T16:20:37.232672+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Simulate many test trajectories from the same distribution as the calibration set, compute the score $s_\\zeta$ for each, and record the empirical fraction for which $d_{\\mathrm{RM}}(x(t),\\bar{x}(t))$ exceeds the right-hand side of (15) at any $t \\in [0,T]$; if that fraction exceeds $\\alpha$ by more than sampling error, the claimed $1-\\alpha$ guarantee fails. A second check is to repeat the experiment after applying the Theorem 2 constraint tightening without the two-step calibration of Remark 1: coverage should drop if exchangeability is genuinely broken.","supporting_citations":[{"cited_title":"On contraction analysis for nonlinear systems,","cited_arxiv_id":null,"evidence_quote":"Supplies the foundational contraction analysis and the definition of incremental stability on which the paper's boundedness notion rests."},{"cited_title":"Control contraction metrics: Convex and intrinsic criteria for nonlinear feedback design,","cited_arxiv_id":null,"evidence_quote":"Provides the control contraction metric conditions and the convex controller synthesis that the paper augments with uncertainty prediction."},{"cited_title":"Robust online motion planning via contraction theory and convex optimization,","cited_arxiv_id":null,"evidence_quote":"Gives the Riemannian contraction bound for perturbed systems and the robust motion planning formulation the paper extends to probabilistic settings."},{"cited_title":"Machine-learning appli- cations of algorithmic randomness,","cited_arxiv_id":null,"evidence_quote":"Introduces the conformal prediction quantile result (Lemma 2) that yields the distribution-free coverage guarantee."},{"cited_title":"V ovk, A","cited_arxiv_id":null,"evidence_quote":"Supplies the book-level treatment of conformal prediction, including exchangeability and quantile bounds, that underpins Lemma 3."}],"review_version":1}