{"id":"3ea21b5f-b20b-44ba-9e43-8dc4c1794079","arxiv_id":"2509.05201","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":6,"one_line_summary":"A perception-aware tube MPC that treats CNN-based perception noise as a bounded zonotope and is solved as an LP, with hardware validation on a mobile robot.","lead":"This paper describes a control method for robots that use a camera and a neural network to locate themselves, where the network's errors can be biased and heavy-tailed rather than Gaussian. It keeps those errors inside a geometric shape called a zonotope and solves the resulting robust control problem as a fast linear program, with tests on a real omnidirectional robot.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Implemented terminal set is a zonotopic inner approximation (Remark 5) with no proof of invariance; stability guarantee holds only for the ellipsoid of Theorem 4, so recursive feasibility of Algorithm 1 is unproven.","rationale":"The reader's weakest_assumption identifies the most load-bearing gap: the paper provides theoretical guarantees for an ellipsoidal terminal set (Theorem 4) but implements a zonotopic approximation (Remark 5) without proving that the zonotope is invariant or satisfies the required inclusions. This is not a cosmetic issue; recursive feasibility and closed-loop stability of the MPC depend on the terminal set being positively invariant under the terminal control law. The paper does not supply a theorem that the zonotope Z_f satisfies (22b), and the cited conversion result [41] does not by itself preserve invariance. Thus, the central claim that Algorithm 1 ensures closed-loop stability is not supported for the algorithm actually run. This is a correctness risk rather than a matter of external consensus, and it is internal to the paper's argument. I agree with the reader's assessment and recommend no change to the conditional verdict, as the issue is substantive but not necessarily fatal if the authors can prove or construct an invariant zonotopic terminal set.","tokens_in":20653,"tokens_out":10149,"duration_ms":109287,"concrete_test":"For the system in Section VI.B.3 (A=I_3, B as reported, K_f as reported), construct the zonotope Z_f used in the experiments and compute (A+BK_f)Z_f via zonotope propagation (Lemma 4). Use Lemma 1's LP to check (A+BK_f)Z_f ⊆ Z_f, Z_f ⊆ Sbar_x, and K_f Z_f ⊆ Sbar_u. If any containment fails, the terminal conditions (22b) are violated, so Algorithm 1 lacks a stability guarantee. Complementary check: simulate the closed-loop MPC from the reported initial condition and detect any infeasibility of (15) or constraint violation.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Theorem 4 proves that the maximum-volume ellipsoid S_f=<P>_E satisfies (A+BK_f)S_f⊆S_f, S_f⊆Sbar_x, and K_f S_f⊆Sbar_u. Remark 5 replaces S_f with an inscribed zonotope Z_f (via [41]) to keep the MPC an LP, and Algorithm 1 (steps 4–5) plus Section VI.B.3 use Z_f as the termal set. However, the paper does not prove that Z_f inherits invariance or the inclusion constraints. Invariance of an ellipsoid does not imply invariance of an inner zonotope; a smaller set need not be invariant under the same dynamics. Consequently, the standard recursive-feasibility argument for MPC stability fails for the executed algorithm: the shifted optimal trajectory plus terminal control may leave Z_f, making (15g) infeasible at the next step. This is an internal inconsistency between the theory and the implementation, directly affecting the central claim of guaranteed closed-loop stability. A secondary concern: the measurement-noise zonotope in (34) is fitted to 95% of the data, so the bounded-noise assumption is not fully validated for the experiments.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a robust tube-based MPC framework for linear systems in which the state estimate is provided by a perception module (CNN) modeled as a noisy sensor with non-Gaussian, possibly biased noise. The uncertainty is represented by constrained zonotopes, and the observer gain and state-feedback gain are designed so that the estimation error and the deviation between observer and nominal state stay in contractive zonotopic sets (Theorems 1 and 2). The terminal cost is based on Minkowski functions of polar sets (Theorem 3), and the terminal set is initially computed as a maximum-volume ellipsoid satisfying invariance and tightened constraint inclusions (Theorem 4). To keep the MPC a linear program, the terminal set is then replaced by an inscribed zonotope (Remark 5, Algorithm 1), and the stage cost is replaced by an infinity/one-norm slack formulation (Remark 6). The method is validated in simulation and on a ROSbot XL with a custom CNN perception module, comparing favorably with a Gaussian-Kalman-based MPC.","tokens_in":20970,"tokens_out":12157,"duration_ms":137812,"significance":"If the gaps identified below are addressed, the paper would be a useful contribution to perception-based control: it tackles a practically important issue (non-Gaussian, biased perception noise), builds a coherent set-membership observer-plus-MPC pipeline, and includes real-hardware validation. Theorems 1-3 are standard containment derivations and appear essentially correct under the stated assumptions. The main value is the integration of constrained zonotopes, Minkowski-Lyapunov terminal costs, and an LP formulation that is fast enough for onboard use. However, the gap between the terminal set for which invariance is proved (an ellipsoid) and the terminal set actually implemented (a zonotope) directly affects the central closed-loop stability claim; until that gap is closed, the theoretical guarantees do not cover the implemented algorithm.","major_comments":[{"comment":"The implemented terminal constraint (15g) uses a zonotopic inner approximation Z_f of the ellipsoidal terminal set S_f from Theorem 4. Theorem 4 proves invariance and constraint inclusions only for S_f: (A+BK_f)S_f⊆S_f, S_f⊆S_bar{x}, and K_fS_f⊆S_bar{u}. Invariance of S_f does not imply invariance of an inner subset Z_f. In particular, the standard recursive-feasibility argument requires that the shifted terminal state, propagated under A+BK_f, remain in Z_f; no such property is proved for the inscribed zonotope. With λ_f=0.95 in Section VI.B.3 and n=3, an inner zonotope that is also inside S_f cannot generally contain λ_fS_f, so the usual sufficient condition for invariance of a subset is not available. The paper should prove invariance and the inclusion conditions for Z_f (e.g., by computing a contractive zonotopic terminal set directly), implement the ellipsoidal terminal set and solv","section":"§V, Remark 5 and Algorithm 1, Step 5; §VI.B.3"},{"comment":"The implemented LP replaces the stage cost in (15a) by a single slack variable β subject to (33b)-(33c). The paper asserts that the invariance and stability results of Theorem 3 remain valid because β is an upper bound on the cost, but no proof is given for the modified problem. The value function of the modified ∞/∞-cost problem is not immediately non-increasing; one must construct a candidate β_{k+1} from the optimal β_k, remove the first stage cost, add the terminal stage cost, and use the terminal inequality (22a). This argument is missing. Since Algorithm 1 solves the modified problem, the stability guarantee for the actual controller is incomplete.","section":"§V, Remark 6 and Eq. (33)"},{"comment":"The measurement-noise zonotope Z_v is computed using 95% of the perception-noise data. Assumption 6 requires v(k)∈Z_v for all k. The remaining 5% of test points lie outside Z_v, so the bounded-noise hypotheses of Theorems 1 and 2 are not satisfied in the experiments. Consequently, the experimental results do not demonstrate the certified bounded-error or closed-loop stability properties for the actual noise distribution. The authors should either enlarge Z_v to contain all observed errors, provide a statistical containment argument compatible with the theory, or explicitly characterize the hardware experiments as heuristic validation outside the assumptions.","section":"§VI.B.2, Eq. (34)"}],"minor_comments":[{"comment":"The notation diag(q_x) and diag(q_u) is dimensionally inconsistent with the proof: the proof requires ∥P^{1/2}q_{x,i}∥_2 ≤ q_{x,i}, i.e. the diagonal entries should be the squares of the scalar bounds. Please clarify the notation.","section":"Theorem 4, Eq. (25c)-(25d)"},{"comment":"Constraint (15h) fixes the initial nominal state to x0, but online only the estimate ˆx(k) is available. The text later uses ¯x(k|k) as a decision variable. Please clarify how (15h) is initialized in Algorithm 1 and how the estimate enters the optimization.","section":"Problem 1, Eq. (15h)"},{"comment":"Step 5 says 'Find the maximum size terminal set S_f using Theorem 4 and Remark 5', but Remark 5 replaces the ellipsoid with a zonotopic approximation. This wording is misleading; the set is no longer the maximum-volume ellipsoid and may not be maximal in any stated sense.","section":"Algorithm 1, Step 5"},{"comment":"Typo: 'constrianed zonotpes' should be 'constrained zonotopes'.","section":"Lemma 1, Section II.C"}],"recommendation":"major_revision","confidential_remarks":"The manuscript is technically interesting and the experimental demonstration is valuable, but the central stability guarantee is currently attached to an ellipsoidal terminal set that is not the one implemented. This is a fixable mismatch, but it must be resolved before the paper can be accepted. I would also ask the authors to provide the missing stability proof for the ∞/∞-cost variant and to reconcile the 95% noise zonotope with the bounded-noise assumption."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"At a glance, the paper does something genuinely useful: it takes set-based estimation with constrained zonotopes, couples it to a tube-based MPC that stays linear, and shows the whole loop running on a real robot with a CNN perception module. The observer and state-deviation bounds (Theorems 1–2) are clean applications of zonotope containment, and the Minkowski–Lyapunov terminal design (Theorems 3–4) is a nice fit for LP-MPC. I'd probably use the observer and constraint-tightening ideas in my own work.\n\nThe catch is the terminal set. The authors prove that the maximum-volume ellipsoid from Theorem 4 is invariant under A+BKf and satisfies the state/input inclusions. Then, to keep the MPC an LP, they replace it with an inscribed zonotope (Remark 5, Algorithm 1). Nothing proves that the zonotope inherits invariance. An inner approximation of an invariant set need not be invariant—the image of the zonotope under A+BKf can leave it while staying inside the ellipsoid. Without invariance of the actual terminal set, the standard recursive-feasibility argument fails, and the closed-loop stability claim is not supported for the algorithm that is actually executed. This is the load-bearing gap, and it's real.\n\nA second issue: the measurement-noise zonotope in (34) is fit to 95% of the data. The bounded-noise assumption (Assumption 6) needs a hard bound. With the Laplace noise in simulation, the bound is exceeded. So the theoretical guarantees apply to a problem whose noise set is slightly larger than what's used. That's a treatable inconsistency, but it should be acknowledged.\n\nSmaller items: a single hardware run, no error bars, and no released code/data make the experimental claims less convincing than they could be.\n\nOverall, the integration is competent and the theory is mostly correct in isolation. The gap between theory and implementation is the main thing. It's fixable—compute a zonotopic terminal set that is actually invariant, or explicitly treat the LP version as a heuristic with experimental support. As is, the paper deserves a serious referee, but it needs revisions before the stability claims can be taken as proved.","headline":"Useful integration of zonotope estimation with LP-MPC, but the stability guarantee doesn't cover the implemented terminal set—fixable, but load-bearing.","tokens_in":21471,"tokens_out":4567,"would_cite":false,"duration_ms":50043,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper argues that vision-based perception noise is biased and heavy-tailed, so robust MPC should bound it with zonotopes instead of assuming Gaussian noise — and shows the design running on a real robot.","keywords":["Perception","state estimation","uncertainty","zonotopes","linear programming","model predictive control","heavy-tailed noise","observer design"],"falsifier":"Recompute the inscribed terminal zonotope $Z_f$ for the paper's hardware design and test $(A+BK_f)Z_f \\subseteq Z_f$ and the inclusions $Z_f \\subseteq S_{\\bar{x}}$, $K_f Z_f \\subseteq S_{\\bar{u}}$; if any vertex maps outside, recursive feasibility is not certified. Alternatively, run the closed loop under a fixed bias plus Laplace noise and check whether the observed estimation error ever leaves the predicted zonotopic bound — Theorem 1 says it cannot.","tokens_in":20537,"feed_emoji":"🤖","tokens_out":9316,"duration_ms":94244,"temperature":0.7,"texified_at":"2026-08-05T20:25:24.779099+00:00","pith_summary":"Perception modules built on neural networks do not produce zero-mean Gaussian errors; in practice their errors are biased, heavy-tailed, and worse far from the training distribution. The paper's claim is that robust control should treat those errors as unknown-but-bounded sets, and it constructs a complete control loop on that principle: a zonotopic observer bounds the estimation error, a tube-based MPC keeps state and input inside tightened constraints, and a Minkowski-function cost makes the online optimization a linear program. The stability argument runs through Minkowski–Lyapunov inequalities and contractive zonotopic invariant sets, with a maximal-volume terminal ellipsoid (replaced by an inscribed zonotope in the implementation) providing recursive feasibility. Simulations and a hardware experiment with a CNN-based camera perception module on an omnidirectional robot support the claim: the perception-aware MPC keeps estimates inside their bound sets and achieves lower cumulative cost than a Gaussian Kalman/LQR baseline under heavy-tailed noise. If the claim holds, it matters because it removes the Gaussian assumption from a safety-critical component of autonomous vehicles without giving up real-time computation.","texify_model":"deepseek-v4-flash","texify_usage":{"total_tokens":4825,"prompt_tokens":812,"completion_tokens":4013,"prompt_tokens_details":{"cached_tokens":0},"prompt_cache_hit_tokens":0,"prompt_cache_miss_tokens":812,"completion_tokens_details":{"reasoning_tokens":3271}},"feed_headline":"Zonotopic MPC tames biased, heavy-tailed perception noise","feed_subtitle":"Set-based state bounds keep a vision-driven robot stable where Gaussian Kalman filters fail.","key_machinery":"The load-bearing object is the constrained zonotope — a polytope written as $\\{c+G\\alpha \\mid |\\alpha|\\le 1, F\\alpha=\\theta\\}$ — used for noise bounds, invariant sets, and cost sets. The argument moves through the Minkowski function $g(S,x)$, the smallest $\\lambda \\ge 0$ with $x \\in \\lambda S$, which becomes the stage and terminal cost; because containment of one constrained zonotope in another can be written as a linear program (Lemma 1), four design steps remain convex: making the error set and the deviation set $\\lambda$-contractive, enforcing the Minkowski–Lyapunov terminal inequality via containment of polar sets, and computing the maximal terminal ellipsoid. The final conversion from that ellipsoid to an inscribe","core_discovery":"The central claim is that for a perception-based observer, the correct uncertainty model is a bounded set rather than a Gaussian distribution. The paper models process noise and perception noise as constrained zonotopes, then designs an observer gain L so that the estimation-error set propagates inside a $\\lambda_L$-contracted version of itself, and a feedback gain K so the state-deviation set does the same. This gives the tube: the true state is always contained in the nominal state plus the sum of the two invariant sets, so tightening the original state and input constraints by that sum yields robust constraint satisfaction. The MPC minimizes a Minkowski-gauge cost (state and input gauges joi","pith_inferences":["A direct invariance proof for the inscribed zonotope terminal set, rather than the ellipsoid it approximates, would close the gap between the stability theorem and the implemented LP; as written, the certificate applies to the ellipsoid.","The zonotopic observer design is modular: it should transfer to other perception modalities (LiDAR, radar, visual odometry) as long as the error admits a bound around the operating region, which the paper motivates but does not test.","One stress test the paper leaves implicit is varying the bias-to-variance ratio of the perception error; the zonotopic design should dominate at high bias, while a tuned Gaussian design may remain competitive at low bias.","The infinity/infinity slack cost could be reused in other set-based MPC settings as a cheap way to avoid the idle/deadbeat artifacts of 1-norm costs, independent of the perception application."],"forward_implications":["Robust constraint satisfaction no longer depends on knowing a Gaussian covariance for perception noise; a zonotopic bound around the learned map is sufficient.","The online MPC can be solved as a linear program, which is fast enough for onboard execution on small robots with a ROS2 stack.","The design exposes explicit tuning parameters — contraction ratios lambda_L, lambda_tilde{x}, and lambda_f — that trade convergence speed against the size of the invariant tubes.","For any perception module whose error is bounded near its training distribution, the same pipeline gives bounded state-estimation error and constraint satisfaction regardless of the noise's mean, color, or tail weight.","The hardware comparison indicates that the advantage over Gaussian designs grows when the noise is biased, because a Kalman-type estimator cannot remove the bias while the zonotopic observer encloses it."],"supporting_citations":[{"why":"Defines constrained zonotopes and their Minkowski-sum/linear-image calculus, the representation used for all noise and invariant sets.","marker":"[30]"},{"why":"Provides Lemma 1, the LP containment condition used to make error, deviation, and terminal sets fit inside their contracted targets.","marker":"[35]"},{"why":"Supplies the Minkowski terminal cost framework that Theorem 3 converts into a polar-set containment condition.","marker":"[33]"},{"why":"Introduces the infinity/infinity cost with a single slack variable, adopted to keep the LP from exhibiting idle or deadbeat behavior.","marker":"[32]"},{"why":"Gives the zonotope-ellipsoid conversion used in Remark 5 to approximate the terminal ellipsoid by an inscribed zonotope while preserving linearity.","marker":"[41]"},{"why":"Establishes the robust perception-based system model with bounded perception error near the operating point, which underlies Assumption 5.","marker":"[10]"}],"fun_headline_variants":["Guaranteed set-based tubes stabilize perception-driven mobile robots","Zonotopic MPC trumps Gaussian filters for heavy-tailed perception errors","Robust MPC with set-based observers beats Kalman on vision robots","Perception-aware controller bounds error sets, not just means","Set-based MPC for vision-driven robots: stability under heavy-tailed noise"],"cache_read_input_tokens":2688,"weakest_assumption_plain":"The stability argument assumes the zonotope used in the algorithm in place of the terminal ellipsoid still satisfies the invariance and constraint-inclusion properties proven only for the ellipsoid; the paper does not prove that inheritance.","fun_headline_variants_meta":{"raw":{"variants":["Guaranteed set-based tubes stabilize perception-driven mobile robots","Zonotopic MPC trumps Gaussian filters for heavy-tailed perception errors","Robust MPC with set-based observers beats Kalman on vision robots","Perception-aware controller bounds error sets, not just means","Set-based MPC for vision-driven robots: stability under heavy-tailed noise"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001254,"raw_usage":{"total_tokens":4980,"prompt_tokens":750,"completion_tokens":4230,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":494,"completion_tokens_details":{"reasoning_tokens":4142}},"tokens_in":494,"tokens_out":4230,"duration_ms":30109,"temperature":1.0,"reasoning_tokens":4142,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-05T05:31:53.129343+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Recompute the inscribed terminal zonotope $Z_f$ for the paper's hardware design and test $(A+BK_f)Z_f \\subseteq Z_f$ and the inclusions $Z_f \\subseteq S_{\\bar{x}}$, $K_f Z_f \\subseteq S_{\\bar{u}}$; if any vertex maps outside, recursive feasibility is not certified. Alternatively, run the closed loop under a fixed bias plus Laplace noise and check whether the observed estimation error ever leaves the predicted zonotopic bound — Theorem 1 says it cannot.","supporting_citations":[{"cited_title":"Computing the maximal positive invariant setfor the constrained zonotopic case,","cited_arxiv_id":null,"evidence_quote":"Provides Lemma 1, the LP containment condition used to make error, deviation, and terminal sets fit inside their contracted targets."},{"cited_title":"Minkowski terminal cost functions for mpc,","cited_arxiv_id":null,"evidence_quote":"Supplies the Minkowski terminal cost framework that Theorem 3 converts into a polar-set containment condition."},{"cited_title":"Model predictive optimal averaging level control,","cited_arxiv_id":null,"evidence_quote":"Introduces the infinity/infinity cost with a single slack variable, adopted to keep the LP from exhibiting idle or deadbeat behavior."},{"cited_title":"Scalable zonotope-ellipsoid conversions using the euclidean zonotope norm,","cited_arxiv_id":null,"evidence_quote":"Gives the zonotope-ellipsoid conversion used in Remark 5 to approximate the terminal ellipsoid by an inscribed zonotope while preserving linearity."},{"cited_title":"Certainty equivalent perception-based control,","cited_arxiv_id":null,"evidence_quote":"Establishes the robust perception-based system model with bounded perception error near the operating point, which underlies Assumption 5."}],"review_version":1}