{"id":"ed5aabbe-fc8f-4e3c-964c-e3b800ba5598","arxiv_id":"2607.07043","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":6.0,"correctness_risk":"unknown","formal_verification":"none","parameter_count":8,"one_line_summary":"A probe-conditioned memory and Koopman MPC framework reduces seal-width tracking error from 0.25-0.40 mm to 0.049 mm in deadband-affected industrial dispensing by retrieving historical actuator-behavior records matched to a 16-move target probe.","lead":"This paper builds a control system for industrial glue-dispensing robots that reuses historical operating data plus a short probe sequence to compensate for actuator deadband during recipe changes. A smart generalist might read it because it cuts commissioning time for manufacturing lines where small pressure corrections silently vanish before reaching the product.","discovery_kind":"unclear","skeptic_critique":{"model":"glm-5.2","headline":"The surrogate's deadband model (Eq. 2) is structurally identical to the controller's probe-estimation model (Eq. 8), so the 5-8x gap partly measures model-plant match rather than genuine control transfer.","rationale":"The reader correctly identified the surrogate-to-reality gap as the most load-bearing concern. I sharpen it: the issue is not just generic fidelity uncertainty but a specific structural identity between the surrogate's deadband model (Eq. 2) and the controller's probe-estimation model (Eq. 8). This identity means the benchmark cannot distinguish 'AK-MPC is a better controller' from 'AK-MPC's internal model matches the test plant's nonlinearity.' The paper is honest about this limitation (§VII explicitly calls the comparison a 'calibrated digital-twin benchmark' and the physical trace a 'sanity check'), which is why CONDITIONAL rather than REJECT is appropriate. The PCM concept, the feasibility-filtering framework, and the ablation methodology are legitimate contributions regardless of whether the magnitude transfers. The proposed concrete test (mismatched deadband structure) is feasible because the surrogate is self-contained and the controller code is released; it would directly measure how much of the gap is structural-match artifact versus genuine control advantage. If the gap persists under structural mismatch, the paper would move toward ACCEPT; if it collapses, the practical magnitude claim needs revision. The reader's verdict of CONDITIONAL with MODERATE confidence is well-calibrated.","tokens_in":18334,"tokens_out":2457,"duration_ms":95898,"concrete_test":"Re-run the 60-case benchmark on a modified surrogate where the deadband uses a smooth sigmoid transition (e.g., D(u) = g·u·(0.5+0.5·tanh(|u|/d−1)) instead of the hard threshold in Eq. 2) and the pressure-width map uses a non-quadratic form (e.g., a higher-order polynomial or the original saturating exponential). If AK-MPC's tracking MAE gap over baselines shrinks below 2x under this mismatched structure, the current 5-8x gap is substantially an artifact of model-plant structural identity rather than a robust control advantage.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The surrogate plant defines delivered pressure through D_θ(u) = sgn(u)g_θ[|u|−d_θ]₊ (Eq. 2). The controller's probe fit (Eq. 8) estimates exactly this parametric family: (d̂, ĝ) = argmin Σ(Δp_eff − sgn(u)g[|u|−d]₊)². On the surrogate, the probe recovers near-true deadband parameters (Lemma 1 bounds the error only by probe noise and calibration residual), while baselines lacking the deadband model cannot. This gives AK-MPC a structural advantage by construction: the 5-8x gap over static inverse, PI, and ARX baselines is measured on a plant whose primary nonlinearity the controller was explicitly designed to model. The PCM ablation (0.0168 mm gain) does not isolate this effect because the no-PCM arm still uses the deadband-aware Koopman structure. The probe-fitted ARX baseline (0.3956 mm, worst performer) uses the same probe data but without the deadband parametric model, confirming that the deadband structure — not the memory or the Koopman lifting — drives most of the gap. The single physical trace (T12, Fig. 8) is acknowledged as a sanity check, not a statistical benchmark (§VII). On a real pneumatic valve, deadband involves stiction, continuous transitions, and pressure-dependent gains that do not follow the clean threshold form, so the structural match that benefits AK-MPC on the surrogate would be absent.","agreement_with_reader":"agree"},"referee_report":{"model":"glm-5.2","summary":"The paper proposes probe-conditioned memory (PCM) for actuator-deadband-aware Koopman MPC (AK-MPC) applied to industrial sealing/dispensing. The core idea is that during recipe-change commissioning, a short 16-move probe estimates actuator deadband, gain, lag, and hysteresis; this descriptor retrieves a nearby historical context that initializes a Koopman predictor and MPC prior. A local low-rank correction adapts the predictor online. The method is evaluated on a calibrated digital twin derived from a physical gluing cell, using 12 target conditions and 5 seeds (60 paired cases) at a 1.00 mm target bead width. AK-MPC achieves 0.0487 mm tracking MAE versus 0.2492-0.3956 mm for calibration-only, PI, RLS-ARX, and probe-fitted ARX baselines. Ablations isolate the PCM contribution (0.0168 mm gain) and online update contribution (0.0671 mm gain). Propositions 1-3 and Theorems 1-2 provide formal bounds on root-only indistinguishability, probe estimation error, predictor mismatch, and finite-horizon cost degradation.","tokens_in":18854,"tokens_out":1827,"duration_ms":180744,"significance":"The paper addresses a practically important problem: fast controller commissioning after recipe changes in dispensing cells where actuator deadband causes delivered-pressure loss invisible to static calibration. The PCM concept—combining static calibration, probe-inferred actuator state, predictor warm start, and fallback into a single retrievable record—is a reasonable and novel contribution to industrial informatics. The formal framework (Propositions 1-2 on root-only indistinguishability, Theorem 1 on retrieved-predictor bounds, Theorem 2 on cost degradation) is internally consistent and provides useful structural insight. The ablation design properly separates PCM from online correction. The inclusion of a physical-cell trace (Fig. 8) alongside the digital-twin benchmark, and the honest framing of the physical trace as a sanity check rather than a statistical benchmark, are commendable. The information-access rules for baselines are clearly stated.","major_comments":[{"comment":"The surrogate plant's deadband model (Eq. 2: D_theta(u) = sgn(u)g_theta[|u|-d_theta]+) is structurally identical to the controller's probe-estimation model (Eq. 8: (d_hat, g_hat) = argmin sum(Delta_p_eff - sgn(u)g[|u|-d]+)^2). This gives AK-MPC a structural advantage by construction on the digital twin: the probe recovers near-true deadband parameters on a plant whose primary nonlinearity the controller was explicitly designed to model. The 5-8x gap over baselines (Table IV) thus partly measures model-plant match rather than genuine control transfer. The authors should explicitly acknowledge this structural match and discuss whether real pneumatic valve deadband (which involves stiction, continuous transitions, pressure-dependent gains) follows the clean threshold form of Eq. 2. The PCM ablation (0.0168 mm, Table VI) does not fully isolate this concern because the no-PCM arm still uses a","section":null},{"comment":"The single physical-cell trace (T12/Pattern-111, Fig. 8) is acknowledged as a 'deployment sanity check rather than a 60-case physical benchmark' (§VII). While this honest framing is appreciated, the central performance claim (0.0487 mm MAE, 5-8x improvement) rests entirely on the self-constructed digital twin. The 12 target conditions (Table III) and 16 PCM source contexts are both generated from the same surrogate. The paper would be substantially strengthened by either (a) running at least 3-5 physical target conditions with paired baselines, or (b) adding a surrogate mismatch sensitivity analysis where the plant deadband structure differs from Eq. 2 (e.g., smooth sigmoid instead of hard threshold) to test robustness of the gap.","section":null},{"comment":"The probe-fitted ARX baseline (0.3956 mm, worst performer in Table IV) uses the same probe data as AK-MPC but without the deadband parametric model. This suggests that the deadband structure—not the memory or Koopman lifting—drives most of the performance gap. However, this baseline is described only briefly in §II and §V. A more detailed comparison isolating the deadband model from the Koopman lifting would clarify the contribution of each component. Specifically, an ablation using a deadband-aware but non-Koopman controller (e.g., deadband-compensated PI) would help separate these effects, as the authors themselves note in §VII.","section":null},{"comment":"Theorem 1 (Eq. 13) bounds the initial predictor mismatch as L_K(||xi*-xi_hat*|| + ||xi_hat*-xi_i*||) + epsilon_K. The covering radius term ||xi_hat*-xi_i*|| depends on the source library (16 contexts) adequately covering the target descriptor space. With only 16 source contexts and 12 targets, the retrieval distance distribution is not reported. The paper should report the mean and worst-case retrieval distances delta_i*(*) across the 60 cases to verify that the source library provides meaningful coverage rather than relying on a small set that may be co-tuned with the targets.","section":null}],"minor_comments":[{"comment":"Eq. (1): the term b_{theta,e} is described as 'episode-level drift' but the subscript 'e' is not explicitly defined. Clarify whether 'e' indexes episodes or is a parameter label.","section":null},{"comment":"Table I: the 'Signal-interface and probe tests' row lists 'private traces, public summaries' which is understandable for confidentiality, but the paper would benefit from at least a summary table of probe-response statistics (e.g., observed deadband range, gain range) across the physical cell.","section":null},{"comment":"Fig. 2: the calibration fit w(p) = alpha - beta*exp(-gamma*p) uses a saturating exponential, but the controller uses a local polynomial surrogate (§II.A). The relationship between these two forms on the compact operating interval should be stated more explicitly, including the polynomial degree and approximation residual.","section":null},{"comment":"The notation d_theta (scalar deadband) vs. d_s_i, d_m_i, d_q_i (descriptors) is flagged by the authors as distinct (§II), but the similarity of symbols may cause confusion. Consider using a different symbol for the scalar deadband (e.g., delta_theta).","section":null},{"comment":"§V: the scaled pressure domain [300, 650] and command bounds [-15, 15] are introduced without explicit conversion to physical units (bar). The relationship to the [0.18, 1.25] bar range in Table I should be stated.","section":null},{"comment":"Table VII: AK-MPC has 52.1% out-of-band rate despite 0.0487 mm MAE. This seems high relative to the reporting band tau_w = 0.035 mm. Clarify whether out-of-band counts any single step exceeding tau_w or a sustained violation, and discuss whether 52.1% is acceptable for the application.","section":null},{"comment":"References [29], [30] appear in the introduction but are numbered after [27], [28] which appear later in the text. Check reference ordering.","section":null},{"comment":"Algorithm 1, line 11: the candidate sequence generation includes 'sampled feasible perturbations' but the sampling distribution is not specified. Clarify how these perturbations are generated.","section":null}],"recommendation":"major_revision","confidential_remarks":"The structural match between the surrogate deadband model (Eq. 2) and the controller's probe-estimation model (Eq. 8) is the most significant concern. It does not invalidate the contribution—the PCM concept and formal framework are sound—but it means the headline 5-8x gap is an optimistic upper bound on real-world improvement. The 0.0168 mm PCM ablation gain is more credible as a transferable result but is modest. The paper is appropriate for this venue if the authors add (1) explicit acknowledgment of the structural match, (2) a robustness test with mismatched deadband structure, and (3) retrieval distance statistics. The single physical trace is insufficient as standalone validation but acceptable as supporting evidence alongside a more thorough surrogate robustness study. The novelty is adequate: PCM as a retrievable record combining calibration, probe, predictor, and controller priors is a reasonable contribution to fast commissioning literature."},"author_rebuttal":null,"desk_editor":{"model":"glm-5.2","letter":"The headline: the 5-8x improvement over baselines is measured on a self-constructed surrogate whose deadband model (Eq. 2) is structurally identical to the controller's probe-estimation model (Eq. 8). That's the one thing you need to know before taking the numbers at face value. The probe-conditioned memory concept itself is legitimate and new, but the magnitude of the gap is inflated by the structural match between plant and controller model. The stress-test note on this point lands squarely. The surrogate uses D_θ(u) = sgn(u)g_θ[|u|−d_θ]₊ and the probe fits exactly that parametric family. On this plant, the probe recovers near-true deadband parameters by construction, while baselines lacking the deadband model cannot. The probe-fitted ARX baseline (0.3956 mm, worst performer) uses the same probe data but without the deadband parametric form, which confirms that the deadband structure—not the memory or the Koopman lifting—drives most of the gap. The PCM ablation (0.0168 mm isolated gain) does not isolate this effect because the no-PCM arm still uses the deadband-aware Koopman structure. So the honest decomposition is: most of the improvement comes from modeling the right nonlinearity on a plant that shares that structure; a smaller but real increment comes from retrieving historical actuator context. What the paper does well: the ablation design is transparent, the scoping is honest (the authors explicitly call the physical trace a sanity check, not a benchmark), and the mathematical framework is internally consistent. Propositions 1-3 and Theorems 1-2 are stated with proofs or sketches. The PCM tuple (Eq. 11) combining pressure root, material/probe descriptors, actuator descriptor, Koopman predictor state, scaler, MPC prior, and fallback filter into a retrievable context is a genuine new combination not present in the cited literature. The physical-cell calibration data (22 samples, 0.010 mm RMSE) anchors the surrogate to real measurements, which is better than a purely synthetic study. The soft spot is proportionate to its load-bearing role: the central performance claim rests entirely on the surrogate, and the single physical trace (T12, Fig. 8) is insufficient to establish transfer. On a real pneumatic valve, deadband involves stiction, continuous transitions, and pressure-dependent gains that do not follow the clean threshold form. The structural advantage would be reduced or absent. The paper would move toward acceptance with either a multi-case physical benchmark or an independent surrogate whose deadband structure differs from the controller's model. Who this is for: control engineers working on dispensing and sealing cells who need fast commissioning after recipe changes. Readers interested in Koopman MPC, case-based retrieval, or deadband compensation will find the combination novel. The paper deserves a serious referee. The concept is real, the execution is careful within its scope, and the limitations are honestly stated. The referee should push hard on the surrogate-plant structural match and demand either physical validation or a mismatched-surrogate robustness check before accepting the headline numbers.","headline":"The 5-8x tracking improvement is measured on a digital twin whose deadband structure matches the controller's internal model by construction. The PCM concept is real and the ablations are honest, but the headline gap partly reflects model-plant match rather than genuine transfer.","tokens_in":19395,"tokens_out":751,"would_cite":true,"duration_ms":110450,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"glm-5.2","headline":"Short probe + stored memory cuts sealing error 5-8x","keywords":["model predictive control","Koopman operator","actuator deadband","probe-conditioned memory","industrial sealing","commissioning","digital twin","pneumatic valves"],"falsifier":"Run the twelve-target, five-seed benchmark on the physical sealing cell rather than the digital twin. If the AK-MPC tracking MAE on the physical cell exceeds the no-memory ablation's 0.0655 mm surrogate figure, or if the paired win rate against the probe-fitted ARX baseline drops below 50%, the central claim that probe-conditioned memory provides actuator-delivery information beyond what the probe alone gives would not hold in the physical regime.","tokens_in":18553,"feed_emoji":"🔧","tokens_out":1382,"duration_ms":141536,"temperature":0.7,"pith_summary":"In industrial sealing and dispensing cells, switching to a narrow bead recipe exposes a hidden problem: the pressure command is calibrated correctly, but the physical valve absorbs small corrective moves through its deadband, so delivered pressure changes only after a direction- and history-dependent threshold is crossed. Standard controllers that rely on the calibrated pressure setpoint, or that try to learn the new dynamics online from scratch, cannot see this delivered-pressure loss quickly enough during the short commissioning window. This paper introduces a controller architecture that combines three elements: a fixed sixteen-move probe executed before closed-loop operation begins, a stored library of historical operating cases that bundles together the pressure setpoint, actuator behavior, predictor state, and controller warm-start, and a Koopman lifted predictive controller that runs a constrained receding-horizon search with a feasibility fallback filter. The probe estimates the current actuator's deadband, gain, lag, and hysteresis; this estimate retrieves the nearest stored case from memory; the retrieved predictor is then corrected online with a low-rank update during the commissioning run. On a calibrated digital twin of a physical gluing cell, the full system achieves 0.0487 mm tracking mean absolute error on a 1.00 mm target, compared to 0.2492-0.3956 mm for calibration-only inverse, adaptive PI, online RLS-ARX, and probe-fitted ARX baselines. Ablation shows that removing the stored memory raises error to 0.0655 mm, and freezing the online correction raises it to 0.1158 mm, isolating the contribution of each component. The central claim is that historical runs become useful for a new recipe when they carry actuator-delivery information that a short probe can verify and refine, rather than carrying only a predictor checkpoint or a static calibration map.","feed_headline":"Short probe + stored memory cuts sealing error 5-8x","feed_subtitle":"A sixteen-move probe retrieves historical actuator behavior, letting a Koopman MPC skip online rediscovery of valve deadband during recipe切换","key_machinery":"The mechanism has three layers. First, a fixed probe sequence of sixteen pressure moves estimates four actuator quantities (deadband width, delivered-pressure gain, lag, and hysteresis) by fitting a deadband-gain map and a transient-response regression. Second, these estimates form a query vector that retrieves the nearest stored operating case from a library, where each case bundles a Koopman lifted predictor (finite matrix acting on a polynomial dictionary), a state scaler, an MPC prior, and a fallback filter. Third, the retrieved predictor is corrected online using a regularized finite-rank update (ridge regression with rank truncation) fit to the closed-loop transition buffer, while a可行性","core_discovery":"The paper identifies actuator deadband as the missing state variable in low-width sealing commissioning: two recipes can share the same calibrated pressure-width map and pressure root yet produce different bead widths because the valve's delivered-pressure response depends on direction and history. The paper shows that a sixteen-move probe can estimate the actuator-delivery descriptor before feedback begins, that this descriptor can retrieve a useful historical controller initialization from a stored library, and that the combination of retrieved memory with a low-rank online correction inside a constrained Koopman MPC reduces tracking error by a factor of five to eight over baselines that缺乏","pith_inferences":["If the deadband-gain map structure used in the surrogate does not capture more complex real-valve behaviors such as stiction, temperature-dependent gain variation, or asymmetric dynamic friction, the probe-estimation error bound in Lemma 1 may loosen enough that the retrieved memory provides a poor prior, collapsing the performance gap toward the no-memory ablation.","The method's value is inversely proportional to how much target data the commissioning window allows: in applications with long production runs where online identification eventually converges, the PCM contribution shrinks toward the 0.0168 mm isolated gain, while in very short runs it may dominate.","The sixteen-move probe length appears tuned to the specific deadband and hysteresis structure of pneumatic valves; processes with faster actuator dynamics or different nonlinearity orders may require a different probe length to achieve the same excitation rank."],"forward_implications":["The approach could reduce commissioning time for any pneumatic or hydraulic process where valve deadband creates a gap between commanded and delivered effort, including paint spraying, adhesive dispensing, and fluid metering.","The probe-conditioned memory structure suggests a general recipe for transfer in physical control: store not just a model checkpoint but the actuator-delivery state that a short target-side probe can verify, making historical data useful across recipes that share hardware but differ in operating regime.","The feasibility fallback filter, which screens MPC candidates against a prediction tube and falls back to a pressure-root or adaptive-PI move when the tube is violated, provides a safety architecture for deploying learned controllers on processes where constraint violation is costly.","The low-rank correction structure, motivated by the observation that pressure kinematics are identical across contexts while only the width-dynamics rows change, suggests that transfer in physically structured systems may require updating far fewer parameters than full model reidentification."],"fun_headline_variants":["Probe-conditioned memory cuts sealing tracking error 5-8x","Deadband-aware Koopman MPC reaches 0.049 mm MAE in industrial sealing","Sixteen-move probe retrieves historical actuator behavior for recipe commissioning","Actuator-deadband state variable key to low-width sealing control","Probe-conditioned memory in Koopman MPC cuts MAE from 0.25 to 0.05 mm"],"cache_read_input_tokens":0,"weakest_assumption_plain":"The entire statistical benchmark runs on a calibrated digital twin whose deadband-gain map and drift model are algebraic surrogates fitted to a physical cell. The only physical-cell closed-loop evidence is a single trace for one target condition. If the surrogate's actuator model does not match real valve behavior across the twelve target conditions, the five-to-eight-fold error reduction over baselines may not transfer to the physical plant.","fun_headline_variants_meta":{"raw":{"variants":["Probe-conditioned memory cuts sealing tracking error 5-8x","Deadband-aware Koopman MPC reaches 0.049 mm MAE in industrial sealing","Sixteen-move probe retrieves historical actuator behavior for recipe commissioning","Actuator-deadband state variable key to low-width sealing control","Probe-conditioned memory in Koopman MPC cuts MAE from 0.25 to 0.05 mm"]},"model":"glm-5.2","effort":"low","cost_usd":0.0,"raw_usage":{"total_tokens":796,"prompt_tokens":706,"completion_tokens":90,"prompt_tokens_details":null},"tokens_in":706,"tokens_out":90,"duration_ms":79677,"temperature":1.0,"reasoning_tokens":null,"cache_read_input_tokens":0,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-07-09T21:03:23.977117+00:00","model_set":{"reader":"glm-5.2"},"falsifier":"Run the twelve-target, five-seed benchmark on the physical sealing cell rather than the digital twin. If the AK-MPC tracking MAE on the physical cell exceeds the no-memory ablation's 0.0655 mm surrogate figure, or if the paired win rate against the probe-fitted ARX baseline drops below 50%, the central claim that probe-conditioned memory provides actuator-delivery information beyond what the probe alone gives would not hold in the physical regime.","supporting_citations":[],"review_version":1}