Pith. sign in

REVIEW 4 major objections 8 minor 30 references

Short probe + stored memory cuts sealing error 5-8x

Reviewed by Pith at T0; open to challenge. T0 means a machine referee read the full paper against a public rubric. the ladder, T0–T4 →

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗

T0 review · glm-5.2

2026-07-09 21:03 UTC pith:PSCL5E2W

load-bearing objection 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. the 4 major comments →

arxiv 2607.07043 v1 pith:PSCL5E2W submitted 2026-07-08 eess.SY cs.SY

Probe-Conditioned Memory for Actuator-Deadband-Aware Koopman MPC in Industrial Sealing

Yue Wu This is my paper
classification eess.SY cs.SY
keywords model predictive controlKoopman operatoractuator deadbandprobe-conditioned memoryindustrial sealingcommissioningdigital twinpneumatic valves
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

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.

Core claim

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

What carries the argument

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

Load-bearing premise

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.

What would settle it

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.

Watch this falsifier — get emailed when new claim-graph text bears on it.

If this is right

  • 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.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • 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.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit.

Referee Report

4 major / 8 minor

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.

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 (4)
  1. 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
  2. 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.
  3. 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.
  4. 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.
minor comments (8)
  1. 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.
  2. 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.
  3. 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.
  4. 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).
  5. §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.
  6. 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.
  7. References [29], [30] appear in the introduction but are numbered after [27], [28] which appear later in the text. Check reference ordering.
  8. 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.

Circularity Check

0 steps flagged

No significant circularity; the structural match between surrogate deadband model and probe estimator is a simulation-validity concern, not a circular derivation.

full rationale

The paper's derivation chain (Propositions 1–3, Lemmas 1–2, Theorems 1–2) consists of standard mathematical results — mean-value theorem bounds, Lipschitz error propagation, recursive prediction-tube bounds, and finite-horizon cost degradation — all proven in-text from stated assumptions without self-citation chains. The author does not cite their own prior work as load-bearing. The one structural concern raised by the skeptic — that the surrogate plant's deadband map D_θ(u) = sgn(u)g_θ[|u|−d_θ]₊ (Eq. 2) matches the controller's probe-estimation model (Eq. 8) — is a modeling-assumption and simulation-to-reality-validity issue, not circularity. The paper does not present the deadband parameters as a 'prediction' or 'first-principles result' that is then validated against the same surrogate. Instead, it explicitly frames the evaluation as a 'calibrated digital-twin study' (§II), acknowledges the single physical trace as a 'sanity check rather than a 60-case physical benchmark' (§VII), and states 'the statistical benchmark remains the calibrated surrogate study' (§VIII). The 5–8× performance gap is an empirical result on the surrogate, not a derived theorem. The PCM ablation (0.0168 mm gain) isolates the memory-retrieval contribution by design; the no-PCM arm retaining the deadband structure is the intended comparison, not a hidden circularity. The structural match between plant and estimator is standard system-identification practice (design the estimator to match expected plant structure), and the paper is transparent about this being a surrogate study. Score 2 reflects that the structural match is a legitimate validity concern but does not constitute circular derivation of results from inputs by construction.

Axiom & Free-Parameter Ledger

8 free parameters · 4 axioms · 2 invented entities

The free parameters are mostly calibration constants and controller hyperparameters. The calibration parameters (alpha, beta, gamma) are fitted to physical data, which is appropriate. The controller constants (adapter rank, retrieval weight, MPC weights) are chosen by the authors without systematic tuning justification, though auxiliary sensitivity runs over adapter rank (2,4,8), retrieval weight (0.3,0.6,0.9), and probe length (8,16,24) are mentioned. The most load-bearing axiom is surrogate fidelity, which is only partially validated by one physical trace.

free parameters (8)
  • alpha (calibration) = 1.62
    Saturating exponential pressure-width fit parameter, fitted to 22 physical calibration samples
  • beta (calibration) = 1.22
    Saturating exponential pressure-width fit parameter, fitted to physical calibration data
  • gamma (calibration) = 1.45
    Saturating exponential pressure-width decay rate, fitted to physical calibration data
  • alpha (retrieval weight) = 0.6
    Weight balancing static vs dynamic descriptor in retrieval metric (Eq. 12), chosen by the authors
  • adapter rank r_delta = 4
    Rank of the low-rank correction in Eq. 19, chosen as a controller constant
  • lambda_delta (ridge regularization) = not stated
    Regularization parameter in the ridge regression update (Eq. 18), value not given in the paper
  • MPC weights q_w, q_p, r_u, r_du = not stated
    Cost function weights in Eq. 20, values not specified in the paper
  • tau_w (reporting band) = 0.035 mm
    Width-error band used for out-of-band reporting and tube filtering, set as benchmark parameter
axioms (4)
  • domain assumption The lifted one-step predictor is locally Lipschitz in the descriptor (Theorem 1)
    Invoked in Theorem 1 to bound initial predictor mismatch; standard for smooth Koopman operators but not proved for the specific finite dictionary used
  • domain assumption The calibration map is monotone on the operating interval with 0 < m_f <= |f'_b(p)|
    Invoked in Lemma 1 and Proposition 1; reasonable for pressure-width curves but assumed rather than verified across all 12 target conditions
  • ad hoc to paper The surrogate digital twin faithfully represents physical plant dynamics including deadband, hysteresis, lag, and drift
    The 60-case benchmark and all performance claims depend on this fidelity; only one physical trace (T12) is shown as validation
  • domain assumption The 16-action probe sequence excites all relevant deadband directions with sufficient margin
    Invoked in Lemma 1; the probe design is fixed but its adequacy across all target actuator modes is assumed
invented entities (2)
  • Probe-conditioned memory (PCM) context record independent evidence
    purpose: Stored operating case tuple linking pressure root, material/probe response, actuator descriptor, Koopman predictor, scaler, MPC prior, and fallback filter for retrieval-based controller initialization
    The PCM tuple (Eq. 11) is a new data structure, but its components are standard control objects. Its effectiveness is tested via ablation (Table VI) showing 0.0168 mm gain, providing falsifiable evidence within the surrogate benchmark.
  • Actuator descriptor z_a = [d, g, l, h]^T independent evidence
    purpose: Four-element vector capturing deadband, delivered-pressure gain, lag, and hysteresis for retrieval matching
    Derived from probe data via Eqs. 8-9; the estimation error is bounded in Lemma 1. The descriptor is a re-parameterization of known actuator nonlinearity concepts rather than a new physical entity.

pith-pipeline@v1.1.0-glm · 21009 in / 3241 out tokens · 318622 ms · 2026-07-09T21:03:23.977117+00:00 · methodology

0 comments
read the original abstract

Industrial sealing and dispensing cells often reuse a pressure chain, nozzle, substrate path, and vision interface across product recipes. For a narrow bead recipe, however, a calibrated static pressure can remain correct while small corrective moves are absorbed by actuator deadband; delivered pressure changes only after a direction- and history-dependent threshold is crossed. Commissioning is defined here as the target setup and retuning interval after such a recipe change. A physical gluing and dispensing cell provides pressure-to-width calibration, a fixed probing sequence, signal-interface limits, residual scales, and actuator bounds. The controller comparison is then run on an anonymized digital twin calibrated from those measurements. The actuator-deadband-aware Koopman model predictive controller (AK-MPC) initializes from probe-conditioned memory (PCM) that links the pressure setpoint to probe-inferred actuator behavior, a predictor, a controller prior, and a fallback filter. During commissioning, a sixteen-move probe selects a nearby historical case, fits the current pressure-width relation, updates a small local dynamic correction, and supplies a feasible receding-horizon pressure policy. In the main \(1.00\) mm benchmark, where delivered-pressure loss is visible in the probe, AK-MPC reaches 0.0487 mm tracking mean absolute error (MAE) over 60 paired cases; the calibration-only inverse, adaptive proportional-integral, online recursive-least-squares ARX, and probe-fitted ARX controllers range from 0.2492 to 0.3956 mm. This large gap reflects the full constrained Koopman-MPC and online-correction workflow. The isolated PCM contribution is measured by ablation: removing PCM raises the error to 0.0655 mm. In this regime, a short actuator characterization makes historical runs useful before much target data are available.

Figures

Figures reproduced from arXiv: 2607.07043 by Yue Wu.

Figure 1
Figure 1. Figure 1: Physical gluing/dispensing cell and surrogate-control signal flow. (a) Physical test cell with pneumatic pressure actuation, glue applicator nozzle, paper [PITH_FULL_IMAGE:figures/full_fig_p003_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: Physical pressure-width calibration for the displayed low-width recipe. [PITH_FULL_IMAGE:figures/full_fig_p004_2.png] view at source ↗
Figure 3
Figure 3. Figure 3: Plant-calibrated sealing surrogate and probe-conditioned memory layer [PITH_FULL_IMAGE:figures/full_fig_p004_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Fast-transfer lifecycle from target probing to local correction, feasible [PITH_FULL_IMAGE:figures/full_fig_p007_4.png] view at source ↗
Figure 5
Figure 5. Figure 5: Fixed 16-action probe and actuator descriptor construction for the [PITH_FULL_IMAGE:figures/full_fig_p008_5.png] view at source ↗
Figure 6
Figure 6. Figure 6: Target-condition design for the 12 1.00 mm stress-recipe targets. The heat map summarizes pressure-root location, actuator deadband, delivered￾pressure gain, lag, hysteresis, and drift variations used in the paired bench￾mark. follows the same trend. The baseline gap follows the delivered￾pressure loss seen in the fixed probe. Pairing by target and seed leaves the ordering unchanged. AK-MPC wins all 60 pai… view at source ↗
Figure 7
Figure 7. Figure 7: Tracking MAE in the main 1.00 mm actuator-deadband benchmark. AK-MPC gives the lowest controller error, and removing probe-conditioned memory increases the error. provides the corresponding constraint/runtime heat map. The probe-conditioned-memory controller also uses less action total variation than static inverse adaptive PI and online RLS￾ARX-MPC, consistent with its explicit model of deadband and deliv… view at source ↗
Figure 8
Figure 8. Figure 8: Closed-loop control performance of the deployed AK-MPC on the physical sealing cell under increasing episode drift for the combined-dynamics [PITH_FULL_IMAGE:figures/full_fig_p011_8.png] view at source ↗
Figure 10
Figure 10. Figure 10: Mechanism ablation results for actuator-deadband validation. Positive [PITH_FULL_IMAGE:figures/full_fig_p011_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Constraint and runtime heat map for the 1.00 mm actuator-deadband benchmark, with values colored by column-normalized scale [PITH_FULL_IMAGE:figures/full_fig_p011_11.png] view at source ↗

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Reference graph

Works this paper leans on

30 extracted references · 30 canonical work pages · 1 internal anchor

  1. [1]

    Visual guidance of a sealant dispensing robot for online detection of complex 3D-curve seams

    Hanh, Le Duc, Thien, Cao Viet. Visual guidance of a sealant dispensing robot for online detection of complex 3D-curve seams. International Journal on Interactive Design and Manufacturing. 16(4), 1525–1532, (2022). doi: 10.1007/s12008-022-00843-y

    work page doi:10.1007/s12008-022-00843-y 2022

  2. [2]

    Force and vision-based system for robotic sealing monitoring

    Pereira, Franco Rocha, Rodrigues, Caio Dimitrov, Souza, Hugo da Silva e, Neto, Jose Oliveira Cruz, Rocha, Matheus Chiaramonte, Barbosa, Gustavo Franco, Shiki, Sidney Bruce, Inoue, Roberto Santos. Force and vision-based system for robotic sealing monitoring. The International Journal of Advanced Manufacturing Technology. 126(1–2), 391–403, (2023). doi: 10....

    work page doi:10.1007/s00170-023-11110-z 2023

  3. [3]

    A multi-purpose robot perception system enabling closed-loop control for zero defect manufacturing in gluing processes of large parts

    Prezas, Loukas, Arkouli, Zoi, Michalos, George, Angelakis, Panagiotis, Gkournelos, Christos, Makris, Sotiris. A multi-purpose robot perception system enabling closed-loop control for zero defect manufacturing in gluing processes of large parts. Robotics and Autonomous Systems. 181, 104778, (2024). doi: 10.1016/j.robot.2024.104778

    work page doi:10.1016/j.robot.2024.104778 2024

  4. [4]

    Digital twin modeling of the robotic gluing system for predicting the quality of glue lines and optimizing gluing parameters

    Kang, Ruihao, Hu, Junshan, Li, Mingyu, Zhang, Qi, Su, Xingtao, Li, Zhengping, Tian, Wei. Digital twin modeling of the robotic gluing system for predicting the quality of glue lines and optimizing gluing parameters. Journal of Manufacturing Systems. 80, 1074–1092, (2025). doi: 10.1016/j.jmsy.2025.05.008

    work page doi:10.1016/j.jmsy.2025.05.008 2025

  5. [5]

    Automated human-in-the-loop commissioning of in- dustrial processes: A glue dispensing case study

    Taets, Jeroen, Jordens, Jeroen, Lefebvre, Tom, Ostyn, Frederik, Creve- coeur, Guillaume. Automated human-in-the-loop commissioning of in- dustrial processes: A glue dispensing case study. The International Journal of Advanced Manufacturing Technology. 136(11–12), 4991– 5005, (2025). doi: 10.1007/s00170-025-15121-w

    work page doi:10.1007/s00170-025-15121-w 2025

  6. [6]

    Data tech- nologies in robotic sealing: A systematic review on inspection and quality control

    Sarti, Giovana, Kawano, Frederico Leoni Franco, Shiki, Sidney Bruce, Malavolta, Alexandre Tacito, Barbosa, Gustavo Franco. Data tech- nologies in robotic sealing: A systematic review on inspection and quality control. The International Journal of Advanced Manufacturing Technology. (2026). doi: 10.1007/s00170-026-18219-x

    work page doi:10.1007/s00170-026-18219-x 2026

  7. [7]

    Experimental identification of the dead zone in proportional directional pneumatic valves

    Valdiero, Antonio Carlos, Bavaresco, Delair, Andrighetto, Pedro Luis. Experimental identification of the dead zone in proportional directional pneumatic valves. International Journal of Fluid Power. 9(1), 27–33, (2008). doi: 10.1080/14399776.2008.10781294

    work page doi:10.1080/14399776.2008.10781294 2008

  8. [8]

    Tian, Yingjun, Su, Renbo, Wang, Xilong, Altin, Nur Banu, Fang, Guoxin, Wang, Charlie C. L.. OpenPneu: Compact platform for pneu- matic actuation with multi-channels. 2023 IEEE/ASME International Conference on Advanced Intelligent Mechatronics, 765–770, (2023). doi: 10.1109/AIM46323.2023.10196254

    work page doi:10.1109/aim46323.2023.10196254 2023

  9. [9]

    BiPneu: Design and control of a bipolar-pressure pneumatic system for soft robots

    Mei, Yu, Zhou, Xinyu, Naik, Vedant, Gao, Alan, Tan, Xiaobo. BiPneu: Design and control of a bipolar-pressure pneumatic system for soft robots. IEEE/ASME Transactions on Mechatronics, early access, (2026). doi: 10.1109/TMECH.2026.3693622

    work page doi:10.1109/tmech.2026.3693622 2026

  10. [10]

    Hysteresis compensation of flexible continuum ma- nipulator using RGBD sensing and temporal convolutional network

    Park, Junhyun, Jang, Seonghyeok, Park, Hyojae, Bae, Seongjun, Hwang, Minho. Hysteresis compensation of flexible continuum ma- nipulator using RGBD sensing and temporal convolutional network. IEEE Robotics and Automation Letters. 9(7), 6091–6098, (2024). doi: 10.1109/LRA.2024.3398501

    work page doi:10.1109/lra.2024.3398501 2024

  11. [11]

    Control pneumatic soft bending actuator with feedforward hysteresis compensation by pneumatic physical reservoir computing

    Shen, Junyi, Miyazaki, Tetsuro, Kawashima, Kenji. Control pneumatic soft bending actuator with feedforward hysteresis compensation by pneumatic physical reservoir computing. IEEE Robotics and Automation Letters. 10(2), 1664–1671, (2025). doi: 10.1109/LRA.2024.3523229

    work page doi:10.1109/lra.2024.3523229 2025

  12. [12]

    Model Pre- dictive Control: Theory, Computation, and Design

    Rawlings, James B., Mayne, David Q., Diehl, Moritz M.. Model Pre- dictive Control: Theory, Computation, and Design. Nob Hill Publishing. (2017)

    work page 2017

  13. [13]

    Joe, Badgwell, Thomas A

    Qin, S. Joe, Badgwell, Thomas A.. A Survey of Industrial Model Predictive Control Technology. Control Engineering Practice. 11(7), 733–764, (2003). doi: 10.1016/S0967-0661(02)00186-7

    work page doi:10.1016/s0967-0661(02)00186-7 2003

  14. [14]

    Online learning-based predictive control of crystallization processes under batch-to-batch parametric drift

    Zheng, Yingzhe, Zhao, Tianyi, Wang, Xiaonan, Wu, Zhe. Online learning-based predictive control of crystallization processes under batch-to-batch parametric drift. AIChE Journal. 68(11), (2022). doi: 10.1002/aic.17815

    work page doi:10.1002/aic.17815 2022

  15. [15]

    Online machine learning modeling and predictive control of nonlinear systems with scheduled mode transitions

    Hu, Cheng, Cao, Yuan, Wu, Zhe. Online machine learning modeling and predictive control of nonlinear systems with scheduled mode transitions. AIChE Journal. 69(2), (2022). doi: 10.1002/aic.17882

    work page doi:10.1002/aic.17882 2022

  16. [16]

    Real-time adaptive sparse- identification-based predictive control of nonlinear processes

    Abdullah, Fahim, Christofides, Panagiotis D.. Real-time adaptive sparse- identification-based predictive control of nonlinear processes. Chem- ical Engineering Research and Design. 196, 750–769, (2023). doi: 10.1016/j.cherd.2023.07.011

    work page doi:10.1016/j.cherd.2023.07.011 2023

  17. [17]

    Machine learning-based predictive control using on-line model lineariza- tion: Application to an experimental electrochemical reactor

    Luo, Junwei, Citmaci, Berkay, Jang, Joon Baek, Abdullah, Fahim, Morales-Guio, Carlos G., Christofides, Panagiotis D.. Machine learning-based predictive control using on-line model lineariza- tion: Application to an experimental electrochemical reactor. Chem- ical Engineering Research and Design. 197, 721–737, (2023). doi: 10.1016/j.cherd.2023.08.017

    work page doi:10.1016/j.cherd.2023.08.017 2023

  18. [18]

    Model predictive control of switched nonlinear systems using online machine learning

    Hu, Cheng, Wu, Zhe. Model predictive control of switched nonlinear systems using online machine learning. Chemical Engineering Research and Design. 209, 221–236, (2024). doi: 10.1016/j.cherd.2024.08.001

    work page doi:10.1016/j.cherd.2024.08.001 2024

  19. [19]

    Process structure-based recurrent neu- ral network modeling for predictive control: A comparative study

    Alhajeri, Mohammed S., Luo, Junwei, Wu, Zhe, Albalawi, Fahad, Christofides, Panagiotis D.. Process structure-based recurrent neu- ral network modeling for predictive control: A comparative study. Chemical Engineering Research and Design. 179, 77–89, (2022). doi: 10.1016/j.cherd.2021.12.046

    work page doi:10.1016/j.cherd.2021.12.046 2022

  20. [20]

    A Data-Driven Approximation of the Koopman Operator: Extending Dynamic Mode Decomposition

    Williams, Matthew O., Kevrekidis, Ioannis G., Rowley, Clarence W.. A Data-Driven Approximation of the Koopman Operator: Extending Dynamic Mode Decomposition. Journal of Nonlinear Science. 25, 1307– 1346, (2015). doi: 10.1007/s00332-015-9258-5. PROBE-CONDITIONED MEMORY FOR ACTUATOR-DEADBAND-AW ARE KOOPMAN MPC 13

    work page internal anchor Pith review doi:10.1007/s00332-015-9258-5 2015

  21. [21]

    Linear Predictors for Nonlinear Dynamical Systems: Koopman Operator Meets Model Predictive Control

    Korda, Milan, Mezi ´c, Igor. Linear Predictors for Nonlinear Dynamical Systems: Koopman Operator Meets Model Predictive Control. Automat- ica. 93, 149–160, (2018). doi: 10.1016/j.automatica.2018.03.046

    work page doi:10.1016/j.automatica.2018.03.046 2018

  22. [22]

    Hybrid Koopman model predictive control of nonlinear systems using multiple EDMD models: An application to a batch pulp digester with feed fluctuation

    Son, Sang Hwan, Choi, Hyun-Kyu, Moon, Jiyoung, Kwon, Joseph Sang- Il. Hybrid Koopman model predictive control of nonlinear systems using multiple EDMD models: An application to a batch pulp digester with feed fluctuation. Control Engineering Practice. 118, 104956, (2022). doi: 10.1016/j.conengprac.2021.104956

    work page doi:10.1016/j.conengprac.2021.104956 2022

  23. [23]

    Koopman operator-based multi-model for predic- tive control

    Lawrynczuk, Maciej. Koopman operator-based multi-model for predic- tive control. Nonlinear Dynamics. 112(12), 9955–9982, (2024). doi: 10.1007/s11071-024-09615-7

    work page doi:10.1007/s11071-024-09615-7 2024

  24. [24]

    Experimental validation of deep Koopman MPC for real-time pasteurization unit control

    Valabek, Patrik, Horvathova, Michaela, Pannocchia, Gabriele, Klauco, Martin. Experimental validation of deep Koopman MPC for real-time pasteurization unit control. Control Engineering Practice. 167, 106663, (2026). doi: 10.1016/j.conengprac.2025.106663

    work page doi:10.1016/j.conengprac.2025.106663 2026

  25. [25]

    Case-Based Reasoning: Foundational Issues, Methodological Variations, and System Approaches

    Aamodt, Agnar, Plaza, Enric. Case-Based Reasoning: Foundational Issues, Methodological Variations, and System Approaches. AI Com- munications. 7(1), 39–59, (1994). doi: 10.3233/AIC-1994-7104

    work page doi:10.3233/aic-1994-7104 1994

  26. [26]

    A Survey on Transfer Learning

    Pan, Sinno Jialin, Yang, Qiang. A Survey on Transfer Learning. IEEE Transactions on Knowledge and Data Engineering. 22(10), 1345–1359, (2010). doi: 10.1109/TKDE.2009.191

    work page doi:10.1109/tkde.2009.191 2010

  27. [27]

    Modeling and predictive control of nonlinear processes using transfer learning method

    Xiao, Ming, Hu, Cheng, Wu, Zhe. Modeling and predictive control of nonlinear processes using transfer learning method. AIChE Journal. 69(7), (2023). doi: 10.1002/aic.18076

    work page doi:10.1002/aic.18076 2023

  28. [28]

    Privacy-preserving federated machine learn- ing modeling and predictive control of heterogeneous nonlinear sys- tems

    Xu, Zeyuan, Wu, Zhe. Privacy-preserving federated machine learn- ing modeling and predictive control of heterogeneous nonlinear sys- tems. Computers & Chemical Engineering. 187, 108749, (2024). doi: 10.1016/j.compchemeng.2024.108749

    work page doi:10.1016/j.compchemeng.2024.108749 2024

  29. [29]

    A digital twin-driven method for online quality control in process industry

    Zhu, Xiaoyang, Ji, Yangjian. A digital twin-driven method for online quality control in process industry. The International Journal of Ad- vanced Manufacturing Technology. 119(5–6), 3045–3064, (2022). doi: 10.1007/s00170-021-08369-5

    work page doi:10.1007/s00170-021-08369-5 2022

  30. [30]

    Dynamic transfer soft sensor for concept drift adaptation

    Zhang, Tianming, Yan, Gaowei, Ren, Mifeng, Cheng, Lan, Li, Rong, Xie, Gang. Dynamic transfer soft sensor for concept drift adaptation. Journal of Process Control. 123, 50–63, (2023). doi: 10.1016/j.jprocont.2023.01.012

    work page doi:10.1016/j.jprocont.2023.01.012 2023