{"id":"f837061f-9382-49dd-b65c-8f307b49a8b7","arxiv_id":"2507.15677","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A data-selection-enhanced, data-driven MPC controller tracks joint trajectories on a three-segment cable-driven arm with 0.541 degree average error, outperforming a PID baseline.","lead":"This paper applies a data-driven model predictive control method to a cable-driven robotic arm, using only input-output data to plan motions. It adds a data selection step that picks the most relevant historical data, cutting solve time per step by about 80% and achieving average tracking error of 0.541 degrees.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The claimed accuracy gain over PID may stem from the inverse-model feedforward inside the MPC, not from the data-driven predictor; the paper lacks an ablation isolating the MPC's contribution.","rationale":"The reader's weakest_assumption concerns the unjustified extension of Willems' lemma to nonlinear systems. This is a valid theoretical gap, but it does not directly refute the empirical result; the closed-loop experiment could still succeed if the linear predictor happens to be locally accurate. The more damaging issue is experimental confounding: the MPC cost explicitly incorporates a reference input u_tar computed by an inverse MLP model (Sec. IV-E). A PID baseline without feedforward is the wrong comparator for isolating the contribution of the data-driven predictor. The paper even states in Sec. III that the inverse model is used to compute the reference input at each step. Because the cost weights R on the input tracking term and Q on the output tracking term are not analyzed, it is unknown whether the optimizer is following the inverse-model feedforward (which would make the G-matrix constraint (10b) nearly redundant) or genuinely using the data-driven prediction. The reported 80% speed improvement also omits the DSA selection overhead, but the accuracy confound is more central because it undermines the paper's quantitative headline. The proposed ablation (feedforward-only vs full MPC) is a single, decisive experiment: if feedforward-only matches the 0.541° error, the claim that the data-driven MPC with DSA achieves this accuracy is not established; if feedforward-only is significantly worse, the MPC predictor is doing real work. This is consistent with the reader's CONDITIONAL verdict, which should remain until such an ablation is provided.","tokens_in":11096,"tokens_out":10300,"duration_ms":114117,"concrete_test":"Run the same five-point tracking experiments with a feedforward-only controller: compute u_tar from the inverse MLP model and apply it directly through the motor PID, without running the MPC optimization or the G-matrix constraint. Compare the average tracking error against the reported 0.541°. If the feedforward-only error is close to 0.541° (e.g., within 0.1°), the data-driven MPC and DSA are not responsible for the improvement. Additionally, report the value of ||u - u_tar||_R relative to ||y - y_tar||_Q at the optimum to show whether the MPC is actually shaping the solution rather than just passing through the feedforward.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central experimental claim is that data-driven MPC with DSA achieves 0.541° average tracking error versus 1.418° for PID. However, the MPC cost (10a) includes ||u_k - u_tar_k||^2_R, where u_tar comes from an inverse MLP model (Sec. IV-E) trained on the same collected data. The PID baseline has no feedforward. If the inverse model is accurate, the feedforward alone can drive tracking, and the G-matrix constraint (10b) may be weakly active. The paper reports no ablation (e.g., inverse model plus motor PID, or MPC without the u_tar term), so the observed gap cannot be attributed to the data-driven predictor or DSA. The nonlinearity gap in Sec. IV-B compounds this: if Eq. (8) is a poor predictor, the optimizer can still produce u close to u_tar, making the data-driven MPC look successful while the predictor contributes little. Thus the central claim is not currently supported as stated.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a data-driven model predictive control (MPC) approach for a flexible cable-driven robotic arm (FCRA). The method constructs an implicit predictor from a Hankel matrix and a G-matrix based on Willems' Fundamental Lemma, integrates it as an equality constraint in an MPC optimization, and introduces a data selection algorithm (DSA) that picks one of several data blocks (collected under different payloads) to reduce per-step computation. An MLP inverse model is trained on the same data to provide reference motor inputs in the MPC cost. The method is evaluated in simulation for hyperparameter tuning and on a three-segment hardware platform for positioning repeatability, single-point tracking, and letter-drawing trajectory tracking. The reported results include an average tracking error of 0.541° versus 1.418° for PID and a per-step solve-time reduction from about 19 ms to 4 ms.","tokens_in":11306,"tokens_out":5610,"duration_ms":60607,"significance":"If the central performance claims hold, the paper would provide a practical demonstration of purely data-driven MPC for a cable-driven arm, with a data-selection mechanism that addresses real-time computational constraints. The hardware validation and the explicit study of hyperparameter effects are strengths. However, the unsupported extension of the LTI data-driven predictor to the nonlinear FCRA and the lack of an ablation separating the data-driven predictor from the inverse-model feedforward currently leave the main tracking-error claim unsubstantiated. The computational speedup claim also needs to account for selection overhead. The work is potentially significant for the robotics community, but these load-bearing issues require substantial revision.","major_comments":[{"comment":"The predictor in Eq. (8) is obtained by applying Willems' Fundamental Lemma, which is an LTI result, to data from a nonlinear cable-driven system. The paper states in Sec. IV-B that 'when the data volume N is large, the method can be effectively extended to nonlinear systems,' but no proof or formal argument is given, and the preceding sentence concedes that the linear-system inferences 'do not necessarily hold for nonlinear systems.' Because the MPC constraint (10b) uses this G-matrix as the plant model, the tracking results in Sec. V rest on an unsubstantiated modeling assumption. Please either provide a rigorous justification (e.g., an approximate behavioral model with a quantified error bound), cite a sound nonlinear extension from the literature, or reframe the method as a heuristic whose validity is established only by the closed-loop experiments, with the experiments designed to isolate the predictor's contribution.","section":"Sec. IV-B, Eq. (8)"},{"comment":"The MPC cost in Eq. (10a) includes the term ||u_k - u_tar_k||^2_R, where u_tar is computed from an inverse MLP model trained on the same collected data (Sec. IV-E), while the PID baseline in Fig. 8 has no such feedforward. Without an ablation that removes or isolates the u_tar term, or a feedforward-equipped PID baseline, the reported reduction from 1.418° to 0.541° in Table III cannot be attributed to the data-driven predictor or the DSA; the inverse-model feedforward alone could plausibly account for most of the improvement. Please add ablation experiments, for example: (i) PID with inverse-model feedforward, (ii) the proposed MPC with the u_tar term removed, and (iii) the proposed MPC without DSA using all data.","section":"Sec. IV-C/E and V-D, Eq. (10a), Table III"},{"comment":"Table III reports a single average error per target position with no trial count, standard deviation, or confidence interval. Given the measured positional repeatability in Table II (STD.DEV up to 2.640 mm, 3-sigma up to 10.380 mm), the claimed margin between 0.541° and 1.418° may lie within run-to-run variability. Report mean ± std over multiple independent repetitions per condition, and clearly state whether these tracking experiments are the same runs used for the positioning test or separate trials.","section":"Sec. V-D, Table III"},{"comment":"The reported solve-time reduction from 19 ms to 4 ms omits the time spent on the DSA selection step, which involves solving (11) for each of the M groups. In addition, the data-length notation is inconsistent: Sec. V-A states N = 120000 with M = 6, while Fig. 6 and inequality (13) use N = 401 and treat N as the block length. Please clarify the relationship among N, Ns, and M, and report end-to-end per-step wall-clock time including data selection, optimization, and communication.","section":"Sec. V-A/V-B, Sec. IV-D"}],"minor_comments":[{"comment":"The tracking-error metric in Eq. (12) is printed as '1/c Ta-1 sum' and appears to lack a normalization by Ta; it also yields squared units if interpreted literally. Please state the exact formula and specify whether Table III reports RMSE in degrees.","section":"Sec. V-B, Eq. (12)"},{"comment":"In Sec. IV-C the decision variable u_k is described as 'motor velocities,' while Sec. IV-A defines the model input as motor angles θ and Sec. IV-B uses input sequence u^d as motor angles. Please clarify whether the input to the Hankel/G-matrix model is the motor angle or the motor velocity, since this is essential for reproducing the MPC implementation.","section":"Sec. IV-C"},{"comment":"In Eq. (11), 'σ_L,th ⊂ σ' should read 'σ_L,th ∈ σ' (or the set σ should be defined); the current notation is mathematically imprecise.","section":"Sec. IV-D, Eq. (11)"},{"comment":"Table I contains a typo in the header 'Load W eight(kg)'; also, the row labels 'Payload1' through 'Payload3' should be explained in the caption or text.","section":"Sec. V-A, Table I"},{"comment":"The parameter listing 'Q = 10000I1X6, R = 70I1X9, S = 0.01I1X6, F = 0.01I1X9, P = 0.1I1X6' uses a nonstandard notation 'I1X6'; please use standard notation such as Q = 10000·I_6, R = 70·I_9, and so on, with the identity-matrix dimension made explicit.","section":"Sec. V-D, paragraph after Eq. (10)"}],"recommendation":"major_revision","confidential_remarks":"The manuscript builds directly on existing DeePC literature but does not fully position its novelty relative to nonlinear data-driven MPC or to prior applications of DeePC on robotic systems. The comparison baseline is weak; comparing against DeePC without data selection and against a feedforward-equipped PID would be far more informative. The confounding role of the MLP inverse-model feedforward in the MPC cost should be resolved before acceptance. The paper is within the scope of the journal if the experimental methodology is strengthened."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"The paper applies data-enabled predictive control to a three-segment cable-driven arm and adds a data-selection heuristic (DSA) that picks among pre-collected payload datasets by least-squares fit to the recent trajectory. That is a genuine new application, and the hardware results are real: 4 ms per solve, 2 mm repeatability, and visibly cleaner letter drawing than a PID baseline. The hyperparameter study, while simple, is a reasonable practical contribution.\n\nBut the stress-test concern holds up. The MPC cost in Eq. (10a) includes a heavy penalty on u_k - u_tar_k, where u_tar comes from an inverse MLP trained on the same collected data. The PID baseline has no feedforward. Without an ablation (inverse model + PID, or MPC without the u_tar term), the observed 0.541° versus 1.418° gap cannot be attributed to the data-driven predictor or the DSA. The G-matrix constraint (10b) might be barely active. The authors should report this before claiming the data-driven predictor is responsible.\n\nOther issues are real but minor. The nonlinearity extension is asserted in one sentence with no argument — fine for a heuristic application, but the paper leans on it. Table III has no error bars or trial counts, so the advantage could be within run-to-run variation. The notation flip between N=120000 and N=401 is confusing. The PID tuning is not described, so the baseline strength is unknown.\n\nThat said, the paper is not a washout. The DSA idea is simple and could be useful beyond this platform. The real-time implementation at 50 Hz is non-trivial. The repeatability numbers are reported with standard deviations, which is more than many robotics papers do.\n\nBottom line: it deserves a serious referee, but the verdict should be major revision. The authors need an ablation, repeated trials, and a clearer statement of what the data-driven predictor actually contributes. If the ablation shows the feedforward is doing the work, the paper still has value as a system integration study, but the central claim must be rewritten.","headline":"Useful application paper with a real FCRA testbed, but the main tracking comparison is confounded by an inverse-model feedforward term that is never ablated, so the central claim overstates what the data-driven predictor contributes.","tokens_in":11820,"tokens_out":1928,"would_cite":false,"duration_ms":22870,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper reports that a purely data-driven model predictive controller, built on input-output data and a data-selection algorithm, tracks joint trajectories on a real three-segment flexible cable-driven robotic arm with an average error…","keywords":["data-driven model predictive control","flexible cable-driven robotic arm","data selection algorithm","Hankel matrix","Fundamental Lemma","trajectory tracking","model-free control","quadratic programming"],"falsifier":"Run the same data-driven MPC on the same three-segment arm in an operating region absent from the training data—for example, with a 3 kg payload or a trajectory faster than any in the stored dataset—and compare the predicted outputs of Eq. (8) with actual joint encoder readings over the prediction horizon. If the G-matrix prediction diverges or the DSA selects a group whose estimated load differs sharply from the true load, the claimed model-free extension would be refuted.","tokens_in":10910,"feed_emoji":"🤖","tokens_out":6754,"duration_ms":64298,"temperature":0.7,"pith_summary":"The paper aims to establish that a flexible cable-driven robotic arm (FCRA) can be controlled accurately without any physical model: a model predictive controller built on stored input-output data can outperform a PID baseline on a real three-segment arm. The controller constructs an implicit predictor from Hankel matrices of past data, embeds it in a quadratic-programming optimization, and uses a data selection algorithm to pick the stored dataset that best matches the current operating condition. On the experimental platform, the data-driven MPC achieves an average tracking error of 0.541 degrees versus 1.418 degrees for PID, and the per-step solve time falls from about 19 ms to 4 ms, enabling 50 Hz closed-loop control. The paper also reports a positioning repeatability of roughly 2.070 mm over five targets repeated thirty times, and it studies how the data length and estimation horizon affect tracking error. If these results hold, they indicate that cable hysteresis, friction, and coupling can be absorbed by data-driven prediction rather than explicit modeling.","feed_headline":"Data-driven control cuts cable-arm tracking error to 0.54 degrees","feed_subtitle":"On a real three-segment arm, per-step solve time drops from about 19 ms to 4 ms, making model-free MPC practical at 50 Hz.","key_machinery":"The load-bearing object is the data-driven predictor $G := Y_f \\begin{bmatrix} U_p \\\\ Y_p \\\\ U_f \\end{bmatrix}^{\\dagger}$, a constant matrix that maps a short past input-output history plus a planned future input sequence to a predicted future output sequence. It is obtained by decomposing the Hankel matrices of a persistently exciting input and its output into past and future blocks, applying the Fundamental Lemma, and eliminating the coefficient vector $K$ through a least-squares pseudo-inverse. The MPC then solves a strictly convex quadratic program whose cost penalizes output error to the reference, terminal error, input deviation, and input/output rate changes, subject to the predictor equation, box constraints, and output-rate limits. The DSA selects which of $M$ stored data sequences is used to build $G$ at each step by minimizing $\\|K_j\\|$ against the current measurement, which makes the predictor locally matched to the current load.","core_discovery":"The central claim is that for a nonlinear, noisy cable-driven arm, a predictor derived from the Fundamental Lemma of behavioral systems theory for linear systems still works if it is built from a sufficiently large and well-chosen dataset. Specifically, the authors form the matrix $G := Y_f \\begin{bmatrix} U_p \\\\ Y_p \\\\ U_f \\end{bmatrix}^{\\dagger}$ — the future output block times the pseudo-inverse of the stacked past-input, past-output, and future-input blocks of the Hankel matrix — which yields the linear predictor $y = G\\, [u_{\\mathrm{ini}};\\, y_{\\mathrm{ini}};\\, u]$. This predictor replaces the physical model inside the MPC cost and constraints. A data selection algorithm (DSA) divides the stored data into groups, solves a small least-squares problem per group to test consistency with the current measurement, and retains the group with smallest coefficient norm; that group also selects the matching inverse reference model. The paper argues in Section IV-B that when the data volume $N$ is large, the least-squares form extends the linear theory to nonlinear systems, and the experimental tracking results are presented as validation of this extension.","pith_inferences":["A natural extension the authors do not state: the minimum $\\|K_j\\|$ used by the DSA is effectively an operating-point or load estimator, so the same computation could provide a real-time payload estimate for adaptive planning.","The $G$-matrix predictor built from the best-matching dataset behaves like a local linearization of the nonlinear arm at the current operating condition; this suggests the method works as a regime-switching linear predictor, which may explain why the LTI-based derivation carries over despite the nonlinearity.","The square-Hankel rule from inequality (13) might serve as a general heuristic for when data-driven MPC predictions stabilize, but that is an extrapolation from the paper's single-arm simulation, not a proven theorem."],"forward_implications":["If the reported numbers are correct, an FCRA can be position-controlled and trajectory-tracked without cable-specific friction or hysteresis models, which removes a major obstacle to deploying such arms in surgery or industry.","The data selection algorithm cuts the per-step computation from about 19 ms to 4 ms, meaning the same planning problem fits comfortably inside a 50 Hz control loop on standard hardware.","The hyperparameter study gives a practical sizing rule: keep the Hankel matrix at least square (data length $N$ satisfies inequality (13)) and increase the estimation horizon $n_{\\mathrm{ini}}$ when measurement noise is significant, since tracking error improves in both cases.","The repeatability experiment suggests that data-driven MPC holds roughly 2 mm positioning accuracy across repeated cycles, with the main drift source being gradual cable elongation rather than the controller itself.","Because the method relies only on motor and joint encoder signals, the same pipeline should transfer to other multi-segment cable-driven robots without re-deriving their kinematics or dynamic models."],"supporting_citations":[{"why":"Supplies the data-driven MPC formulation with stability and robustness guarantees that this paper adapts; the Hankel trajectory representation in Eq. (4) is taken from this line of work.","marker":"[7]"},{"why":"Provides the persistency-of-excitation result, the Fundamental Lemma, that justifies representing all system trajectories by the Hankel equation used to build the predictor.","marker":"[24]"},{"why":"The source of the least-squares elimination of the coefficient vector K, which turns Eq. (6) into the explicit G-matrix predictor used in Eq. (8).","marker":"[20]"},{"why":"The OSQP operator-splitting solver used to solve the strictly convex quadratic program at each MPC step, enabling the reported 4 ms solution time.","marker":"[25]"},{"why":"The PID control scheme used as the baseline; the paper's main tracking comparison (1.418 degrees vs 0.541 degrees) is against this method.","marker":"[6]"},{"why":"The Twist Snake cable-driven arm whose 4.9 mm positioning accuracy is the comparison point for the paper's 2.070 mm repeatability result.","marker":"[26]"}],"fun_headline_variants":["Model-free MPC cuts cable-arm tracking error to 0.54 degrees","Data-selection slashes solve time 80% while arm tracks at 0.54 deg","Cable-arm MPC uses data, not physics, to hit 0.54-degree tracking","Real-time data-driven MPC takes 4 ms steps and cuts arm error to 0.54 degrees","Flexible arm control: data-selected MPC beats PID tracking by 2.6x"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that the Fundamental Lemma's exact representation of linear time-invariant systems remains a valid approximate predictor for the nonlinear, noisy cable-driven arm once the dataset is large; the paper states this extension in Section IV-B without a formal proof, so the tracking and speed results depend on it.","fun_headline_variants_meta":{"raw":{"variants":["Model-free MPC cuts cable-arm tracking error to 0.54 degrees","Data-selection slashes solve time 80% while arm tracks at 0.54 deg","Cable-arm MPC uses data, not physics, to hit 0.54-degree tracking","Real-time data-driven MPC takes 4 ms steps and cuts arm error to 0.54 degrees","Flexible arm control: data-selected MPC beats PID tracking by 2.6x"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.001368,"raw_usage":{"total_tokens":5581,"prompt_tokens":1012,"completion_tokens":4569,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":628,"completion_tokens_details":{"reasoning_tokens":4454}},"tokens_in":628,"tokens_out":4569,"duration_ms":33686,"temperature":1.0,"reasoning_tokens":4454,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T15:25:53.782002+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run the same data-driven MPC on the same three-segment arm in an operating region absent from the training data—for example, with a 3 kg payload or a trajectory faster than any in the stored dataset—and compare the predicted outputs of Eq. (8) with actual joint encoder readings over the prediction horizon. If the G-matrix prediction diverges or the DSA selects a group whose estimated load differs sharply from the true load, the claimed model-free extension would be refuted.","supporting_citations":[{"cited_title":"Data-enabled predictive control for grid-connected power converters,","cited_arxiv_id":null,"evidence_quote":"The source of the least-squares elimination of the coefficient vector K, which turns Eq. (6) into the explicit G-matrix predictor used in Eq. (8)."},{"cited_title":"Design, kinematics and control of a modular cable-driven manipulator for fine manipulation,","cited_arxiv_id":null,"evidence_quote":"The PID control scheme used as the baseline; the paper's main tracking comparison (1.418 degrees vs 0.541 degrees) is against this method."},{"cited_title":"Twist snake: Plastic table-top cable- driven robotic arm with all motors located at the base link,","cited_arxiv_id":null,"evidence_quote":"The Twist Snake cable-driven arm whose 4.9 mm positioning accuracy is the comparison point for the paper's 2.070 mm repeatability result."}],"review_version":1}