{"id":"bbc38d9a-b25b-4e81-9cf9-c0e88dfcf688","arxiv_id":"2504.12428","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":7,"one_line_summary":"A learned predictor of the delayed state change reduces the soft robot arm's tracking error by up to 64% compared to a baseline sliding-mode controller.","lead":"This paper combines an online learning algorithm with a memory compression network to predict a soft robot arm's motion 0.14 seconds ahead, reducing tracking errors. It shows that a learned delay compensator lets a two-module soft robot follow circular paths more accurately, especially at high control gains, which matters for assistive care robots.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Reported 64% improvement may be attributable to a trivial linear predictor (x + d·x̂dot) rather than to the learned nonlinear Smith predictor; no such control baseline is reported.","rationale":"The reader's weakest assumption concerns fixed delay and the training target, and the reader's rationale focuses on reporting gaps such as exclusions, undisclosed hyperparameters, and missing code/data. I agree those are real issues, but the more load-bearing gap is the absence of a trivial linear predictor control condition. The proposed method's input includes estimated velocity and its target is a one-delay state difference, so a fixed predictor x + d·x̂dot is the natural minimal baseline for any delay-compensation claim. Without it, the experimental comparison cannot distinguish learned nonlinear delay compensation from simple phase-lead compensation. This is a concrete, testable omission: adding that baseline would either confirm that the learning component is responsible or show that a much simpler method obtains the same benefit. It does not require rejecting the paper; the statistical analysis and reported improvements appear internally consistent. The appropriate verdict remains conditional pending this control experiment and the reader's already-identified reporting improvements.","tokens_in":10318,"tokens_out":19340,"duration_ms":208112,"concrete_test":"Run the identical 60-second circular tracking protocol (spiral buildup plus constant-diameter phase) at low, medium, and high gains with a fixed predictor x_p(t) = x(t) + d·x̂dot(t), using the same observer velocity estimate and no KRLST or LDN. If the fixed predictor's stable-phase XY RMS error at high gain is within 5–10% of the LDN-3 value in Table I, then the 64% improvement over the baseline is not specific to the learned nonlinear Smith predictor and the central claim requires qualification.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the KRLST/LDN-based predictor reduces XY RMS tracking error by up to 64% over the STSMC baseline (Table I). However, the paper does not compare against the simplest possible delay-compensating predictor: x_p(t) = x(t) + d·x̂dot(t). This matters because the KRLST input in §IV-B explicitly includes the observer-estimated velocity x̂dot(t), and the training target in §IV-C is x(t) − x(t−d), which for smooth trajectories is approximately d·x̂dot(t). The learned mapping can therefore approximate a fixed linear predictor almost trivially, and the reported improvement could reflect generic phase-lead/delay compensation rather than the proposed nonlinear learned predictor. The experiments in Table I only compare the learning methods against the baseline STSMC without any predictor, so the specific contribution of the learning-based nonlinear component is not isolated. If the simple linear predictor achieves comparable tracking error, the paper's central claim about the learning-based approximation is not established, even though the numerical improvement over the baseline may be real.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a learning-based approximation of a nonlinear Smith predictor to compensate for a fixed 0.14-second input-output delay in a two-module soft robotic arm. The method combines a Legendre Delay Network (LDN) for compressing input history with a Kernel Recursive Least Squares Tracker (KRLST) for online nonlinear function approximation. The predicted pose is formed by adding the learned correction to the measured pose, and the result is used in the error calculation of a super-twisting sliding-mode controller (STSMC). Experiments on a physical robot show that the learning-based variants reduce XY RMS tracking error compared to the baseline STSMC, with reductions up to 42% at medium gains and 64% at high gains in the stable phase, with statistical significance reported via ANOVA and Tukey post-hoc tests. The paper also presents separate modeling errors showing that the learned predictor reduces the delay-period pose-change prediction error relative to a No-Prediction baseline.","tokens_in":10553,"tokens_out":5012,"duration_ms":51075,"significance":"If the central claim is established, the paper makes a useful engineering contribution: an online, computationally efficient delay-compensation layer that improves tracking of a soft robot without requiring an accurate analytical model. The strongest elements are the physical experiments with 20 runs per condition, the use of ANOVA and Tukey tests to support the main comparisons, and the direct measurement of prediction error (Table II) that quantifies the learned predictor's accuracy. The method's ability to maintain stable tracking at higher controller gains is practically relevant for assistive robotics. However, the paper does not isolate the nonlinear learned component from a much simpler linear predictor, and several experimental details needed for replication are missing; these issues must be addressed before the central claim can be accepted.","major_comments":[{"comment":"The paper does not compare the proposed learning-based predictor against a simple linear predictor of the form x_p(t) = x(t) + d * x_hat_dot(t). Since the KRLST input includes the observed pose and observer-estimated velocity (Section IV-B), and the training target is x(t) - x(t-d) (Section IV-C), which for smooth trajectories is approximately d * x_dot(t), the learned mapping could trivially approximate this linear predictor. Without a control experiment using that linear predictor, the reported 42-64% improvements in Table I cannot be attributed to the nonlinear learned Smith predictor; they may only demonstrate the benefit of generic phase-lead delay compensation. I request this baseline to be added and discussed.","section":"Section IV-C, Table I"},{"comment":"The paper states that KRLST parameters sigma^2, nu, lambda, and the dictionary size M were selected through a systematic tuning process, but it does not report the actual parameter values or ranges, only that adjustments of 20-50% were made around initial values. This omission prevents replication and makes it impossible to assess the sensitivity of the reported 64% improvement to hyperparameter choices. The specific values used in the final experiments should be listed.","section":"Section V-B"},{"comment":"The experimental protocol says that runs showing 'anomalous behavior due to sporadic computer lag or external disturbances' were excluded, and that 20 successful experiments were retained per condition. The number of excluded runs is not reported. Without this count, the possibility of selection bias in the retained runs cannot be evaluated. Please quantify the exclusions and state the exclusion criteria more precisely.","section":"Section V-C"},{"comment":"The derivation of the nonlinear Smith predictor contains an inconsistent approximation: the text states 'We also establish Xp(t-d) ≈ X(t), leveraging the fact that an ideal predictor at t-d would have predicted the state at time t.' For an ideal predictor, Xp(t-d) should be the actual state at t-d, not the state at t; the state at t is the predicted output at the end of the integration window, not the initial condition. Since the integral in Eq. (4) is never computed and the learning target uses the actual x(t) - x(t-d), this may be a conceptual error that does not affect the implementation, but it should be corrected or clarified because it is part of the method's motivation.","section":"Section III-D"}],"minor_comments":[{"comment":"The acronym for Kernel Recursive Least Squares Tracker is written as 'KRLST' in most places but as 'KRSLT' in the abstract and in Section II-C; please make the spelling consistent.","section":"Abstract and Section II-C"},{"comment":"The word 'analitically' should be 'analytically'.","section":"Section IV-C"},{"comment":"The sentence 'the improved modeling of LDN-3 during the transient phase may explain its slightly advantage at tracking' should read 'its slight advantage at tracking'.","section":"Section VI-B"},{"comment":"The definition of the No-Pred case as 'the difference between current and future states' is confusing, because the training target used in Section IV-C is a past pose difference x(t) - x(t-d). Clarify the time indexing used for the No-Pred error in Table II.","section":"Section V-C"},{"comment":"The tables report only the mean ± standard error. Reporting the standard deviation or confidence intervals for the RMS errors would help readers assess the dispersion across the 20 runs.","section":"Table I and Table II"}],"recommendation":"major_revision","confidential_remarks":"The missing linear-predictor baseline is the key technical concern; without it, the central claim is not yet established. The paper would be strengthened by reporting the actual KRLST hyperparameters and the number of excluded runs. I recommend major revision rather than rejection because the issues are addressable within the manuscript's scope and the experimental data appear to support a real improvement over the no-prediction baseline."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Quick take: this is a competent experimental paper that shows a meaningful reduction in tracking error on a two-module soft arm, but the central attribution—that the KRLST/LDN learned nonlinear Smith predictor is what gets you the gain—is not fully supported because there's no comparison against the trivial linear predictor x + d·ẋ̂. That's the one thing worth pushing on.\n\nWhat's genuinely here: the combination of LDN for input history compression and KRLST for online learning is clean, and the experiments are properly done. Twenty runs, ANOVA/Tukey, consistent error reductions across medium and high gains. The improvement at high gains (27.6 to ~9.9 mm) is substantial, and Table II confirms the learned predictor is much better than no prediction at estimating the delay-induced state change. The practical value for soft robot control is real.\n\nWhere it gets soft: the stress-test lands. The training target is x(t) - x(t-d), which for smooth trajectories is approximately d·ẋ(t). The input features include the observer-estimated velocity. So KRLST can approximate the fixed linear predictor d·ẋ̂ almost trivially. Without running the controller with that linear predictor as a baseline, you can't tell whether the learning is doing anything beyond phase-lead. That is a load-bearing omission because the paper's claimed novelty is learning the nonlinear SP integral. If the linear predictor gets the same tracking performance, the contribution shrinks to \"velocity feedforward works,\" which is not new. This needs to be fixed before I'd sign off.\n\nThe other issues are minor: KRLST hyperparameters (σ², ν, λ, M) are described vaguely in the tuning process, the number of excluded runs isn't quantified, and there's no code or data. Those are addressable. The fixed-delay assumption is fine for this setup.\n\nBottom line: worth sending to review, but the referee should insist on the linear-predictor control baseline and full parameter reporting. I'd accept, conditional on that.","headline":"Real experimental improvement, but a missing linear-predictor baseline undercuts the claimed nonlinear contribution.","tokens_in":11105,"tokens_out":2757,"would_cite":false,"duration_ms":29724,"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":"Learned compensation for a nonlinear Smith predictor reduces the XY tracking error of a two-module soft robot arm by up to 64% at high controller gains, allowing stable high-gain motion despite the arm's 0.14-second delay.","keywords":["soft robot control","Smith predictor","delay compensation","Kernel Recursive Least Squares","Legendre Delay Network","sliding mode control","online learning","assistive robotics"],"falsifier":"Keep the controller and gains fixed but add an extra, unaccounted communication latency (say 0.05 s) or hang a payload that changes the arm's effective delay; if the learned predictor's modeling error reverts to the No-Pred level and tracking error returns to the baseline, the delay-compensation claim is specific to the fixed 0.14 s delay. A cleaner direct test is to run the identical KRLST/LDN predictor on a simulated nonlinear system with known delay, where the true Smith integral $\\Delta x_p(t)$ can be computed exactly, and check whether $\\hat{y}(t)$ matches it rather than merely matching $x(t)-x(t-d)$.","tokens_in":10099,"feed_emoji":"🦾","tokens_out":7174,"duration_ms":66675,"temperature":0.7,"pith_summary":"The paper tries to establish that a learning-based approximation of a nonlinear Smith predictor can compensate the fixed 0.14-second input-output delay of a two-module soft robot arm, and that doing so lets the existing super-twisting sliding-mode controller run at higher gains without the oscillations that normally appear. The method trains a Kernel Recursive Least Squares Tracker online to reproduce the Smith predictor's integral term, using a Legendre Delay Network to compress the recent command history into three state variables. On a circular XY tracking task, the learned predictor reduced stable-phase RMS tracking error from 27.6 mm to 9.9 mm at high controller gains (a 64% reduction) and from 14.4 mm to 8.3 mm at medium gains, with the differences statistically significant at p < 0.0001. If correct, the approach is a way to get accurate, stable tracking from soft arms without an accurate analytical model of their nonlinear dynamics.","feed_headline":"Learning predictor cuts soft-robot tracking error 64%","feed_subtitle":"Online kernel learning approximates a Smith predictor, compensating the arm's 0.14s delay and stabilizing high-gain tracking.","key_machinery":"The load-bearing object is the learning predictor block: a KRLST regressor that takes as input the 6D pose, the observer-estimated 6D velocity, and either a 3-dimensional LDN memory state $m(t) \\in \\mathbb{R}^3$ or the raw actuator history $u[t-d,t]$ (3 or 7 steps), and outputs $\\hat{y}(t)$. Training uses shifted data: at time $t$ the input is $\\{X(t-d), u[t-2d,t-d]\\}$ and the target is the measured pose difference $x(t) - x(t-d)$, so the regressor learns the delayed state change that the Smith predictor integral would compute. The LDN is the linear system $\\theta \\dot{m}(t) = A m(t) + B u(t)$ with memory length $\\theta = d = 0.14$ s and $p = 3$, derived from a Pad\\'e approximation of the delay, which compresses seven samples of control history into three states. KRLST supplies online adaptation, a bounded dictionary of size 80, and a forgetting factor, so the learned compensation tracks changing robot dynamics.","core_discovery":"The central claim is that the integral term $\\Delta x_p(t) = \\int_{t-d}^{t} F_x(X_p(\\tau), u(\\tau))\\,d\\tau$ of a nonlinear Smith predictor, which is the predicted pose change over the delay, can be approximated online by a learned mapping $\\hat{y}(t) \\approx \\Delta x_p(t)$ whose inputs are the current complete state $X(t) = [x(t), \\dot{\\hat{x}}(t)]$ and a compressed history of recent control actions, and that substituting $x_p(t) = x(t) + \\hat{y}(t)$ into the sliding-mode error stabilizes high-gain tracking. The authors argue this is justified because the integral depends only on the control history $u[t-d,t]$ and the predicted state at time $t-d$, which is approximated by the current measured state. They report that all learning variants significantly outperform the baseline controller in the stable phase at medium and high gains, with the LDN-compressed variant statistically better at modeling the delay-period state change, and that both the 3-state compressed history and the full 7-step history give comparable tracking performance.","pith_inferences":["Beyond the paper: because the training target is the measured closed-loop state difference, the learned predictor may partly absorb whatever controller was active during data collection, so the reported gains might not transfer to new gains, trajectories, or users without retraining.","Beyond the paper: the closeness of Hist-3, Hist-7, and LDN-3 suggests the 0.14 s delay is short enough that the most recent three control steps carry most of the information, so the compression advantage would likely show up mainly for longer or variable delays.","Beyond the paper: with a fixed known delay of 0.14 s at 50 Hz, a simple tapped delay line could also supply the history; the learning block's real contribution may be absorbing model mismatch rather than compressing memory, a separation the current ablation cannot make."],"forward_implications":["At medium and high controller gains, the learning-based methods lowered stable-phase XY RMS error from 14.4 mm to 8.3 mm and from 27.6 mm to 9.9 mm, respectively, reductions of 42% and 64% relative to the sliding-mode baseline.","The LDN-compressed variant matched the raw 7-step history in stable tracking while using only 3 memory states, and was statistically better at modeling the delayed state change (2.35 vs 2.55 mm RMS at medium gain).","Because KRLST adapts online with a forgetting mechanism, the compensation does not require an accurate analytical model of the soft arm's nonlinear dynamics.","At low gains all methods perform similarly, so the demonstrated benefit is specifically the ability to raise controller gains without the oscillation that degrades the baseline.","The authors report qualitatively smoother motion in preliminary human-guided trajectories, pointing toward assistive tasks where reference trajectories are not fixed in advance."],"supporting_citations":[{"why":"Defines the Smith predictor structure whose integral term the paper learns to approximate.","marker":"[9]"},{"why":"Provides the Legendre Delay Network state-space memory used to compress the input history over the delay window.","marker":"[10]"},{"why":"Provides the Kernel Recursive Least Squares Tracker algorithm with its dictionary and forgetting mechanism for online regression.","marker":"[11]"},{"why":"Supplies the SINDYc-identified first-order model and the baseline super-twisting sliding-mode controller that the learned predictor augments.","marker":"[12]"}],"fun_headline_variants":["Learning-based delay compensation improves soft robot tracking","Online kernel learning approximates Smith predictor for soft robots","KRLST and LDN reduce soft robot tracking error","Soft robot arm tracks better with learned delay compensation","Efficient online learning stabilizes soft robot control"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The approach assumes the input-output delay is a fixed, known 0.14 seconds and that the measured pose difference $x(t)-x(t-d)$ is a faithful training target for the Smith predictor integral; if the delay varies with load or the measured state difference is corrupted by disturbances, the learned predictor compensates the wrong quantity.","fun_headline_variants_meta":{"raw":{"variants":["Learning-based delay compensation improves soft robot tracking","Online kernel learning approximates Smith predictor for soft robots","KRLST and LDN reduce soft robot tracking error","Soft robot arm tracks better with learned delay compensation","Efficient online learning stabilizes soft robot control"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000686,"raw_usage":{"total_tokens":3110,"prompt_tokens":946,"completion_tokens":2164,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":562,"completion_tokens_details":{"reasoning_tokens":2091}},"tokens_in":562,"tokens_out":2164,"duration_ms":18467,"temperature":1.0,"reasoning_tokens":2091,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-16T12:32:12.230066+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Keep the controller and gains fixed but add an extra, unaccounted communication latency (say 0.05 s) or hang a payload that changes the arm's effective delay; if the learned predictor's modeling error reverts to the No-Pred level and tracking error returns to the baseline, the delay-compensation claim is specific to the fixed 0.14 s delay. A cleaner direct test is to run the identical KRLST/LDN predictor on a simulated nonlinear system with known delay, where the true Smith integral $\\Delta x_p(t)$ can be computed exactly, and check whether $\\hat{y}(t)$ matches it rather than merely matching $x(t)-x(t-d)$.","supporting_citations":[{"cited_title":"A process-model control for line ar systems with delay,","cited_arxiv_id":null,"evidence_quote":"Defines the Smith predictor structure whose integral term the paper learns to approximate."},{"cited_title":"Legendre mem ory units: Continuous-time representation in recurrent neural netwo rks,","cited_arxiv_id":null,"evidence_quote":"Provides the Legendre Delay Network state-space memory used to compress the input history over the delay window."},{"cited_title":"Kernel recursive least-squares tracker for time-varying regress ion,","cited_arxiv_id":null,"evidence_quote":"Provides the Kernel Recursive Least Squares Tracker algorithm with its dictionary and forgetting mechanism for online regression."},{"cited_title":"Sliding-mode control of a soft robot based on data-driven s parse identification,","cited_arxiv_id":null,"evidence_quote":"Supplies the SINDYc-identified first-order model and the baseline super-twisting sliding-mode controller that the learned predictor augments."}],"review_version":1}