{"id":"8ec866f1-c6ab-46f9-8701-b3b2a4897ead","arxiv_id":"2411.13859","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A hybrid offline-online neural predictor with gradient-descent optimization provides real-time nonlinear model predictive control for a 22-ton hydraulic excavator, improving tracking and flow efficiency over PID.","lead":"This paper presents a data-driven model predictive controller that uses a neural network to predict a hydraulic excavator's motion several steps ahead, plus an online network that learns from prediction errors during operation. Tests on a 22-ton excavator show the controller tracking target motions and using hydraulic flow more efficiently than a standard PID controller.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Optimizer reliability is the soft spot: the cost is not proven convex and fixed 30-step GD has no convergence guarantee; a restart/iteration check would settle it.","rationale":"The real-machine demonstration is meaningful supporting evidence and I do not question the empirical results as reported. The central methodological claim, however, depends on an unproven hinge: that the NMPC cost can be treated as convex enough for a fixed 30-iteration gradient descent to provide a satisfactory control sequence. Eq. (10) is quadratic in the outputs of ReLU MLPs, making the cost piecewise quadratic and generally nonconvex in U, and Eq. (11) adds a discrete engine-speed constraint that plain gradient descent cannot optimize. The paper gives neither a proof nor an ablation showing the optimizer converges, and the online model H shifts the cost surface during operation. This is the same weakest assumption the reader identified, and it justifies a conditional rather than accept verdict. The proposed restart/convergence check would settle the concern directly if the authors release the trained models and data; without those artifacts, the check cannot be independently run, which should be acknowledged as a limitation.","tokens_in":19268,"tokens_out":7854,"duration_ms":85734,"concrete_test":"Run the Section VI.C heavy-load trajectory through the proposed NMPC with the paper's 30 GD iterations, and again with 300 iterations and with 10 random restarts per cycle, keeping all other settings fixed; report achieved cost J and tracking ARMSE for each variant. If either variant changes J or ARMSE by more than a few percent, the fixed-budget result is not converged and the claimed reliability of the optimizer is unsupported. The authors should also state explicitly how the discrete engine-speed constraint in Eq. (11) is handled; if an unspecified gear-selection rule was used, that rule must be part of the published method.","verdict_should_be":"UNCHANGED","load_bearing_attack":"Section IV.D and Algorithm 1 rest the central claim on a gradient-descent solver for a cost that is asserted, not shown, to be 'generally convex.' Eq. (10) is quadratic in the outputs of ReLU MLPs G and H, so as a function of U it is piecewise quadratic and generally nonconvex; moreover, Eq. (11) imposes the discrete constraint ω_engine ∈ {low, medium, high} and a 1 s switching rule, which vanilla GD (8) cannot directly optimize. No mixed-integer or rounding procedure is described, and no evidence is given that the fixed 30 iterations with the error-dependent learning rate (12) reach a satisfactory solution, especially while H is updated online each cycle. If the optimizer stalls in a poor local region or ignores the discrete gear decision, the optimized control sequence is not what the paper claims, and the reported 50 Hz performance is a heuristic outcome rather than a validated NMPC solution.","agreement_with_reader":"agree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a data-driven nonlinear model predictive control (NMPC) framework for a 22-ton hydraulic excavator. The dynamics model combines an offline single-shot multi-step prediction (SSMP) network built from LSTM and MLP with an online MLP that learns prediction mismatches caused by load or environmental interaction. The control sequence is obtained by gradient descent with an adaptive learning rate on a cost function that penalizes joint position error, velocity error, and engine speed, subject to input saturation, three discrete engine gear levels, and a one-second gear-switching constraint. The controller applies the average of the optimized control sequence rather than its first element. The authors validate the approach in simulation with a PID-like baseline and on the real excavator, reporting improved tracking and flow efficiency under load.","tokens_in":19370,"tokens_out":2466,"duration_ms":27074,"significance":"If the claims hold, the paper would make a useful engineering contribution: it demonstrates a full-scale industrial validation of a learning-based NMPC, proposes an interesting offline-online hybrid prediction architecture that avoids retraining the LSTM online, and gives practical computational timings showing that the scheme can run at 50 Hz on a laptop CPU. The real-machine experiment on a 22-ton excavator is a notable strength, as is the explicit treatment of energy efficiency via engine-speed optimization. However, the paper's central methodological claims—that the cost is convex and that fixed-iteration gradient descent reliably solves the optimization—are not justified, and the empirical evaluation lacks a clear train/test separation, error bars, and a comparison against another NMPC solver or baseline. These issues limit the strength of the generalizability claims until addressed.","major_comments":[{"comment":"The statement that the cost function is 'generally convex' is not supported. The cost in Eq. (10) is a quadratic function of the outputs of ReLU MLPs G and H, which are piecewise-linear functions of U; as a function of the control sequence, J is piecewise quadratic and generally nonconvex. The paper provides no proof, and the self-stated limitation in the Conclusions ('has yet to undergo detailed theoretical validation') is directly relevant. The reviewer should ask for either a proof for a restricted setting or, more practically, empirical evidence of optimizer reliability: report the distribution of the terminal cost over many NMPC cycles, test multiple random initializations or restarts, and compare against a standard nonlinear optimizer (e.g., SQP or IPOPT) on a subset of cycles. Without this, the 50 Hz real-time result is a heuristic outcome rather than a validated NMPC solution.","section":"§IV.D, Eq. (10)–(12), Algorithm 1"},{"comment":"The optimization problem in Eq. (7)–(11) contains the discrete constraint ω_engine ∈ {low, medium, high} and the 1 s switching constraint t_switch > 1 s, but the gradient descent update in Eq. (8) operates on continuous variables and no rounding, projection, or mixed-integer handling is described. A plain gradient step cannot move between discrete gear levels in a meaningful way, and the switching constraint is not included in the cost or in the gradient. The paper should explain how the gear decision is actually made in Algorithm 1, and how the discrete constraint is enforced during the 30 iterations. A concrete demonstration that the discrete decision is not merely post-hoc rounding would be needed to support the claimed multi-objective optimization.","section":"§IV.D, Eq. (11) and Eq. (8)"},{"comment":"The evaluation does not clearly separate training and test trajectories. The text says that random sinusoidal signals generate motion trajectories and that Tables II–IV report ARMSE on 'various random trajectories,' but it is not stated whether these trajectories were excluded from the offline training set or from the online model's update data. Without this separation, the reported ARMSE values may partly reflect in-sample fitting. The authors should specify the train/test split, report the number of evaluation trajectories, and include error bars or confidence intervals across repeated runs. This is important because the paper claims robustness and generalization, but Tables II–IV currently show single point estimates.","section":"§VI, Tables II–IV and Fig. 7"},{"comment":"The control comparison is only against a PID baseline (with dead-zone compensation). The PID controller cannot optimize engine speed, so the comparison conflates the benefit of having engine speed as an optimized control input with the benefit of the proposed NMPC formulation itself. To support the claim that the proposed NMPC is effective as an NMPC, the paper should compare against at least one standard NMPC baseline using the same cost function and constraints—for example, an SQP-based NMPC with the same predictive model, or a linearized-MPC variant. Without such a baseline, the experimental plots do not isolate the contribution of the proposed optimization method.","section":"§VI.C, Figs. 8–10"}],"minor_comments":[{"comment":"In Eq. (6), the derivatives ∂M/∂U appear in the update rules for wH and bH; these should presumably be ∂M/∂wH and ∂M/∂bH. Please correct the notation.","section":"§IV.C, Eq. (6)"},{"comment":"The sentence 'Conversely, when the positional error is minimal, the learning rate is decreased appropriately to prevent the risk of over-regulation signals' is repeated verbatim two sentences later. Please remove the duplicate.","section":"§IV.C, paragraph after Eq. (12)"},{"comment":"The offline MLP input dimension is stated as (128 + 4i) × 1, while the online MLP input is stated as (13h + 4i) × 1. Please clarify why the offline model uses the LSTM hidden state (128) rather than the raw history length h, and define all dimensions consistently in the notation table.","section":"§V.B"},{"comment":"The baseline in Table II is referred to as 'DBN [24]' in the experiment but the Conclusions describe the comparison as against 'standard MLP.' Please make the baseline identification consistent.","section":"§VI.A and Conclusions"},{"comment":"The y-axis label 'AMRSE' appears to be a typo for 'ARMSE'. Also, the figure caption says 'real-time AMRSE' while the text discusses ARMSE over a sliding window; please clarify the definition used in the plot.","section":"§VI.B, Fig. 7"},{"comment":"The sentence 'To enable remote control functionality, we use two laptops, referred to as transmission and control' is clear, but it is followed by 'The user datagram protocol (UDP) is designed to meet our communication needs.' The phrase 'is designed' should be 'was used' or 'is used' for clarity.","section":"§V.A"}],"recommendation":"major_revision","confidential_remarks":"The paper has a genuine real-machine validation that is rare in this area, and the hybrid offline-online prediction idea is worth publishing if the optimization and evaluation concerns are addressed. The main risk is that the optimizer, which is the core of the NMPC claim, is not demonstrated to be reliable for the nonconvex, mixed-integer problem it is asked to solve. The authors themselves acknowledge the absence of theoretical validation, which is appropriate, but the empirical side needs to close that gap with stronger evidence than the current plots. I would not reject on the basis of missing theory alone, but I would not accept without a concrete convergence study and a clear train/test protocol."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Worth a look if you care about learning-based control on real heavy machinery. The standout fact is that this controller ran on an actual 22-ton hydraulic excavator at 50 Hz, with measured computation times of 5–15 ms in Table VI. That alone separates it from most DD-NMPC papers, which stop at simulation. The energy-efficiency plots also show real reductions in oil overflow relative to a compensated PID baseline, which is the kind of evidence an industrial reader will care about.\n\nWhat is genuinely new is the specific combination: a single-shot LSTM-MLP that predicts multi-step state increments (not states directly), an online MLP that learns the residual mismatch between the offline model and reality, and a gradient-descent NMPC that uses an adaptive learning rate and outputs the average of the control sequence. Each component is known, but putting them together in this architecture and demonstrating it on a load-carrying excavator is a legitimate engineering contribution.\n\nThe paper does several things well. The data-collection strategy is sensible, mixing open-loop and closed-loop data with a safety monitor. The offline/online hybrid is pragmatic: rather than retraining the LSTM online, they train a small MLP on the prediction error, which is a reasonable way to handle load-induced dynamics changes. The computational complexity analysis is honest, with measured timings repeated 1000 times.\n\nThe soft spots are real but concentrated in the optimization section. Section IV.D says the cost function is \"generally convex\" without proof, and Eq. (10) is quadratic in the outputs of ReLU networks, so as a function of the control sequence it is piecewise quadratic and generally nonconvex. More concretely, the discrete gear constraint ω_engine in {low, medium, high} with a 1 s dwell time appears in Eq. (11), but the gradient update in Eq. (8) has no mechanism to handle discrete variables; no rounding or mixed-integer scheme is described. Fixed 30 iterations with no convergence check is also asserted, not justified. That is the load-bearing weakness, and the stress-test note correctly identifies it. It does not sink the paper because the real-machine results suggest the optimizer behaved acceptably in practice, but the paper should explain what it actually does about the discrete gears and ideally add a restart or iteration-termination check.\n\nAlso worth flagging: Tables II–IV do not clearly separate training and test trajectories, there are no error bars, and no code or data are provided. The conclusion honestly admits there is no detailed theoretical validation, which is good to see. These issues lower confidence but do not undermine the central empirical claim.\n\nWho is this for? People working on control of hydraulic construction equipment, and researchers looking for a data-driven NMPC case study with real hardware. It deserves a serious referee, but the major revision should focus on the optimizer: either prove convexity, handle the discrete gear decision explicitly, or present a fallback strategy. I would not desk-reject this.","headline":"A practical data-driven NMPC with real 22-ton excavator validation; the optimizer story is under-supported, but the engineering result is credible and worth refereeing.","tokens_in":19957,"tokens_out":1665,"would_cite":true,"duration_ms":20322,"reading_group":"yes","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"By combining an offline LSTM-MLP predictor with an online mismatch-correcting network, the paper claims a data-driven NMPC can control both motion and energy use of a 22-ton hydraulic excavator at 50 Hz.","keywords":["data-driven NMPC","nonlinear model predictive control","LSTM-MLP prediction","online learning","hydraulic excavator","real-time control","energy management","multi-step prediction"],"falsifier":"On a logged or simulated trajectory of the same excavator, rerun the NMPC optimization from several different initial control sequences, including nonzero starts, and compare the resulting commands; if the gradient descent settles into materially different control sequences, or tracking error diverges on one start, the fixed-iteration assumption is not reliable. A cheaper check is to record the cost function value over the 30 iterations and see whether it is still changing substantially at iteration 30.","tokens_in":18960,"feed_emoji":"🏗️","tokens_out":5814,"duration_ms":52948,"temperature":0.7,"pith_summary":"The paper claims that a data-driven nonlinear model predictive controller can take over the motion and power management of a heavy hydraulic robot, and it demonstrates this on a 22-ton excavator. Instead of building a physics-based model, the controller learns a neural predictor offline from collected motion data and couples it with a small online network that learns prediction mismatches when the machine picks up or drops loads. The predictive model outputs the whole horizon in one shot—predicting changes in joint states rather than absolute states—so it avoids error accumulation and keeps the optimization cheap. The control sequence is found by gradient descent with an adaptive learning rate, and the average of the sequence is applied rather than the first value. The paper argues that if this works as demonstrated, the approach generalizes to other industrial systems that need nonlinear control plus energy management.","feed_headline":"Neural controller steers a 22-ton excavator at 50 Hz","feed_subtitle":"An LSTM-MLP model predicts and corrects in one pass, so the machine adapts to loads and saves energy.","key_machinery":"The load-bearing object is the SSMP (single-shot multi-step prediction) model, a layered LSTM-MLP network that maps a block of historical states and a block of future control inputs to a block of predicted state increments $\\Delta \\hat{X}_{t+1:t+i}$. An LSTM encodes the history, and an MLP combines that encoding with the planned future inputs, so the derivative $\\partial G/\\partial U$ is available through the chain rule; this makes gradient-descent optimization practical. Around it, a small online MLP $H$ is trained online on the mismatch between the offline prediction and the measured state, and the NMPC cost function is minimized over the full input sequence. Two operational choices carry much of the argument: predicting state changes instead of absolute states, and applying the average of the optimized control sequence instead of the first element.","core_discovery":"On the paper's own terms, the central discovery is that a hybrid offline-online neural predictor can serve as the dynamics model inside a real-time NMPC for a heavy hydraulic excavator, without needing explicit physical equations. The offline LSTM-MLP model, called SSMP, is trained to predict the sequence of state changes over a horizon in one forward pass; the online MLP learns the difference between the offline model's predictions and the measured states during operation, acting as a correction for load-induced changes. Because the whole predictor is built from layers whose derivatives are chain-rule computable, the NMPC cost function—combining tracking error, velocity error, and engine speed—can be minimized by gradient descent with a learning rate that shrinks when position error is small. The controller then applies the mean of the optimized control sequence rather than its first element. Simulations and experiments with a 22-ton hydraulic excavator are presented as evidence that this scheme tracks reference trajectories, handles a 1500 kg load at the end joint, reduces hydraulic flow overflow compared with fixed-gear PID control, and computes fast enough for 50 Hz command updates.","pith_inferences":["An extension the paper leaves implicit: the same offline-online predictor could be applied to other load-varying hydraulic machines such as telehandlers or wheel loaders, but the claimed generality would need testing on machines with different actuator dynamics and sensor suites.","The paper does not analyze what happens when the gradient descent hits a poor local optimum; one could stress-test the controller by initializing the optimization from several different control sequences and checking whether the final commands and tracking stay consistent.","The adaptive learning rate in Eq. (12) is a heuristic that scales the step by position error; a testable refinement would compare it with a line-search or momentum rule on the same excavator dataset to see whether the fixed 30-iteration budget remains adequate.","The average-of-sequence control law is unusual relative to standard receding-horizon MPC; one implication is that it effectively low-pass filters the optimized sequence, which might trade aggressiveness for stability—an effect worth isolating by ablating exactly this choice."],"forward_implications":["According to the paper, the SSMP predictor reports consistently lower ARMSE on random trajectories than the DBN multi-step baseline trained on the same data, so direct multi-step change prediction is presented as a more accurate predictive model.","With the online compensation model, prediction error under an added end-effector load falls by at least 50% across joints and gear settings, so the hybrid model is claimed to adapt to load changes without retraining the offline network.","The NMPC tracks reference joint trajectories under no-load and 1500 kg load conditions while choosing among low, medium, and high engine gears, and it cuts flow overflow compared with fixed-gear PID; the paper uses this to argue the method handles both motion control and energy efficiency.","Timing measurements on a laptop CPU put one NMPC update between about 5 and 16 ms for the tested configurations, which the paper says is fast enough for the 50 Hz excavator interface.","Because the cost function can be written in convex form and solved by gradient descent, the paper claims the same controller structure can be reused for other industrial systems with multiple objectives, including energy management."],"supporting_citations":[{"why":"Supplies the LSTM-MPC approach for learning-based predictive control that the SSMP model builds on and contrasts with.","marker":"[21]"},{"why":"Provides the DBN multi-step prediction baseline used for the ARMSE comparisons in the offline model evaluation.","marker":"[24]"},{"why":"Introduces the free-and-forced response decomposition for echo-state-network NMPC that motivates the offline-online hybrid predictor.","marker":"[27]"},{"why":"Describes offline-plus-online active learning for dynamics models, which informs the paper's online mismatch-compensation strategy.","marker":"[22]"},{"why":"Supplies the free-and-forced response viewpoint for practical nonlinear predictive control used to justify the predictor structure.","marker":"[35]"},{"why":"Provides the dead-zone-compensated PID baseline used for comparison in both simulation and real excavator experiments.","marker":"[37]"},{"why":"Shows derivative-based optimization of an MLP NARMAX model for NMPC, the lineage of the gradient-descent solver used here.","marker":"[23]"}],"fun_headline_variants":["Hybrid neural model powers real-time control for heavy excavator","LSTM-MLP predictive control guides 22-ton excavator in real time","Data-driven NMPC learns on the fly for heavy hydraulic robots","Neural NMPC adapts to loads, saving energy in excavators","One-pass predictor enables fast NMPC for 22-ton excavator"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that thirty iterations of gradient descent with the adaptive learning rate always produce a good enough control sequence from the zero initial guess; the paper gives no proof, and if the optimization stalls in a bad local optimum the commands to the 22-ton machine would degrade.","fun_headline_variants_meta":{"raw":{"variants":["Hybrid neural model powers real-time control for heavy excavator","LSTM-MLP predictive control guides 22-ton excavator in real time","Data-driven NMPC learns on the fly for heavy hydraulic robots","Neural NMPC adapts to loads, saving energy in excavators","One-pass predictor enables fast NMPC for 22-ton excavator"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000655,"raw_usage":{"total_tokens":3048,"prompt_tokens":1040,"completion_tokens":2008,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":656,"completion_tokens_details":{"reasoning_tokens":1913}},"tokens_in":656,"tokens_out":2008,"duration_ms":13297,"temperature":1.0,"reasoning_tokens":1913,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T15:47:48.694309+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"On a logged or simulated trajectory of the same excavator, rerun the NMPC optimization from several different initial control sequences, including nonzero starts, and compare the resulting commands; if the gradient descent settles into materially different control sequences, or tracking error diverges on one start, the fixed-iteration assumption is not reliable. A cheaper check is to record the cost function value over the 30 iterations and see whether it is still changing substantially at iteration 30.","supporting_citations":[{"cited_title":"Lstm-mpc: A deep learning based predictive control method for multimode process control,","cited_arxiv_id":null,"evidence_quote":"Supplies the LSTM-MPC approach for learning-based predictive control that the SSMP model builds on and contrasts with."},{"cited_title":"Deep learning-based model predictive control for continuous stirred-tank reactor system,","cited_arxiv_id":null,"evidence_quote":"Provides the DBN multi-step prediction baseline used for the ARMSE comparisons in the offline model evaluation."},{"cited_title":"Echo state networks for practical nonlinear model predictive control of unknown dynamic systems,","cited_arxiv_id":null,"evidence_quote":"Introduces the free-and-forced response decomposition for echo-state-network NMPC that motivates the offline-online hybrid predictor."},{"cited_title":"Active learn- ing of discrete-time dynamics for uncertainty-aware model predictive control,","cited_arxiv_id":null,"evidence_quote":"Describes offline-plus-online active learning for dynamics models, which informs the paper's online mismatch-compensation strategy."},{"cited_title":"A practical approach to predictive control for nonlinear processes,","cited_arxiv_id":null,"evidence_quote":"Supplies the free-and-forced response viewpoint for practical nonlinear predictive control used to justify the predictor structure."},{"cited_title":"Robust adaptive precision motion control of hydraulic actuators with valve dead-zone compensation,","cited_arxiv_id":null,"evidence_quote":"Provides the dead-zone-compensated PID baseline used for comparison in both simulation and real excavator experiments."},{"cited_title":"Neural-network- based nonlinear model predictive control for piezoelectric actuators,","cited_arxiv_id":null,"evidence_quote":"Shows derivative-based optimization of an MLP NARMAX model for NMPC, the lineage of the gradient-descent solver used here."}],"review_version":1}