{"id":"13b08da8-6faa-47e6-8db1-00f03ef5fc63","arxiv_id":"2508.00928","paper_version":1,"verdict":"REJECT","confidence":"HIGH","novelty_score":4.0,"correctness_risk":"high","formal_verification":"none","parameter_count":2,"one_line_summary":"A tuned MPC model of head-neck control reproduces lateral perturbation responses using muscle effort and partial somatosensory feedback, without corrective orientation integrators.","lead":"This paper extends a model predictive control framework to simulate how the head and neck stabilize under sideways (lateral) perturbations, comparing the model against human experimental data. The finding is that muscle effort plus partial somatosensory feedback fits the data best, without needing extra postural integrators.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Claimed validation is in-sample: Eq. 2 tunes the MPC weights on the same Forbes data used as validation in Section 5, with no held-out data or error metrics; the acknowledged 39.86 mm posture error adds parameter-compensation risk.","rationale":"The reader's weakest_assumption highlights plant fidelity (the head CG-T1 displacement exceeding the physiological bound), which is a valid and concrete concern. However, the more load-bearing problem is that the validation is in-sample: Eq. 2 defines the tuning objective on the same lateral perturbation dataset used as the validation in Section 5, so the fit cannot demonstrate generalization. The posture error compounds this by giving the fitted weights a plausible alternative explanation as compensation for plant inaccuracies. I therefore agree with the reader's REJECT verdict and would keep it unchanged. A leave-one-participant-out cross-validation with reported quantitative error metrics would directly test whether the claimed accurate reproduction holds out of sample, and would separate genuine CNS strategy from overfitting and parameter compensation.","tokens_in":7293,"tokens_out":3431,"duration_ms":41343,"concrete_test":"Perform leave-one-participant-out cross-validation on the Forbes lateral-translation dataset: for each held-out participant, re-run the Section 2.3 high-level optimization on the remaining participants' data, then compute RMSE and FRF gain/phase errors between simulated and experimental responses for the held-out participant using the \"muscle effort + partial somatosensory\" configuration. Report the distribution of held-out errors. If held-out errors are comparable to training errors, the in-sample fit is not overfitting; if they are substantially larger, the claimed \"experimental validation\" (Abstract, Section 6) is not established.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that muscle effort plus partial somatosensory feedback \"accurately reproduces\" human head-neck responses and captures CNS decision making (Abstract, Section 6). The presented evidence does not support this because the validation is in-sample: the high-level optimization (Section 2.3, Eq. 2) minimizes RMSE between simulated and experimental signals from the same Forbes lateral-translation dataset used as the \"validation\" in Fig. 3, and the full weight vector W (Eq. 3) is the output of that optimization. No held-out subjects/trials, cross-validation, or quantitative RMSE/FRF error values are reported; the text only states that the muscle configuration \"generally provides a closer match.\" Without an out-of-sample check, the fitted weights could simply memorize the tuning dataset, making \"experimental validation\" circular with respect to the tuning objective. Second, the plant itself is acknowledged to be physiologically inaccurate: Section 4 reports a steady-state head CG-T1 displacement of 39.86 mm, exceeding the 34.5 mm upper bound, with lower neck joint error of about 10 degrees in the chosen configuration. Because W is optimized on this flawed plant, the resulting weight ratios (e.g., Wty1/Wty2 ~ 5, Wwx1 largest) may be compensating for plant errors rather than revealing a genuine CNS cost function. Thus the core empirical claim — accurate reproduction of human lateral head-neck responses and sufficiency of the proposed cost terms — is currently unsupported.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The manuscript extends an MPC-based head-neck postural control framework from sagittal-plane perturbations to lateral (coronal-plane) perturbations. The model uses a Simscape biomechanical plant, re-tunes the MPC cost weights through a high-level optimization against a lateral-perturbation dataset from Forbes et al., and compares two configurations (muscle effort with partial somatosensory feedback, with or without a head-in-space integrator at the upper neck joint) against experimental human data in the time and frequency domains. The authors conclude that the muscle-effort configuration best reproduces the human responses and that these cost terms are sufficient to capture CNS decision making during lateral head-neck perturbations.","tokens_in":7531,"tokens_out":4574,"duration_ms":51733,"significance":"If the central validation claim held, a computationally efficient MPC-based head-neck model with a plausible CNS cost structure would be a useful tool for automated-vehicle comfort assessment. The paper is transparent about the model's posture error and provides the full optimized weight vector, runtime information (RTF 8-11), and a feature importance analysis. However, the reported agreement is an in-sample fit to the same dataset used for tuning, and no quantitative error metrics are provided; these issues directly undermine the abstract's claim of experimental validation. The paper's value is therefore more as a modeling proposal than as a validated predictive model.","major_comments":[{"comment":"The validation is in-sample. The high-level optimization in Eq. (2) minimizes the RMSE between simulated and experimental signals from the Forbes lateral-translation dataset, and Fig. 3 then compares the model with 'experimental human data' from that same dataset. No held-out subjects or trials, cross-validation, or out-of-sample evaluation is reported. The agreement shown in Fig. 3 is therefore a restatement of the tuning objective, not an independent validation. The abstract's claim that the model 'can accurately reproduce dynamic responses' requires a genuine out-of-sample test and per-condition error metrics.","section":"Section 2.3, Eq. (2); Section 5, Fig. 3"},{"comment":"The plant model has an acknowledged, load-bearing posture error: the steady-state head CG-T1 displacement is 39.86 mm, exceeding the 34.5 mm upper bound cited from the literature, and the selected configuration leaves a lower-neck joint error of approximately 10 degrees. Because the MPC weights W are optimized on this plant, the optimized weight ratios (e.g., Wty1/Wty2 about 5, and the dominant Wwx1) may be compensating for plant inaccuracy rather than revealing an intrinsic CNS cost function. The Section 6 conclusion that muscle effort plus partial somatosensory feedback 'are sufficient to capture the CNS decision making' is not supported without demonstrating robustness of the fitted weights to plant variations or correcting the plant posture error.","section":"Section 4"},{"comment":"The performance comparison is only qualitative: the text states that the muscle configuration 'generally provides a closer match' to the experimental responses. No RMSE values, frequency-response gain/phase errors, confidence intervals, or subject-trial statistics are reported. Without numerical error metrics, the central claim of 'accurate' reproduction cannot be quantified, and the claimed superiority of one configuration over the other cannot be substantiated.","section":"Section 5"}],"minor_comments":[{"comment":"In the sentence listing weight changes, the second mention of 'Wwy1 by a factor of approximately 25.6' should refer to Wwy2; Table 1 shows Wwy2 changing from 1.62 to 41.40.","section":"Section 4, paragraph after Table 1"},{"comment":"The phrase 'withinl interval k' is a typo and should read 'within interval k'.","section":"Section 2.2, after Eq. (1)"},{"comment":"The reference to '(Fig. 3)' for the experimental setup points to the later results figure. The experimental setup should be shown in a dedicated figure or the cross-reference corrected.","section":"Section 3"},{"comment":"The notation r E[.] is nonstandard; please specify the expectation is taken over (time points, trials, subjects) and how it is estimated from the discrete data.","section":"Eq. (2)"}],"recommendation":"reject","confidential_remarks":"The manuscript relies heavily on the prior framework in [9], which is cited as an SSRN preprint; the editor may wish to verify that this prior work has undergone peer review or that the present manuscript is self-contained. Additionally, the random forest feature importance analysis in Table 2 is presented without details on training or validation, and it does not constitute a test of the model's predictive performance; it should not be used to support the validation claim."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Short version: the paper is a lateral-perturbation extension of the same authors' MPC head-neck framework, and the new material is the coronal-plane dataset, a locked lower-neck yaw simplification, and an RTF around 8-11. That is a legitimate incremental step for occupant-modeling work, and the real-time factor is a genuinely useful number if it generalizes beyond their desktop.\n\nWhat I credit: the authors are explicit about the fitting process. Eq. (2) says the ten MPC weights are tuned by minimizing RMSE against the Forbes lateral-translation dataset, and Section 5 then shows the model against the same dataset. They also disclose the posture problem: steady-state head CG-T1 displacement of 39.86 mm, above the 34.5 mm healthy bound, with ~10 deg lower neck error in the chosen configuration. There is no hidden claim here. The random-forest feature importance table is honestly described as derived from the optimizer history, which makes it about the fitted model, not about CNS priorities.\n\nThe soft spot is load-bearing: the 'experimental validation' in the abstract is in-sample fit. Without held-out subjects, held-out trials, or at least a cross-validation split, the reported agreement only shows that the optimizer could find weights that reproduce the data it was given. That is a fitting result, not a validation. The claim that muscle effort plus partial somatosensory feedback 'captures CNS decision making' cannot be separated from the fact that the plant is acknowledged to be off in a way that could shift the optimal weights. It is not circular in the logical sense, but it is circular as evidence.\n\nI also note that the 'proven already for anterior-posterior perturbations' rests on [9], an SSRN preprint, not a peer-reviewed source. So the foundational claim is thinner than the sentence suggests.\n\nWho this is for: people building fast surrogate head-neck models for AV comfort and motion-sickness studies. They can use it as a modeling case study, not as a validated CNS mechanism. The right next step is out-of-sample testing - hold out subjects, report per-subject RMSE and FRF error, and reframe the abstract as 'fits' rather than 'validates.' I'd send it back for that revision, not desk-reject it out of hand.","headline":"Incremental MPC extension with a useful RTF number, but the abstract overclaims validation: the weights are fit on the same Forbes data used for the comparison, so the agreement is in-sample.","tokens_in":8117,"tokens_out":2747,"would_cite":false,"duration_ms":32757,"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":"This paper argues that a model-predictive controller minimizing muscle effort and the mismatch between predicted and sensed joint motion is sufficient to reproduce human head-neck responses to lateral trunk perturbations, without…","keywords":["model predictive control","head-neck dynamics","lateral perturbations","postural stabilization","somatosensory feedback","muscle effort","automated vehicles","compensatory postural adjustments"],"falsifier":"Correct the model so the static head center-of-gravity to T1 displacement falls below the literature upper bound of 34.5 mm and re-run the weight optimization against the same lateral perturbation dataset; if the dynamic fit degrades substantially or the optimized weights shift dramatically, the claimed muscle-effort-plus-partial-somatosensory-feedback strategy is partly an artifact of the plant. Alternatively, record neck muscle activity during the same perturbations and check whether the predicted distribution of effort between lower and upper neck joints matches the measured activation pattern.","tokens_in":7025,"feed_emoji":"🧠","tokens_out":11102,"duration_ms":122115,"temperature":0.7,"pith_summary":"The paper tries to show that a single model-predictive control strategy—choose muscle torques by simulating a short future horizon of head-neck motion and penalize both muscular effort and the gap between predicted and sensed feedback—can reproduce how humans stabilize the head when the trunk is pushed sideways. This matters for automated vehicles, where occupants may be reading or watching a screen with no warning of the vehicle's motion; the head carries the visual and vestibular sensors, so a validated control law that predicts head motion would let engineers assess comfort and motion sickness without running human experiments for every maneuver. The model is validated against published human data from lateral seat perturbations with the torso restrained, comparing time and frequency responses. The authors find that muscle effort combined with partial somatosensory feedback alone gives the best dynamic fit, and that adding corrective integrators for head-on-trunk or head-in-space orientation fixes steady-state posture but degrades the dynamic response. Their conclusion is that these two cost terms are sufficient to capture the central nervous system's decision making during laterally perturbed head-neck dynamics.","feed_headline":"Muscle effort and sensory feedback match head-neck lateral response","feed_subtitle":"If right, the same control policy could predict occupant head motion in automated vehicles.","key_machinery":"The central object is the model predictive controller itself, treated as the central nervous system's decision process. At each time step it predicts a horizon of head-neck trajectories from a simplified two-joint biomechanical model (lower neck representing T1–C7 and upper neck representing C0–C1, with lower-neck yaw locked) and selects joint torques that minimize a weighted sum of muscle effort and somatosensory conflict, where somatosensory conflict is the difference between the plant's sensed joint motion and the internal model's prediction of it. A high-level optimization tunes the ten weights in the cost function by minimizing the root-mean-square error between simulated and experimental lateral perturbation responses, and a multi-collocation prediction horizon with 10 ms integration makes the computation fast enough to run at 8–11 times real time.","core_discovery":"The paper's central claim is that a model predictive controller whose cost function contains muscle-effort terms and partial somatosensory conflict terms reproduces the dynamic responses of the human head-neck system under lateral trunk perturbations. Muscle effort is represented by weighted joint torques, and partial somatosensory conflict is the weighted difference between predicted and sensed joint angular velocities for a chosen subset of neck joints. The controller's ten weights are fitted by a multi-objective genetic algorithm against the average human response in a lateral translation experiment. The best overall fit comes from the configuration without head-in-space or head-on-trunk integrators; the only visible exception is a slightly better low-frequency yaw gain when an upper-neck head-in-space integrator is active. The optimized weights emphasize lower-neck pitch muscle effort and lower-neck roll-rate somatosensory error, and the residual head center-of-gravity to T1 forward displacement (39.86 mm) exceeds the upper literature bound (34.5 mm), which the paper attributes to the chosen initial posture and the simplified biomechanical plant.","pith_inferences":["If the fitted weights are read as estimates of a real neural cost function, the large lower-neck pitch effort weight and lower-neck roll-rate sensory weight predict that the T1–C7 region is the dominant stabilizing actuator for lateral perturbations, a claim the paper does not test with electromyography.","A natural extension is to repeat the lateral perturbation experiment with eyes open or with a visual cue of the perturbation; the framework predicts that visual conflict terms would enter the cost function or the somatosensory weights would shift, which is testable.","Because the head CG–T1 displacement overshoots the anatomical range, the fitted weights may be partially absorbing plant error; re-identifying neck stiffness and damping from independent measurements would show whether the same weights still fit the data.","The feature-importance analysis suggests that perturbation spectra concentrated near the lower-neck roll resonance would most sharply discriminate between alternative weight vectors, so a follow-up experiment could be designed specifically around that frequency band."],"forward_implications":["The same MPC cost structure that handled anterior-posterior perturbations transfers to lateral perturbations with only re-tuned weights, suggesting a domain-general CNS objective rather than scenario-specific control.","Lower-neck control carries most of the explanatory load; models or experiments that omit lower-neck roll and pitch feedback should predict head-neck lateral responses less accurately.","Corrective head-on-trunk and head-in-space integrators are not needed for dynamic lateral responses, so real-time occupant-simulation pipelines can omit them.","The real-time factor of 8–11 makes the model usable as a virtual occupant in vehicle dynamics simulations for comfort and motion-sickness assessment.","The model's static head CG–T1 displacement overshoots the physiological range even though its dynamics match human data, marking posture prediction as the part of the model most in need of further work."],"supporting_citations":[{"why":"Supplies the original MPC postural-control framework, the cost-function structure, and the baseline weights this paper re-tunes for lateral perturbations.","marker":"[9]"},{"why":"Provides the published human lateral-trunk perturbation dataset used for experimental validation.","marker":"[6]"},{"why":"Provides the active-inference/sensory-conflict rationale for minimizing the gap between predicted and sensed feedback.","marker":"[14]"},{"why":"Supplies the head-in-space and head-on-trunk integrator approach that the paper tests and then rejects.","marker":"[5]"},{"why":"Supports the simplification of locking lower-neck yaw by showing axial neck mobility is concentrated in the upper neck.","marker":"[12]"},{"why":"Supplies the head center-of-gravity anterior displacement range used to choose and assess the model's posture.","marker":"[16]"},{"why":"Provides the observer-theory sensory-conflict concept that motivates the 'surprise factor' framing and the MPC objective.","marker":"[11]"}],"fun_headline_variants":["Muscle effort plus sensory feedback bests head-neck model","MPC with muscle effort and partial sensory matches head-neck data","Muscle effort and sensory cues control head-neck model","Predicting head motion via MPC with muscle and sensory terms"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The whole result rests on the assumption that the simplified two-joint neck model is a faithful stand-in for real head-neck mechanics during side-to-side motion; if that assumption is wrong, the fitted control weights could be compensating for model error rather than revealing the central nervous system's real priorities.","fun_headline_variants_meta":{"raw":{"variants":["Muscle effort plus sensory feedback bests head-neck model","MPC with muscle effort and partial sensory matches head-neck data","Muscle effort and sensory cues control head-neck model","Predicting head motion via MPC with muscle and sensory terms"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000664,"raw_usage":{"total_tokens":3009,"prompt_tokens":896,"completion_tokens":2113,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":512,"completion_tokens_details":{"reasoning_tokens":2043}},"tokens_in":512,"tokens_out":2113,"duration_ms":17120,"temperature":1.0,"reasoning_tokens":2043,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-06T11:24:52.744028+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Correct the model so the static head center-of-gravity to T1 displacement falls below the literature upper bound of 34.5 mm and re-run the weight optimization against the same lateral perturbation dataset; if the dynamic fit degrades substantially or the optimized weights shift dramatically, the claimed muscle-effort-plus-partial-somatosensory-feedback strategy is partly an artifact of the plant. Alternatively, record neck muscle activity during the same perturbations and check whether the predicted distribution of effort between lower and upper neck joints matches the measured activation pattern.","supporting_citations":[{"cited_title":"SSRN Electronic Journal DOI 10.2139/ssrn.5095834","cited_arxiv_id":null,"evidence_quote":"Supplies the original MPC postural-control framework, the cost-function structure, and the baseline weights this paper re-tunes for lateral perturbations."},{"cited_title":"Phd thesis, Delft University of Technology, The Netherlands","cited_arxiv_id":null,"evidence_quote":"Provides the published human lateral-trunk perturbation dataset used for experimental validation."},{"cited_title":"The MIT Press","cited_arxiv_id":null,"evidence_quote":"Provides the active-inference/sensory-conflict rationale for minimizing the gap between predicted and sensed feedback."},{"cited_title":"Multibody System Dynamics","cited_arxiv_id":null,"evidence_quote":"Supplies the head-in-space and head-on-trunk integrator approach that the paper tests and then rejects."},{"cited_title":"Spine 26(24):2692–2700","cited_arxiv_id":null,"evidence_quote":"Supports the simplification of locking lower-neck yaw by showing axial neck mobility is concentrated in the upper neck."},{"cited_title":"Journal of biomechanics 42:1177–92","cited_arxiv_id":null,"evidence_quote":"Supplies the head center-of-gravity anterior displacement range used to choose and assess the model's posture."},{"cited_title":"In: Ellis SR (ed) Pictorial Communication in Real and Virtual Environments, Taylor & Francis, London, pp 362–376","cited_arxiv_id":null,"evidence_quote":"Provides the observer-theory sensory-conflict concept that motivates the 'surprise factor' framing and the MPC objective."}],"review_version":1}