{"id":"a7eb7f04-3074-40a0-aa07-6b6cbd46d997","arxiv_id":"2411.16891","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":3,"one_line_summary":"Using ground reaction forces in simple acceleration models improves short-horizon prediction of the body's center of mass during everyday activities, with 250 ms as a practical limit.","lead":"This study tested three ways to predict where a person's center of mass will be up to 625 milliseconds ahead, using motion capture and force plate data from 10 people doing 14 everyday movements. Predictions that use ground reaction forces beat a zero-acceleration baseline at short horizons, and errors grow quadratically with prediction length, suggesting a 250 millisecond limit for reliable use.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"Abstract overstates ADA evidence: Table 2 shows Zero-Constant at T=250 ms is not significant (p=1), so the claim that both constant and cubic profiles improve direction accuracy at 125 and 250 ms is unsupported as written.","rationale":"I read the paper as an offline evaluation of three acceleration assumptions for CoM prediction using ideal initial conditions and GRF inputs. For the central claim to hold, the statistical comparisons in Section 3 must support the abstract's statements. The AE comparison does, and the Cubic ADA comparison does, but the Constant ADA comparison at 250 ms does not: Table 2 lists p=1. The reader's identified weakest assumption about unmeasured chair/contact forces is a legitimate secondary concern, but it is less decisive for the positive claim because unmeasured external forces would tend to degrade the GRF-based predictions rather than artificially inflate their advantage; a robustness check excluding sit-to-stand and stand-to-sit would be useful but is unlikely to reverse the Cubic result. The most load-bearing soft spot is therefore the internal mismatch between the abstract's joint claim and the reported pairwise tests. This does not warrant rejection: the core result that GRF information, at least via the cubic profile, improves short-horizon prediction is supported, so the reader's CONDITIONAL verdict stands, with the condition that the abstract be corrected. Other aspects, including the quadratic error growth, the Oracle lower bound, and the 250 ms recommendation, are secondary and appropriately hedged in the discussion.","tokens_in":14609,"tokens_out":11507,"duration_ms":115433,"concrete_test":"Independently recompute the Bonferroni-adjusted Welch t-test for ADA between Zero and Constant at T=250 ms using the per-subject average direction accuracy values. If the adjusted p-value is not <0.05, the abstract's claim that both constant and cubic profiles improve direction accuracy at 250 ms is not supported, and the abstract should attribute the 250 ms accuracy benefit to the Cubic profile only.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's headline claim (abstract; repeated in the reader's strongest_claim) asserts that 'the constant and cubic profiles, which utilize the GRFs, outperform the zero-acceleration assumption in ... accuracy ... at horizons of 125 and 250 ms (p<0.034, d>1.44).' The paper's own post-hoc tests in Section 3 and Table 2 do not support this for the Constant profile at 250 ms. For ADA at T=250 ms, Zero-Const has p=1 (no significant difference), while only Zero-Cubic is significant (p=0.034, d=1.44). The results text explicitly states that differences were found between Zero-Const only at T=125 ms and between Zero-Cubic at all T. Thus the central statistical claim, as worded, attributes a 250 ms direction-accuracy benefit to both GRF-based profiles, but the evidence supports only the Cubic profile for that metric and horizon. This is an internal inconsistency between the abstract and the reported analysis. It does not overturn the AE results or the Cubic-based evidence for GRF benefit, but it means the strong joint claim in the abstract should be corrected or the supporting test for Constant at 250 ms must be shown to be significant under an appropriate analysis.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"This manuscript studies short-horizon prediction of the whole-body center-of-mass (CoM) position during 14 non-cyclic activities performed by 10 healthy adults. The CoM is modeled as a double integrator driven by acceleration inputs derived from ground reaction forces, and four acceleration assumptions are compared: zero, constant, cubic-to-zero, and an oracle that uses future measured accelerations. For horizons of 125 to 625 ms, the authors report that average and maximum position errors grow quadratically with horizon length while direction accuracy decreases linearly, and that GRF-based constant and cubic profiles improve prediction at short horizons relative to the zero-acceleration baseline. The paper proposes 250 ms as a practical horizon threshold for intent inference in wearable robotics.","tokens_in":14882,"tokens_out":7654,"duration_ms":73912,"significance":"If the results hold, the paper provides a transparent, parameter-free baseline for CoM prediction and a concrete argument that even a single instantaneous GRF sample improves short-horizon prediction relative to a constant-velocity assumption. The statistical analysis is careful: Welch ANOVA, Bonferroni-corrected post-hoc tests, F-tests for trend degree, weighted least squares, confidence intervals, and reported effect sizes. The a priori definition of the acceleration profiles and the use of an oracle reference reduce circularity concerns, and the quadratic-error explanation in Appendix C is explicit and testable. The broad qualitative conclusions are likely robust, but the abstract overstates the direction-accuracy evidence for the constant profile, and the GRF-derived acceleration assumption is not satisfied for all analyzed activities due to unmeasured chair-contact forces.","major_comments":[{"comment":"The abstract claims that \"the constant and cubic profiles, which utilize the GRFs, outperform the zero-acceleration assumption in ... accuracy ... at horizons of 125 and 250 ms (p<0.034, d>1.44).\" Table 2 shows that for ADA at T=250 ms the Zero-Const comparison has p=1.00 with no reported effect size, so the Constant profile does not significantly improve direction accuracy at 250 ms; only the Zero-Cubic comparison is significant there (p=0.034, d=1.44). The results text in Section 3 also states that differences were found between Zero-Const only at T=125 ms. The abstract and the corresponding sentence in Section 4.2 should be corrected to distinguish the Constant and Cubic profiles, e.g., Constant improving direction accuracy only at 125 ms and Cubic improving it at all tested horizons.","section":"Abstract and Section 3, Table 2"},{"comment":"Equation (B.1) defines the acceleration input using ground reaction forces and gravity as the only external forces, but the protocol includes Sit to Stand, Stand to Sit, and Foot on Chair, during which the chair exerts reaction forces on the body that are not measured by the force plates. For samples in which the participant is still in contact with the chair, u[1] and the Oracle inputs are not the true CoM acceleration, so the error comparisons involving the Constant, Cubic, and Oracle profiles can be biased for those activities. The authors should either quantify the extent of chair-contact samples through a sensitivity analysis that excludes those activities or phases, or add an explicit limitation with evidence that the conclusions are unchanged; as written, the assumption in Eq. (B.1) is not satisfied for all analyzed data.","section":"Section 2.1 and Appendix B, Eq. (B.1)"}],"minor_comments":[{"comment":"The abstract reports R2>0.930 for the quadratic position-error fits, while Fig. 3b shows R2=0.918 for the Constant maximum-error fit; these numbers should be reconciled.","section":"Abstract and Fig. 3b"},{"comment":"The sentence \"with the Zero having a significantly larger ADA than Cubic at all levels of T\" appears to have the comparison reversed; Table 2 and Fig. 4a indicate that Cubic has significantly higher ADA than Zero, so the wording should be corrected.","section":"Section 4.2"},{"comment":"The manual labeling of force plate contact events is not described with any reliability measure; a brief statement on inter- or intra-rater consistency, or an automated rule-based alternative, would strengthen reproducibility.","section":"Section 2.1"},{"comment":"The choice of 125 ms as the minimum horizon is justified by a self-cited prior conference paper (Noghani and Bolivar-Nieto, 2024); restating the relevant evidence or providing an independent rationale would make the horizon grid self-contained.","section":"Section 2.2"},{"comment":"The statement that at T=250 ms the profiles achieved \"AE <0.57 cm, ME <11.7 cm\" should specify that these are mean values and should indicate whether the bounds apply to all profiles or only to the GRF-based profiles.","section":"Section 4.3"}],"recommendation":"major_revision","confidential_remarks":"The two major concerns are both addressable: the abstract needs a precise statement of which profile improves direction accuracy at which horizon, and the chair-contact issue needs either a sensitivity analysis or a clearly justified limitation. The paper's central idea is sound and the statistical execution is strong; with those revisions it would be suitable for publication."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Read this one if you care about simple, interpretable CoM prediction for wearable robotics. The headline result — GRF-based acceleration profiles beat zero-acceleration at short horizons — is real and well supported. But the abstract overstates the direction-accuracy evidence: Table 2 shows the Constant profile is not significantly better than Zero at 250 ms for ADA (p = 1), so the claim that both Constant and Cubic outperform Zero in accuracy at 125 and 250 ms is not what the data shows.\n\nWhat is genuinely new: the systematic comparison of zero, constant, and cubic-to-zero acceleration assumptions on 14 non-cyclic activities, across five horizons, using whole-body marker data as ground truth. The experiment is clean and the statistics are careful — Welch ANOVA, Bonferroni-corrected post-hoc tests, effect sizes, weighted least squares for the trend fits. The quadratic error growth is not just a fit; it follows from the double-integrator model (Appendix C) and matches the data with R² > 0.918. The Oracle profile provides a sensible lower bound. The AE benefit of GRF profiles at 125–250 ms is large and unambiguous (p < 0.001, d > 3.23).\n\nSoft spots, in order of importance. First, the abstract's accuracy claim: as written it attributes a 250 ms ADA benefit to both Constant and Cubic, but only Cubic is significant there. That is an internal inconsistency and should be corrected. Second, the 250 ms threshold is offered as a recommendation without a formal test; that is a descriptive observation, not a demonstrated cutoff. Third, chair contact during sit-to-stand and stand-to-sit means GRF-only acceleration misses external forces; the limitation is acknowledged but not quantified. That said, any bias would likely make the GRF profiles look worse, not better, so the central result is conservative. Missing code/data is a transparency issue but not a correctness issue.\n\nWho is this for? Researchers building simple CoM or intent predictors for exoskeletons and prostheses. It is not a breakthrough method, but it is a useful empirical guideline. It deserves a serious referee: send it to review with a request to fix the abstract and table alignment.","headline":"Solid empirical comparison with an abstract that overstates one accuracy finding; the AE result holds, the ADA claim needs correction.","tokens_in":15403,"tokens_out":2219,"would_cite":true,"duration_ms":22281,"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":"The paper shows that incorporating ground reaction forces into center-of-mass prediction, through constant or cubic acceleration profiles, cuts position error and improves direction accuracy at 125–250 ms horizons, and identifies 250 ms…","keywords":["center of mass prediction","ground reaction forces","non-cyclic activities","intent inference","wearable robots","acceleration profile","prediction horizon","double integrator"],"falsifier":"Re-run the prediction pipeline with an instrumented chair or handrail that records contact forces during sit-to-stand and stand-to-sit, and compare ground-truth CoM acceleration from force plates plus contact forces against the whole-body marker estimate; if the zero-acceleration baseline then matches the GRF-based profiles at 125 and 250 ms, the reported benefit would be shown to depend on unmeasured external contacts.","tokens_in":14415,"feed_emoji":"🦿","tokens_out":5963,"duration_ms":49530,"temperature":0.7,"pith_summary":"The paper asks whether a wearable system can predict where the whole-body center of mass (CoM) will be a fraction of a second from now, using only the current state and a guessed future acceleration, and whether adding ground reaction forces (GRFs) helps. It analyzes data from 10 healthy adults performing 14 non-cyclic activities, comparing three acceleration forecasts—zero, constant, and a cubic that decays to zero—across horizons from 125 to 625 ms. The central result is that GRF-informed constant and cubic profiles clearly beat the zero-acceleration assumption at 125 and 250 ms horizons, in both position error and direction accuracy, and that 250 ms is the point where prediction quality starts to degrade sharply. A sympathetic reader would care because CoM position is a proxy for movement intent, and a 250 ms lookahead window is about the timescale on which exoskeletons and prostheses need to act.","feed_headline":"Ground reaction forces sharpen center-of-mass prediction at 250 ms","feed_subtitle":"Constant and cubic profiles beat zero-acceleration on error and direction, marking a practical wearable-robot intent horizon.","key_machinery":"The load-bearing mechanism is a discrete-time double-integrator state-space model of the CoM, $\\ddot{p}=u$, where the input $u$ is the GRF-derived acceleration (net external force divided by mass, with gravity removed). Given current position and velocity, the future position is computed by propagating the state with one of three assumed acceleration profiles over the horizon: zero, constant at the first sample, or a cubic that starts at the first sample and decays to zero with zero jerk at both ends. The cubic profile is motivated by minimum-jerk humanoid planning; the oracle profile uses the future GRFs directly and serves as a lower bound. Because the error formula (Appendix C) reduces to a double integral of the difference between assumed and true acceleration, a constant mismatch under the simplifying assumptions produces the observed quadratic error growth in horizon length.","core_discovery":"The authors claim that future center-of-mass position over short horizons can be predicted by forward integration of a double-integrator model, and that the quality of that prediction is governed mainly by horizon length and by whether the assumed acceleration profile uses ground reaction forces. Using whole-body marker trajectories as ground truth, they find that position error grows quadratically with horizon (R² > 0.930 for all profiles) and direction accuracy falls linearly (R² > 0.615). At horizons of 125 and 250 ms, the constant and cubic-to-zero profiles—both seeded with the GRF-derived acceleration at the first sample—outperform the zero-acceleration baseline on average error and average direction accuracy, with statistically significant differences and large effect sizes; at longer horizons the advantage fades and sometimes reverses for worst-case error. The authors therefore treat 250 ms as a threshold for practical predictive intent inference.","pith_inferences":["Editorial inference: the one-sample advantage of GRF information suggests that even partially available GRFs in instrumented footwear or bionic limbs could improve short-horizon intent inference without a full motion-capture ground truth.","Editorial inference: because the quadratic error growth follows from double integration of a constant acceleration mismatch, the same scaling should appear whenever a point-mass CoM model is forward-integrated with a mismatched acceleration forecast; testing this in a different dataset (e.g., walking perturbations) would check the generality.","Editorial inference: chair contact forces during sit-to-stand and stand-to-sit are a plausible unmeasured external force; adding an instrumented chair or handrail would reveal whether the reported GRF benefit changes when contacts are not on the force plates.","Editorial inference: the near-oracle performance at 125 ms aligns with neuromuscular reaction time, suggesting that a control loop running at this timescale might be the natural operating point for balance-assist controllers."],"forward_implications":["At horizons up to 250 ms, GRF-informed constant and cubic profiles reduce average CoM position error and improve direction accuracy relative to zero-acceleration prediction, with the largest effects at 125 ms.","Prediction error grows quadratically and direction accuracy declines linearly with horizon for all profiles, so extending the prediction window beyond a few hundred milliseconds carries a steep accuracy cost.","The 250 ms horizon is proposed as a practical threshold for CoM-based intent prediction in applications such as lower-limb wearable robots.","At longer horizons, zero-acceleration can produce lower maximum errors than GRF-based profiles, because a single early acceleration sample becomes an unreliable forecast.","Even the oracle profile, which uses the true future GRFs, accumulates integration drift and loses direction accuracy, so no constant-profile method can fully escape horizon-dependent degradation."],"supporting_citations":[{"why":"Supplies the constant-acceleration forecasting convention used by the Const profile.","marker":"(Murphy, 2023)"},{"why":"Motivates the minimum-jerk cubic-to-zero profile for CoM acceleration.","marker":"(Van Heerden, 2017)"},{"why":"Provides the full-body musculoskeletal model used to compute the marker-based CoM reference in OpenSim.","marker":"(Rajagopal et al., 2016)"},{"why":"Provides the OpenSim platform used for inverse kinematics and CoM estimation.","marker":"(Delp et al., 2007)"},{"why":"Source of the 14 balance-test activities adapted for data collection.","marker":"(Berg et al., 1992)"},{"why":"Prior data-driven CoM prediction work whose quadratic error trend the paper compares with.","marker":"(Leestma et al., 2024)"},{"why":"Prior work establishing the 125 ms minimum horizon length used in this study.","marker":"(Noghani and Bolívar-Nieto, 2024)"},{"why":"Supports the wearable-robot motivation by noting that GRFs are partially available in bionic legs.","marker":"(Azocar et al., 2020)"}],"fun_headline_variants":["GRFs boost CoM prediction at short horizons","250ms threshold for CoM prediction with GRFs","Quadratic error growth limits CoM prediction horizon","GRFs sharpen CoM prediction up to 250ms","Predicting CoM: GRFs matter for short horizons"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The comparisons assume that the whole-body marker-based OpenSim CoM estimate is an unbiased ground truth and that the measured ground reaction forces plus gravity are the only external forces acting on the body, so unmeasured chair or handrail contacts during activities like sit-to-stand and stand-to-sit would violate the acceleration input.","fun_headline_variants_meta":{"raw":{"variants":["GRFs boost CoM prediction at short horizons","250ms threshold for CoM prediction with GRFs","Quadratic error growth limits CoM prediction horizon","GRFs sharpen CoM prediction up to 250ms","Predicting CoM: GRFs matter for short horizons"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000346,"raw_usage":{"total_tokens":1958,"prompt_tokens":1066,"completion_tokens":892,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":682,"completion_tokens_details":{"reasoning_tokens":816}},"tokens_in":682,"tokens_out":892,"duration_ms":5841,"temperature":1.0,"reasoning_tokens":816,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T12:45:38.812861+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Re-run the prediction pipeline with an instrumented chair or handrail that records contact forces during sit-to-stand and stand-to-sit, and compare ground-truth CoM acceleration from force plates plus contact forces against the whole-body marker estimate; if the zero-acceleration baseline then matches the GRF-based profiles at 125 and 250 ms, the reported benefit would be shown to depend on unmeasured external contacts.","supporting_citations":[],"review_version":1}