{"id":"3796200c-8dba-4cca-a71a-381e5f446109","arxiv_id":"2411.14701","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":4,"one_line_summary":"A personalised skeletal human model with soft-body feet, generated from motion capture data and trained with a walking policy, reproduces measured ground reaction forces in simulation.","lead":"Researchers built a personalised digital human that walks in a physics simulator, with soft deformable feet shaped from motion capture of a real subject. The soft feet produce ground reaction forces resembling the real walker's measured forces, which could help simulate human interaction with assistive robots.","discovery_kind":"extension","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The reported GRF advantage of the soft-foot model is confounded by post-hoc phase offsets and stiffness tuning; the EM metric is computed after aligning to the same data it claims to predict.","rationale":"The reader's verdict is CONDITIONAL, which is appropriate. I agree with the core concern: the validation protocol is confounded by applying phase offsets and tuning material properties against the same GRF data that is later used to compute the EM metric. This is the most load-bearing weakness because the paper's only clear quantitative advantage of the soft-foot model over the skeleton baseline is the GRF EM value. The joint-angle analysis actually shows the skeleton-only model tracks the reference as well as or better than model E on most joints, so if the GRF advantage evaporates under a clean held-out test, the central claim loses its empirical support. I did not take up the secondary concern about the realism of the 15-vertex uniform-material flex feet because, while relevant to generality, it is not needed to identify the methodological flaw; the post-hoc alignment alone prevents the reported numbers from supporting the claim. A held-out evaluation with reported offset values would settle whether the soft-body contact model genuinely predicts measured forces. If that test confirms the advantage, the conditional verdict could be upgraded; if not, the claim should be weakened to a demonstration of feasibility rather than accuracy.","tokens_in":8755,"tokens_out":3242,"duration_ms":34049,"concrete_test":"Re-run the evaluation with a held-out protocol: hold out two of the six left-foot gait cycles; tune the flex stiffness, foot position, and any phase offset on the remaining four cycles; report offset values; then compute EM on the two held-out cycles for model E and model SK. If the held-out EM of model E is not clearly above model SK's held-out EM, the reported superiority of the soft-body feet is an artifact of post-hoc fitting rather than a predictive property of the model.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The central claim is that the soft-body feet generate GRF comparable to real measured data. The evidence for this is Table 4, where model E (EM=0.628) beats the skeleton-only model SK (EM=0.558). But Section 3.2 states that for models A, B, and E, 'offsets were applied to better match the GRF profiles,' and the offset values are never reported. The EM metric is then computed on these offset-adjusted curves. Because the offsets are chosen on the same recordings used to evaluate match, the EM values are fit statistics, not prediction accuracies. This is compounded by Section 2.3, which says flex stiffness and foot position were fine-tuned using GRF and joint-angle results, again on the same data. So the improvement from model SK to model E could be entirely an artifact of aligning and tuning to the reference GRF, and the abstract's 'comparable to real measured data' claim is not supported by an unbiased evaluation. The joint-angle results are also weaker than the GRF claim: Table 5 shows model SK has higher mean R2 than model E on five of six joints, so the soft-foot model does not improve kinematic fidelity; the GRF improvement is the only claimed benefit, and it is confounded.","agreement_with_reader":"partial"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a personalised 3D human digital twin in MuJoCo that augments an existing skeletal (SK) model with soft-body feet represented as simplified flex objects. The feet are generated from motion-captured SMPL body shapes, reduced to roughly 15 vertices, attached to the skeleton with hard pins, and assigned hyperelastic material properties. A walking control policy is trained for several variants: the original skeleton (SK), two softer-foot models (A and B), and a final stiffened/repositioned model (E). The models are evaluated against recorded ground reaction forces (GRF) using an experimental-match (EM) metric and against joint angles using a linear-fit method (LFM). The paper reports that model E achieves the best GRF match (average EM 0.628 vs 0.558 for SK) and joint-angle results close to the SK model and reference motion.","tokens_in":8943,"tokens_out":4348,"duration_ms":42026,"significance":"If the evaluation were unbiased, the paper would make a useful engineering contribution: it demonstrates a low-cost way to add personalised soft-body contact to an existing skeletal model while training the control policy only on kinematic motion capture. The pipeline is concrete and reproducible in principle, and the use of real subject data, multiple stiffness variants, EM, and LFM metrics are strengths. However, the central empirical claim is not currently supported because the material properties, foot position, and phase offsets are tuned against the same recordings used for evaluation, and no held-out validation is performed. The kinematic results also show that model E is generally worse than the bare skeleton, so the soft-foot benefit rests almost entirely on the confounded GRF comparison. The contribution is therefore promising but requires a substantially strengthened evaluation before the reported claims can be accepted.","major_comments":[{"comment":"The central GRF comparison is confounded because the EM values for models A, B, and E are computed only after hand-applied phase offsets were used to 'better match the GRF profiles,' and the offset values are not reported. Because these offsets are chosen on the four gait cycles from the same recordings on which EM is then evaluated, the EM values in Table 4 are fit statistics rather than unbiased predictions; consequently, the improvement of model E over model SK (0.628 vs 0.558) does not by itself support the abstract's claim that the soft-body feet generate GRF comparable to real measured data. Please report the offsets, justify them independently of the evaluation data (e.g., from a calibration gait cycle), and recompute EM without data-dependent alignment.","section":"§3.2 and Table 4"},{"comment":"The material stiffness and foot position were fine-tuned using the GRF and joint-angle results of the same recordings used for the evaluation, as stated in Section 3.1: 'the material properties of the flex and the position of the flex foot were fine-tuned to improve the GRF and joint angle results.' This creates a direct circularity for the main empirical claim: the reported match of model E may largely reflect fitting to the test data. Please describe the tuning procedure (search grid, number of configurations tried, stopping criterion), and validate the final model on held-out gait cycles or subjects not used during tuning.","section":"§2.3 and §3.1"},{"comment":"The joint-angle results do not support the claim that model E 'closely follow[s] joint angle results of the bare skeletal model.' In Table 5, the mean R² of model SK is higher than that of model E for five of the six joints (e.g., L Ankle 0.950 vs 0.838; R Knee 0.928 vs 0.918), and model E only exceeds SK on R Ankle (0.945 vs 0.893). The paper should either soften the kinematic claim or provide an analysis showing that the differences are within measurement noise and not systematically worse.","section":"§3.3 and Table 5"},{"comment":"The GRF evaluation is limited to one foot and a small number of recordings: six complete left-foot recordings, with the right foot excluded because only three complete recordings were available. With n=6, the mean and standard deviation of the reference GRF are noisy, and no statistical test is provided for the EM differences across models; the per-cycle EM ranges for SK (0.54–0.58) and E (0.60–0.67) do not overlap, but it is unclear whether this is significant given the small sample. At a minimum, report per-recording values, the number of subjects, and a confidence interval or test for the EM comparison.","section":"§3.2"}],"minor_comments":[{"comment":"The connective tissue Young's modulus range is printed as '1.5 × 10^6 – 2.25 × 10^5 kPa', which is in descending order and uses units that appear inconsistent with the muscle and fat rows; please verify the values and units.","section":"Table 1"},{"comment":"The sentence 'The higher EM values of model A over model SK confirm this' contradicts Table 4, where model A has average EM 0.463 ± 0.013 and model SK 0.558 ± 0.022; please correct the text or the table.","section":"§3.2"},{"comment":"The statement that 'the model took on average 11.7 seconds to train and around 4200 iterations to converge' alongside 'the base skeletal model required 5.3 seconds per iteration' suggests a missing 'per iteration' in the first quantity; please clarify.","section":"§3.1"},{"comment":"Please add a data/code availability statement or clarify whether the models, motion-capture recordings, and evaluation scripts are available for replication.","section":"General"}],"recommendation":"major_revision","confidential_remarks":null},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Colleague,\n\nThe good news first: this paper gives people working in MuJoCo a concrete recipe for adding deformable, personalised feet to an existing skeletal model. Taking SMPL foot meshes, decimating them to 15 vertices, converting them to flex objects, and attaching them with hard pins is exactly the kind of integration detail that is normally buried in appendices. The comparison between the elasticity plugin and direct flex property definition is also useful. The authors are honest about the simplifications: they only used the left foot, only six recordings, and they acknowledge the shape mismatch.\n\nThe soft spot is the evaluation. The central claim is that the soft-body feet generate GRF comparable to measured data. But material stiffness and foot position were fine-tuned against the same GRF and joint-angle data used for evaluation (Section 2.3, Section 3.1). Then for models A, B, and E, offsets were applied to the GRF curves before computing the EM metric, and the offset values are not reported. That means the EM numbers in Table 4 are fit statistics, not prediction accuracies. The improvement from SK (0.558) to E (0.628) could easily be an artifact of aligning and tuning to the reference. The joint-angle results are not a saving grace: Table 5 shows model SK has higher R2 than model E on five of six joints, so the soft-foot model does not improve kinematic fidelity.\n\nNone of this means the approach is wrong. The idea is plausible and the integration is a legitimate contribution for rehabilitation robotics. But the evaluation protocol needs to change. The authors should hold out one or more gait cycles (or subjects) for tuning, report the offsets explicitly, and ideally release code and data so others can reproduce the pipeline. With a blind evaluation, the paper would be solid.\n\nWorth a serious referee. Send it to peer review, ask for revisions. It is not a desk reject, but the abstract's claim of dynamic accuracy is currently overstated.","headline":"Useful MuJoCo integration for personalised soft-body feet, but the main GRF result is partly an artifact of tuning and offsets applied to the same data used for evaluation.","tokens_in":9550,"tokens_out":1940,"would_cite":false,"duration_ms":44776,"reading_group":"maybe","serious_thinker":"yes","would_accept_peer_review":true},"rs_alignment":null,"lean_confirmation":null,"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"The paper claims that adding personalised soft-body feet to a skeletal human model lets a walking controller trained only on kinematic motion capture produce ground reaction forces comparable to measured data.","keywords":["3D human digital twin","soft-body simulation","MuJoCo flex","ground reaction force","walking control policy","motion capture","joint angles","SMPL"],"falsifier":"Run model E with its fixed stiffness, damping, and foot position on a second subject's motion capture without re-tuning any parameters; if the vertical ground-reaction-force experimental match falls below the bare-skeleton baseline, the claim that personalised soft feet generate realistic forces from kinematic-only training would be refuted. A simpler check is to re-evaluate model E on the same subject without applying the reported phase offset and record how much the experimental match drops.","tokens_in":8481,"feed_emoji":"🦶","tokens_out":5126,"duration_ms":47833,"temperature":0.7,"pith_summary":"The paper claims that a personalised 3D human digital twin, built from motion capture data, can produce dynamically realistic walking in simulation when soft-body feet are added to a skeletal model. The walking control policy is trained only on kinematic motion capture, yet the soft feet generate vertical ground reaction forces comparable to force-plate measurements, with the best model reaching an average experimental-match score of 0.628 versus 0.558 for the bare skeleton. The authors argue this shows that simplified deformable feet can carry the contact dynamics needed for faithful human-robot interaction simulations without requiring force data during training. If true, this opens a route to personalised digital twins for studying assistive robots and rehabilitation scenarios using only motion capture of individual subjects.","feed_headline":"Soft-body feet let a digital twin walk with realistic ground forces","feed_subtitle":"Hyperelastic foot meshes let a kinematic-only walking controller reproduce measured ground reaction forces.","key_machinery":"The central object is the MuJoCo flex object: a deformable mesh simulated with hyperelastic material properties, here a 15-vertex simplified foot generated from an SMPL body shape and attached to the skeletal model by hard pins, chosen because welds and connects are soft contacts that caused oscillation. The mechanism is that the flex foot deforms under ground contact and generates the ground reaction force; tuning its stiffness (2000 to 12000) and damping (100) plus shifting foot position lets the deformation produce a smooth, M-shaped force curve without the skeleton itself touching the floor.","core_discovery":"On the paper's own terms, the central discovery is that a uniform, 15-vertex hyperelastic foot mesh pinned to a skeleton is sufficient for a kinematic-only trained walking policy to reproduce the measured double-bump vertical ground reaction force profile. By stiffening the foot material from model A through model E and adjusting foot position, the authors eliminate skeleton-floor contact spikes and bring joint angles closer to the reference, with model E outperforming the bare skeleton on the experimental-match metric. The paper attributes residual phase differences between simulated and measured forces to shape simplification of the foot, and applies offsets when comparing curves.","pith_inferences":["The paper leaves open whether the phase offsets applied when comparing simulated to measured ground reaction forces are legitimate gait-phase alignment or a post-hoc correction; removing the offsets and reporting the drop in experimental match would clarify how much of the result comes from the model itself.","The stiffness and foot position were tuned on the same subject's force data used for evaluation, so a held-out subject test would show whether the personalisation procedure generalises or merely fits.","Because only the left foot had enough complete recordings to form a reliable standard deviation, the force comparison is one-footed; extending the capture protocol to the right foot would verify the claim for both limbs.","The 15-vertex simplification is likely the main source of the phase offset and the initial gradual rise in force, so increasing mesh resolution while keeping the material model fixed could test whether the offset shrinks."],"forward_implications":["If the claim is correct, personalised digital twins can be generated from motion capture alone and used to predict ground reaction forces in scenarios where force plates are not available.","The same soft-foot approach could be transferred to other skeletal or musculoskeletal models in the same physics engine, since only the foot contact needs to be replaced.","Stiffness tuning of the soft foot becomes a practical lever for matching contact dynamics: too soft feet cause intermittent skeleton-floor contact and force spikes, while too stiff feet approximate the bare skeleton.","The reported improvement in experimental match from 0.558 to 0.628 suggests that soft-body contact can add dynamic realism on top of kinematic fidelity without changing the control objective."],"supporting_citations":[{"why":"Supplies the SMPL parametric body model from which the personalised foot meshes are derived.","marker":"[13]"},{"why":"Provides the MuJoCo documentation defining flex objects and the pin, weld, and connect contact options used to attach the soft feet.","marker":"[15]"},{"why":"Describes the existing skeletal walking control policy pipeline that the soft-foot models extend and benchmark against.","marker":"[12]"},{"why":"Supplies the motion capture database providing the subject's walking data and force plate recordings.","marker":"[11]"},{"why":"Introduces the experimental-match metric for ground reaction force comparison and the precedent for phase-offset issues.","marker":"[20]"},{"why":"Provides the linear fit method used for joint angle waveform similarity analysis.","marker":"[21]"},{"why":"Describes the pipeline for building the personalised skeletal model from motion capture data.","marker":"[10]"},{"why":"Supports the viability of hyperelastic material models for approximating soft human tissue.","marker":"[9]"},{"why":"Provides ranges of human soft tissue material properties used to set the flex foot parameters.","marker":"[14]"}],"fun_headline_variants":["Soft-body feet give digital twins realistic ground forces","Hyperelastic foot mesh fixes walking simulation forces","Digital twin walks softer with flexible feet","Soft feet let digital twin match real ground forces","Uniform hyperelastic feet improve walking controller forces"],"cache_read_input_tokens":3200,"weakest_assumption_plain":"The load-bearing premise is that a 15-vertex, uniformly elastic foot pinned to the skeleton behaves like a real human foot in contact, and that the stiffness increases and phase offsets used during evaluation are legitimate corrections rather than tuning to the measured force data.","fun_headline_variants_meta":{"raw":{"variants":["Soft-body feet give digital twins realistic ground forces","Hyperelastic foot mesh fixes walking simulation forces","Digital twin walks softer with flexible feet","Soft feet let digital twin match real ground forces","Uniform hyperelastic feet improve walking controller forces"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000472,"raw_usage":{"total_tokens":2281,"prompt_tokens":814,"completion_tokens":1467,"prompt_tokens_details":{"cached_tokens":384},"prompt_cache_hit_tokens":384,"prompt_cache_miss_tokens":430,"completion_tokens_details":{"reasoning_tokens":1400}},"tokens_in":430,"tokens_out":1467,"duration_ms":10735,"temperature":1.0,"reasoning_tokens":1400,"cache_read_input_tokens":384,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-12T14:59:57.031501+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Run model E with its fixed stiffness, damping, and foot position on a second subject's motion capture without re-tuning any parameters; if the vertical ground-reaction-force experimental match falls below the bare-skeleton baseline, the claim that personalised soft feet generate realistic forces from kinematic-only training would be refuted. A simpler check is to re-evaluate model E on the same subject without applying the reported phase offset and record how much the experimental match drops.","supporting_citations":[{"cited_title":"MuJoCo Documentation: Overview,","cited_arxiv_id":null,"evidence_quote":"Provides the MuJoCo documentation defining flex objects and the pin, weld, and connect contact options used to attach the soft feet."},{"cited_title":"Assessment of Waveform Similarity in Clinical Gait Data: The Linear Fit Method,","cited_arxiv_id":null,"evidence_quote":"Provides the linear fit method used for joint angle waveform similarity analysis."},{"cited_title":"Investigation of Modeling Differences between OpenSim and Visual3D for Gait Analysis of Healthy Gait,","cited_arxiv_id":null,"evidence_quote":"Describes the pipeline for building the personalised skeletal model from motion capture data."}],"review_version":1}