Chapter Four · failure evidence
What Biomechanical Modeling got wrong, from 40 dissertations
Biomechanical modeling efforts frequently fail when simplified kinematic, rigid-body, or muscle representations cannot capture complex dynamic human movements. These failures span simulation instability, loss of tracking compliance, poor cross-task generalization, and estimation inaccuracies compared to empirical baseline measurements. These records come from PhD theses at 15 institutions, 2021 to 2026. Each links to its thesis. They were extracted by language models reading the full text, so treat each as a lead to read, not a verdict.
Wearable sensor and kinematic estimation methods suffer from movement variability and perform worse than standard baselines
Inertial and kinematic estimations frequently degraded joint force predictions, motion tracking, and event detection due to stride variability, segment bias, and motion artifacts. Heuristic and empirical sensor models regularly underperformed compared to direct optical baselines, machine learning models, and true physical object endpoints.
Tried and failed
kinematic inertial sensor estimation applied to ground reaction force estimation. Outcome: worse than baseline. Reason: high stride-by-stride kinematic variability and non-sagittal compensatory movements degraded accuracy relative to direct force sensing
Tried and failed
conditioning models on observed behavioral responses applied to neural motion prediction models. Outcome: worse than baseline. Reason: empirical response endpoints provided worse cross-validated likelihood than true physical object endpoints
Mechanisms of Multi-Object Working Memory and Motion Prediction in the Primate Brain · MIT
Lost to a baseline
Individually-fit variable-similarity motion mapping did not achieve statistically significant performance improvements over the population-fit variable-similarity motion mapping across trajectory metrics (such as minimum jerk distance, peak speed, and robot path inefficiency).
Increasing Transparency and Presence of Teleoperation Systems Through Human-Centered Design · Penn
Lost to a baseline
Heuristic rules-based algorithm achieved lower overall gait event identification accuracy across locomotion modes (94.87%) compared to the unsupervised BP-AR-HMM (99.6%).
Machine Learning and Wearable Sensors for the Estimation of Biomechanical Variables Outside the Laboratory · Scholars' Bank
Lost to a baseline
IMU-based musculoskeletal models underestimated L5/S1 anteroposterior and mediolateral shear forces by 31.1% and 52.6% compared to optical motion capture baseline.
Tried and failed
kinematic model calibration using subject data applied to joint angle estimation in motion capture. Outcome: worse than baseline. Reason: introduced significant systematic bias across segments instead of improving ground truth agreement
Tried and failed
multimodal classification omitting primary kinematic modality applied to repetitive behavior classification. Outcome: worse than baseline. Reason: physiological signals lacked sufficient discriminative power without dominant motion features
Explainable and Robust Data-Driven Machine Learning Methods for Digital Healthcare Monitoring · Virginia Tech
Lost to a baseline
Saeed et al. (SS, FT) scored 0.939 weighted F1 on MotionSense, beating the thesis's motion prediction SS, FT method at 0.934 F1.
Self-supervised Learning for IMU-based Human Activity Recognition · Queens University Institutional Repository
Considered and rejected
Considered and rejected: Rejected naive peak detection for stride segmentation under exoskeleton pulsing due to motion artifacts and tissue compliance, adopting genetic algorithm optimization.
Effects of Mechanical Interventions on Human Locomotion · MIT
Musculoskeletal and muscle dynamics formulations fail to reproduce realistic physiological forces and activations
Simulations omitting muscle dynamic properties or applying non-linear Hill parameterizations produced unstable gaits, excessive reserve forces, and worse error than linear baselines. Furthermore, analytical dynamic inversions and biophysical muscle models could not handle redundant non-linear activations or eccentric contraction conditions.
Tried and failed
reducing muscle fiber length to simulate contracture applied to musculoskeletal gait simulation. Reason: simulated gait became unstable and fell before producing toe gait kinematic patterns
Tried and failed
Active muscle volume modeling applied to locomotion metabolic energy expenditure prediction. Outcome: did not generalise. Reason: Failed to predict metabolic cost and muscle activation differences across varying mechanical footwear conditions
Long-term Effects of Altered Foot-Ankle Mechanics on Locomotion Neuromechanics and Energetics · Georgia Tech
Tried and failed
musculoskeletal simulation omitting muscle force-length-velocity dynamics applied to biomechanical gait simulation. Outcome: unstable. Reason: neglecting dynamic muscle properties caused reserve actuator forces to exceed acceptable validity thresholds
Predicting lower limb kinematics and kinetics from internal measurement units using deep learning · Imperial
Considered and rejected
Considered and rejected: Rejected computing control inputs via continuous analytical inverse dynamics because redundant, non-linear activation dynamics and ODE-based muscle contraction models prohibit direct mathematical inversion.
Neuro-musculoskeletal Models: A Tool to Study the Contribution of Muscle Dynamics to Biological Motor Control · Publikationssystem UB Tuebingen
Considered and rejected
Considered and rejected: Rejected using microscopic/biophysical Huxley-type muscle models with explicit damping because they cannot predict mechanical response during eccentric contractions.
Tunable Mechanical Damping in Fast Perturbed Locomotion: Biomechanical Simulations and Biorobotic Applications · Publikationssystem UB Tuebingen
Tried and failed
non-linear Hill muscle model parameterisation applied to cutting speed biomechanics modeling. Outcome: worse than baseline. Reason: produced worse RMSE compared to a simpler linear model
The biomechanics and energetics of foraging in leaf-cutter ants · Imperial
Tried and failed
Optimization-based musculoskeletal simulation with Hill-type models applied to shoulder muscle activation during overhead tasks. Outcome: did not generalise. Reason: Poor pattern similarity predicting posterior deltoid activation across varied task elevations.
Evaluating Model-Estimated Shoulder Muscle Activity During Overhead Work with Varied Task Demands and Exoskeleton Use · Virginia Tech
Locomotion control and movement generation schemes fail due to rigid tracking or oversimplified state representations
Rigid high-gain tracking and low-dimensional reinforcement learning controllers restricted natural biomechanical variability, induced user discomfort, or failed to converge on whole-body coordination. Enforcing artificial curvature boundaries and unconstrained physics simulations similarly yielded energetically suboptimal postures or excessive computational expense.
Tried and failed
passive mechanical design for functional locomotion applied to transfemoral prosthetic gait adaptation. Outcome: did not generalise. Reason: fixed mechanical properties could not accommodate high-speed walking demands for all users
Tried and failed
rigid high-gain trajectory tracking without compliance applied to human gait assistance control. Outcome: worse than baseline. Reason: Restricted natural biomechanical variability and movement patterns, causing severe user discomfort.
Tried and failed
model-free reinforcement learning with reduced-order state applied to bipedal locomotion control. Outcome: did not converge. Reason: simplified low-dimensional state representation lacked sufficient dynamics information for whole-body motion synthesis
Towards versatile, high-performing and interactive humanoids · UT Austin
Tried and failed
trajectory optimization with non-zero initial boundary curvature applied to undulatory locomotion gait generation. Outcome: worse than baseline. Reason: enforced non-zero boundary curvature produced energetically suboptimal intermediate S-shaped body configurations
Gait optimality for undulatory locomotion with applications to C. elegans phenotyping · Imperial
Considered and rejected
Considered and rejected: Rejected non-hierarchical/flat reinforcement learning (E2E and PMTG) for gait discovery due to suboptimal convergence (standing still or 3-legged locomotion).
Agile Legged Robots through Reinforcement Learning and Optimal Control · ResearchWorks
Tried and failed
motor averaging model applied to obstacle avoidance reaching under target uncertainty. Outcome: worse than baseline. Reason: failed to predict movement direction deflections compared to performance optimization models
Understanding How Environmental Dynamics and Uncertainty Affect Adaptive Changes in Motor Planning · Harvard
Considered and rejected
Considered and rejected: Rejected dynamics-based/physics-based simulation modeling in favor of a kinematic model due to dynamic simulation being computationally costly (limiting interactivity) and lacking subtle nuances of expressive human movement.
A Motion Control Scheme for Animating Expressive Arm Movements · Penn
Models trained on constrained or low-dimensional tasks fail to generalize to dynamic multi-axis movements
Machine learning architectures trained on symmetric or single-axis exercises failed when applied to free dynamic movements with simultaneous multi-axis kinematics. Similarly, aligning pathological gaits to normative references or predicting outcomes across individuals broke down due to idiosyncratic mechanics and inconsistent joint kinematics.
Tried and failed
latent space alignment to normative reference profile applied to pathological gait recovery assessment. Outcome: did not generalise. Reason: normative latent representation movement negatively correlated with specific localized functional joint kinematics
Gait signatures: data-driven discovery of individual-specific neuromechanical dynamics · Georgia Tech
Tried and failed
cross-task machine learning model training applied to muscle fascicle length estimation during movement. Outcome: did not generalise. Reason: varying mechanical loads and ranges of motion across different dynamic movement tasks
Machine Learning and Biomechanical Sensing Toward Real-Time In-The-Loop Gait and Joint Health Optimization · Georgia Tech
Tried and failed
training on constrained tasks to predict unconstrained tasks applied to time-series activity classification. Outcome: did not generalise. Reason: constrained symmetric training data lacked kinematic variability present in free-dynamic movements
Armband EMG-based Lifting Detection and Load Classification Algorithms using Static and Dynamic Lifting Trials · Virginia Tech
Tried and failed
training on single-axis motions for multi-axis prediction applied to continuous kinematic regression from myoelectric signals. Outcome: did not generalise. Reason: isolated single-degree-of-freedom training data failed to represent muscle activation patterns during combined simultaneous movements
Considered and rejected
Considered and rejected: Predicting shot outcome from shooter body kinematics was rejected due to idiosyncratic movement across individuals producing weak predictive accuracy.
Basketball Shooting as a Model System for Understanding Skill Learning and Motor Variability · Harvard
Rigid-body assumptions fail to represent non-rigid anatomical segments and tissue deformation
Treating complex anatomical structures like the foot or ventral nerve tissues as single rigid segments violated underlying non-rigid kinematics and multi-column motion. These assumptions caused instantaneous center of rotation misalignments and could not capture necessary non-rigid tissue deformations.
Considered and rejected
Considered and rejected: Rejected using human-subject motion analysis/inverse dynamics to derive foot stiffness due to high individual gait asymmetry confounding, non-rigid body segment violations, and instantaneous center of rotation misalignment.
Considered and rejected
Considered and rejected: Rejected treating tarsals and metatarsals as single rigid segments (as in standard KU-Leuven model) because it overlooks independent medial/lateral column kinematics and relative metatarsal motions.
Strain estimations of the plantar fascia and other ligaments of the foot: Implications for plantar fasciitis · Iowa State
Considered and rejected
Considered and rejected: Standard parametric motion models (e.g., affine transformations) for VNC image registration; rejected because they could not capture complex non-rigid tissue deformations.
Left open by the authors
Problems the authors named and did not get to.
Left open
Model mechanical interactions driving biophysical movements in sea-urchin gastrulation and zebrafish epiboly using multi-node lateral vertex simulations. Blocker: Lacks specific biological parameter values and formal model adaptations for sea-urchin and zebrafish morphology.
Cellular Mechanics of Cephalic Furrow Formation in the Drosophila Embryo · Texas Tech
Left open
Extend the prosthetic foot design framework to model multi-planar motion such as inversion/eversion for uneven terrain and squatting. Blocker: No concrete 3D kinematic/biomechanical models or optimization formulation specified
Development and Validation of a Passive Prosthetic Foot Design Framework based on Lower Leg Dynamics · MIT
Left open
Improve cross-subject generalization for stance/swing phase duration prediction and frontal plane ankle kinematics/torque. Blocker: Requires multi-sensor HDEMG and biomechanical locomotion data across multiple subjects or a powered lower-limb prosthesis testing apparatus.
A multi-sensor, hybrid model- and signal-based control system for powered lower limb prostheses · Imperial
Left open
Implement Bayesian deep learning networks on multi-IMU gait data to predict distributions of foot placements with uncertainty estimates. Blocker: Lacks access to the multi-IMU dataset and motion capture ground truth collected for the thesis.
A Deep-learning based Approach for Foot Placement Prediction · Virginia Tech
Left open
Simulate muscle contributions to balance control across varied step widths during unperturbed steady-state walking using OpenSim musculoskeletal models. Blocker: Requires kinematic, kinetic, and EMG gait data across varied step widths or dynamic predictive simulation setup.
Fall detection and balance control during human movement · UT Austin
Left open
Train neural networks on simulated or public gait data to predict muscle and joint reaction forces, bypassing musculoskeletal modeling. Blocker: None
Predicting lower limb kinematics and kinetics from internal measurement units using deep learning · Imperial
Left open
Investigate correlations between center of mass position, limb musculature reconstructions, and locomotion implications across synapsid taxa. Blocker: Lacks specific target taxa, anatomical muscle reconstruction models, and formalized biomechanical framework
Reconstructing Body Size and Center of Mass in Synapsids · Harvard
Left open
Develop biomechanical demand estimation models using only markerless motion capture kinematic data, eliminating the need for insole pressure measurement kinetics. Blocker: Requires paired markerless motion capture kinematic data and ground-truth biomechanical kinetics/exposures.
Evaluation of Markerless Motion Capture to Assess Physical Exposures During Material Handling Tasks · Virginia Tech
Left open
Compute lower-limb joint quasi-stiffness from perturbation walking kinematics and kinetics data and compare it to steady-state gait. Blocker: Requires experimental biomechanical gait perturbation data (motion capture and force plates) or specialized perturbation simulations.
An assessment of lower-limb joint quasi-stiffness during healthy and transtibial amputee walking · UT Austin
Left open
Model the effects of aging on lumbar spine tissue loading and biomechanics using modified musculoskeletal and finite element models. Blocker: Lacks specific aged-subject motion capture datasets and age-degraded tissue material properties to parameterize the models
Multi-scale Modeling of the Lumbar Spine During Manual Material Lifting Tasks · Texas Tech
Checking a claim in this area?
We can run the same search on any method or claim. If nothing turns up, we will say so, and that proves nothing on its own.