Biomechanically Accurate Skeletons

• Biomechanically accurate skeletons are digital anatomical models designed to closely replicate real skeletal geometry, joint structure, and biomechanics.
• They employ advanced imaging modalities, precise landmark annotation, and kinematic modeling to ensure anatomical fidelity for simulations and clinical applications.
• Automated mesh processing and learning-based methods enable scalable reconstruction and validation, improving computational efficiency and biomechanical realism.

Biomechanically accurate skeletons are digital anatomical models or mesh representations in which the geometry, joint definition, and structural parameters of bones are constructed or inferred to closely replicate real biological skeletal architecture and articulation. Such skeletons are foundational for advanced biomechanical simulation, clinical modeling, robotics, motion analysis, computer graphics, and learning-based pose estimation. Research in this area encompasses methodologies for generating high-resolution meshes from imaging, regressing skeletal structure from external morphology, calibrating joint parameters against experimental data, and validating biomechanical plausibility in simulated and real-world tasks.

1. Anatomical Data Acquisition and Mesh Construction

Biomechanically accurate skeletal models fundamentally depend on the quality and resolution of anatomical input data. High-fidelity bone meshes are commonly derived from cross-sectional medical imaging modalities, such as computed tomography (CT) or X-ray scans. For example, the construction of detailed human lower body meshes began with manual segmentation of CT images from a 59-year-old female cadaver, with the CT pixel size at 0.33 mm and inter-slice spacing at 1.0 mm (Sreenivasa et al., 2018). Segmentation was performed using MITK, followed by mesh generation and refinement via Matlab and the Iso2Mesh toolbox.

Salient anatomical features—including regions of high curvature at joints—are preserved during processing, while mesh simplification is applied in morphologically simple bone shafts for computational efficiency. Strategic mesh resampling and smoothing (conceptually analogous to low-pass filtering) optimize the trade-off between geometric detail and mesh complexity. Meshes can be linearly rescaled using anthropometric regression, such as De Leva’s equations, via transformations of mesh vertex coordinates: $$\text{ScalingFactor} \propto \text{Subject\_Height} / \text{Reference\_Height}$$ This workflow ensures that digitally reconstructed skeletons remain scalable and adaptable to various subject dimensions or simulation needs.

2. Landmark Annotation, Kinematic Modeling, and Joint Axes

A key criterion for biomechanical fidelity is the annotation of bony landmarks and the precise geometric definition of joint axes. Landmark sets—such as the femoral head, iliac crest, condyles, and malleoli—are annotated on bone meshes following international biomechanical conventions (e.g., Wu et al. for the hip/ankle, Grood and Suntay for the knee) (Sreenivasa et al., 2018). These landmarks enable explicit determination of joint centers and serve as anchor points for modeling joint coordinate systems and articulatory degrees of freedom (DoF).

A rigorous kinematic skeleton assigns rotation axes orthogonally at joint locations, and the mesh is structured to reflect realistic joint motion ranges. For example, a lower body kinematic model may be constructed with 12 internal DoF (distributed across the hip, knee, and ankle) and 6 global DoF at the sacrum, supporting full 3D articulation and physiological motion limits. In more advanced work, constraints on joint motion may also be empirically embedded, for example by incorporating observed kinematic coupling or path-dependent center of rotation migration, as seen in empirically-based multibody dynamics (Ehsani, 2023).

3. Mesh Processing Frameworks and Automation

To render biomechanically accurate skeletons broadly usable, script-based frameworks with graphical interfaces streamline mesh visualization, resampling, and subject-specific scaling. These toolkits enable researchers to interactively inspect anatomical detail, adapt mesh complexity to task requirements, and export models in standard formats (OBJ, STL) or kinematic definitions (Lua) (Sreenivasa et al., 2018). Automated pipelines for mesh processing are critical for generating subject-adaptive, simulation-ready skeletons without extensive manual intervention.

Recent extensions include probabilistic and data-driven mappings from body surface meshes to internal skeletal structures. For example, OSSO (Keller et al., 2022) learns a regression from STAR body shape parameters to a pose-normalized skeleton subspace, validated on large paired datasets of skin and bone meshes derived from DXA scans. This allows for generalization to new subjects and the creation of anatomically coherent internal skeletons solely from surface geometry.

4. Empirical Modeling, Validation, and Biomechanical Rigour

True biomechanical accuracy extends beyond geometric plausibility to include empirical calibration, experimental validation, and biomechanical constraint enforcement. For musculoskeletal simulations—including pediatric spine modeling (Schmid et al., 2019) or hand biomechanics (Tasi et al., 2019)—models are built by calibrating segmental lengths, inertias, mass distributions, and muscle-tendon properties to literature or population databases, with age/gender dependencies and alignment angles adjusted based on normative data.

Model validation proceeds by simulating experimental tasks and comparing outputs (joint torques, muscle forces, disc compressibility, intradiscal pressure, etc.) against in vivo or in vitro datasets. Statistical metrics such as Pearson’s r (e.g., r ≥ 0.82 for trunk muscle strength; r = 0.77 for disc compressibility (Schmid et al., 2019)) confirm the suitability of these models for clinical and biomechanical prediction.

Zoo of approaches for embedding joint biophysics and empirical kinematics include:

• Direct embedding of experimental joint rhythms (e.g., shoulder motion coupling (Ehsani, 2023)).
• Enforcement of physiological joint limits and articulation patterns during pose estimation, as in the SKEL model and its ViT-based HSMR framework (Keller et al., 8 Sep 2025, Xia et al., 27 Mar 2025).
• Reparameterization techniques that align 3D kinematic sequences (on SE(3)) using biomechanically-informed cost functions, as in GORA-S (Mitchel et al., 2018).

5. Data-Driven and Learning-Based Advancements

Contemporary approaches leverage learning-based methods for inferring biomechanically plausible skeletons from limited or indirect data, such as RGB images, silhouettes, or external meshes. Regression and deep learning frameworks (e.g., OSSO, BioPose, EA-RAS) have demonstrated the ability to infer internal bone structures from outer skin/mesh geometry (Keller et al., 2022, Koleini et al., 14 Jan 2025, Peng et al., 3 Sep 2024). These models typically operate in a multi-stage fashion:

• Fit a surface parametric model (e.g., STAR or SMPL) to the observed data.
• Regress or optimize internal skeleton parameters using a statistical shape model or dedicated regressor, mapping body shape coefficients to bone shape coefficients.
• Impose biomechanical constraints via energy minimization (e.g., alignment of inferred joints with surface-derived pose or preservation of inter-bone relationships).

BioPose integrates multi-query transformers and neural inverse kinematics to recover fine-grained anatomical pose from monocular imagery, using virtual markers sampled on the mesh as mediators for pose refinement. Anatomically detailed models such as SKEL re-rig the surface mesh with a physics-inspired skeleton, reducing the number of DoF to those present in human articulation (46 for SKEL vs. 72 in unconstrained SMPL) and enforcing joint-specific limits (Keller et al., 8 Sep 2025).

Single-stage, real-time approaches such as EA-RAS combine dual-branch regression (for body and bone) with lightweight optimization to offer rapid and anatomically realistic skeleton inference, achieving per-subject reconstruction errors on the order of tens of millimeters and speed improvements over traditional staged pipelines by two or three orders of magnitude (Peng et al., 3 Sep 2024).

6. Applications, Accessibility, and Impact

Biomechanically accurate skeletons are foundational across domains:

• Computational biomechanics: for subject-specific musculoskeletal simulations, torque and muscle force estimation, and range of motion analysis.
• Clinical application: for movement disorder analysis, surgical planning, design of orthopedic interventions, and educational visualization.
• Computer vision and animation: enabling plausible character articulation, robust pose estimation, and data augmentation for vision models; facilitating high-fidelity rigging and skinning in arbitrary mesh/skeleton configurations (Hong et al., 17 Mar 2025).
• Robotics and human–robot interaction: for simulating physically plausible human movement, action recognition, and teaching via imitation.

Open-data resources and script repositories, such as the BMFToolkit (Sreenivasa et al., 2018), BioAMASS and SKEL (Keller et al., 8 Sep 2025), and OSSO (Keller et al., 2022), are made publicly accessible to foster reproducibility and adaptation to particular research contexts.

7. Limitations, Generalization, and Future Directions

Despite recent advances, several limitations persist:

• Imaging-based approaches remain constrained by the quality, modality, and population variety of underlying data. The representativity of female, pediatric, or non-standard morphologies continues to be a priority (e.g., the explicit focus on an adult human female lower body (Sreenivasa et al., 2018)).
• Mapping from skin/mesh to internal bone structure is inherently ambiguous and sensitive to pose, occlusion, and clothing; further development of multi-modal data integration (e.g., RGB + depth, multi-view) is required (Peng et al., 3 Sep 2024).
• Muscle actuation, passive joint property modeling (stiffness, damping), and dynamic simulation in small animals or non-human morphologies, as in Drosophila (Özdil et al., 8 Sep 2025), require further adaptation of anatomical priors and parameter optimization pipelines.
• Enforcement of joint constraints and articulation within neural models is still under active investigation, with empirical and physics-inspired priors showing potential for bridging kinematic realism and learning-based estimation (Keller et al., 8 Sep 2025, Xia et al., 27 Mar 2025).
• Real-time pipelines balancing anatomical accuracy and computational efficiency are being improved via plug-and-play, self-supervised, and regressor-optimization hybrid frameworks.

The field is moving toward integrating physics-aware modeling, empirical validation, and scalable learning to enable biomechanics "in the wild"—bridging detailed anatomical fidelity with the needs of computational efficiency and generalization across diverse populations and applications.