Neural-Network Joint–Muscle Mapping (NN-JMM)

Updated 19 May 2026

Neural-Network Joint–Muscle Mapping (NN-JMM) is a computational approach that uses deep neural networks with embedded physiological constraints to map neural and kinematic signals to musculoskeletal states.
It integrates data-driven learning with physics-informed losses and multiscale transfer strategies, achieving significant accuracy improvements in joint kinematics and muscle force predictions.
The method supports applications in real-time biomechanics, prosthesis control, and robotics, while ongoing research addresses scalability, personalization, and interpretability challenges.

Neural-Network Joint–Muscle Mapping (NN-JMM) encompasses a class of computational approaches that learn nonlinear mappings between neural, kinematic, and musculoskeletal states using deep neural networks, often with embedded physiological structure and constraints derived from musculoskeletal dynamics. This paradigm enables the prediction or estimation of joint kinematics, muscle states (activations, forces, lengths), and physiological parameters from multimodal sensor input such as sEMG, IMC, motion capture, or even ultrasound. Recent advances emphasize the fusion of data-driven learning with physics-informed losses, multiscale transfer-learning, and label-free or low-shot learning pipelines, establishing NN-JMM as a foundational methodology in computational biomechanics and neural engineering.

1. Formal Problem Definition

NN-JMM refers to a family of mappings—parametrized by neural networks—between surface or intracortical neural signals, joint kinematics, and musculoskeletal states. Depending on the specific task and experimental design, key formalizations include:

Forward mapping: From muscle activations (e.g., sEMG) to joint kinematics and muscle/tendon forces. For example,

$\mathcal{F}_\theta: \left\{ u(t) \in \mathbb{R}^{N_{a}} \right\}_{t=1}^T \longrightarrow \left\{ q(t) \in \mathbb{R}^{N_{\mathrm{dof}}}, \; F(t) \in \mathbb{R}^{N_{a}}, \; p \in \mathbb{R}^{N_p} \right\}$

as realized in MR PI-RNN (Taneja et al., 2023).

Inverse mapping: From joint kinematics (and optionally joint torques/forces) to muscle activations/forces, typically under-actuated and highly redundant (Ma et al., 2024).
Direct muscle-to-joint/force mapping: From time windows of sEMG envelopes to muscle forces and joint angles via fully connected or convolutional architectures (Zhang et al., 2022, Shi et al., 2023).
Sensor fusion or cross-modal prediction: From unconventional signals (e.g., 2D ultrasound image pairs) to drift-free estimation of EMG, joint angles, and joint moments (Cunningham et al., 2019), or from IMC to OMC-driven MSK outputs (Dasgupta et al., 2022).

Underlying these mappings is a neural representation $\mathcal{F}_\theta$ that must be expressive enough to capture nontrivial biomechanical relationships, but also constrained to ensure physiological plausibility, often via embedded ordinary differential equations or analytical muscle models.

2. Core Architectures and Physics-Informed Embedding

NN-JMM methods implement a diverse range of neural architectures, selected based on input/output modalities, data regime, and real-time constraints.

Recurrent networks: Deep LSTM, GRU, BiGRU, and hybrid RNN-convolutional generators are employed for sequence-to-sequence prediction using time history of EMG or kinematics (Schmidt et al., 2022, Shi et al., 2023, Ma et al., 2024, Taneja et al., 2023). These capture temporal dependencies intrinsic to muscle activation, contraction, and motion.
Feed-forward and convolutional models: For static mappings or short time windows, deep fully connected networks and 1D/2D CNNs perform feature extraction (e.g., for sEMG or US images) followed by regression of muscle or joint states (Shi et al., 2023, Zhang et al., 2022, Cunningham et al., 2019, Ma et al., 2024, Kumar et al., 2023).
GANs with physics-informed policy gradients: Generative models incorporate Lagrangian mechanics as soft constraints on output plausibility, guiding feature decoding and improving performance under low-shot regimes (Shi et al., 2023).
Multi-resolution and transfer-learning pipelines: Wavelet decomposition of mixed-frequency signals enables scale-wise sequential training via transfer of network parameters, improving generalizability and denoising robustness (Taneja et al., 2023).
Biological interpretability: Some models impose structure consistent with muscle synergies (low-dimensional modulatory bottlenecks), or separately model direct versus synergistic neural pathways in mapping from TMS-induced fields to multimuscle EMG (Akbar et al., 2020).

Central to recent advances is the embedding of physics-informed constraints as differentiable loss terms, which may include:

Forward-dynamics ODE residuals enforcing that predicted kinematics and forces satisfy musculoskeletal equations of motion.
Hill-type muscle contraction models coupling activation, length, velocity, and force.
Static-optimization terms penalizing physiologically non-plausible activations or forces.
Moment-arm and geometric consistency, capturing the effects of musculoskeletal architecture.

3. Training Procedures, Data Regimes, and Parameter Identification

NN-JMM frameworks are designed to accommodate various data regimes, loss formulations, and personalized parameter estimation.

Label-free and weakly-supervised learning: By embedding physics-informed constraints, NN-JMM models can be trained with little or no ground-truth labels for muscle forces or activations, relying instead on the minimization of joint trajectory error and residual dynamics (Ma et al., 2024, Zhang et al., 2022, Shi et al., 2023).
Low-shot and transfer-learning: Policy-gradient GANs and transfer-sequential pipelines promote extrapolation from small sample sets, essential for subject-specific or clinical deployment (Shi et al., 2023, Taneja et al., 2023, Schmidt et al., 2022).
Parameter identification: NN-JMM often treats key physiological muscle-tendon parameters (e.g., maximal isometric force, optimal fiber length, activation nonlinearity) as trainable variables. Constraints and automatic differentiation enable joint optimization for accurate identification within physiologically plausible bounds (Taneja et al., 2023, Shi et al., 2023, Ma et al., 2024).
Loss functions: Typically, the total loss is a weighted sum of data misfit (e.g., joint angle MSE), physics residuals, force/activation consistency, regularization (e.g., L1/L2, dropout), and in some cases "zero-line" scores or boundary penalties on output domains (Ma et al., 2024, Taneja et al., 2023, Schmidt et al., 2022, Akbar et al., 2020).

4. Quantitative Performance and Comparative Evaluation

NN-JMM methods consistently report high performance on both synthetic and empirical datasets, with robustness across held-out subjects, tasks, and experimental conditions. Representative metrics include RMSE (in joint angles and muscle forces), coefficient of determination ( $R^2$ ), zero-line scores, and symmetry-based percentage errors.

Accuracy improvements: Multiscale and physics-informed NN-JMM yield 50–70% reductions in MSE compared to single-scale or purely data-driven models, as well as significant increases in $R^2$ (e.g., 0.60 to ≈0.88 for joint angle prediction in elbow flexion-extension) (Taneja et al., 2023).
Label-free and data-efficient learning: Label-free NN-JMM performs comparably or superior to supervised deep networks trained on static optimization-derived labels, with prediction speed gains >5000× over classical optimization-based methods (Ma et al., 2024, Ma et al., 2024).
Generalization and robustness: Cross-session performance remains stable with minimal drop under kinematic variations, different subjects, motions, or noise (Taneja et al., 2023, Shi et al., 2023, Shi et al., 2023, Schmidt et al., 2022).
Muscle parameter identification: Physiologically consistent parameter identification is routinely achieved within literature-reported ranges (e.g., recovered muscle forces within <1% of ground truth) (Taneja et al., 2023, Shi et al., 2023).

Model/Study	Input/Output	Key Metric (Test)	Quantitative Result
MR PI-RNN (Taneja et al., 2023)	sEMG→Kinematics,p	MSE, $R^2$ (joint angle)	MSE: $2.3\times10^{-4}$ , $R^2\approx0.88$
BiGRU (Ma et al., 2024)	Kinematics→a,F	$R^2$ (a), $R^2$ (F)	Knee: $R^2$ (a/F): 0.96/0.93
GAN (Shi et al., 2023)	sEMG→F,θ	RMSE, $\mathcal{F}_\theta$ 0 (force)	Knee RMSE: 11.3 N, $\mathcal{F}_\theta$ 1: 0.88
FNN+Physics (Shi et al., 2023)	sEMG→F,θ,param	$\mathcal{F}_\theta$ 2 (angle/force)	$\mathcal{F}_\theta$ 3 angle: 0.96–0.99
PINN (Kumar et al., 2023)	sEMG→ $\mathcal{F}_\theta$ 4	$\mathcal{F}_\theta$ 5 (angle)	Elbow $\mathcal{F}_\theta$ 6: 0.72–0.95 (vs. ANN 0.15–0.67)

5. Applications and Extensions

NN-JMM serves as a foundation for a wide spectrum of applications across biomechanics, clinical diagnostics, robotics, and human–machine interfaces:

Human-in-the-loop control: Fast, physiologically plausible mapping for prosthesis and exoskeleton actuation, with explicit support for "label-free" closed-loop deployment (Ma et al., 2024, Shi et al., 2023).
Biomechanical analysis: Real-time estimation of unobservable MSK quantities (muscle/tendon force, activation, parameter identification) from wearable sensors (Taneja et al., 2023, Shi et al., 2023).
Robust musculoskeletal robotics: Adaptive online NN-JMM for tendon-driven humanoids, incorporating muscle redundancy, rupture detection, online mapping updates via vision and force sensing (Kawaharazuka et al., 2024, Kawaharazuka et al., 2024).
Low-shot and personalized MSK modeling: Physics-informed regularization decreases data requirements, enabling subject-specific mapping with limited calibration data (Shi et al., 2023, Schmidt et al., 2022).
Multimodal and cross-modal state estimation: Direct mapping from ultrasound images to joint and muscle states for non-invasive, generalizable diagnostics (Cunningham et al., 2019).
Physiology-driven model structure: Embedding explicit kinematic, muscle synergy, and anatomical mechanisms increases interpretability and transferability (Akbar et al., 2020, Shi et al., 2023).

6. Limitations and Future Directions

Despite substantial advances, several limitations and open research directions are evident:

Scalability and generalization: Many current NN-JMM frameworks are restricted to single-DOF joints or limited muscle groups; extension to multi-DOF, open-chain, or whole-body models is a critical next step (Taneja et al., 2023, Ma et al., 2024, Shi et al., 2023).
Personalization and minimal reliance on legacy software: Existing approaches often require OpenSim or similar to provide initial MSK geometry, moment-arm polynomials, or parameter ranges; fully data-driven geometric estimation and cross-modal transfer remain active areas of development (Zhang et al., 2022).
Dynamic, task-dependent loss weighting: Static loss weights may limit adaptation; meta-learning and curriculum loss-scheduling could further optimize training (Ma et al., 2024).
Model interpretability and biophysical plausibility: Embedding additional constraints (e.g., energy efficiency, injury/overuse criteria) or exploiting neural ODEs may improve interpretability and extend to regimes of non-stationary or disordered physiology (Shi et al., 2023).
Robust modularity to actuator/sensor failure: Autoencoder-based online learning and latent-space control paradigms enable robust function retention in the presence of muscle or tendon failure, but integration with distributed sensorimotor feedback is ongoing (Kawaharazuka et al., 2024).

A plausible implication is that future NN-JMM systems will tightly couple physics-based modeling, low-resource and multimodal learning, and online adaptation, under increasingly unconstrained, complex, and personalized biomechanical scenarios.