Abstract
Accurate estimation of muscle activation torque from surface electromyogram (sEMG) signals is critical for the intuitive control of assistive robotic devices such as exoskeletons. Traditional physics-based models are often too complex for real-time application, while pure machine learning approaches require prohibitive amounts of training data. This paper presents Forward-MuscleNET, a hybrid modeling framework that combines shallow neural networks with physics-based muscle torque generators. By offloading the modeling of musculoskeletal geometry to mathematical components, the system achieves high accuracy with minimal training data, specifically reaching a 0.1% Root Mean Squared Error (RMSE) in subject-specific applications.
Introduction
Controlling assistive robots requires a precise understanding of a user’s intended movement. This intention is typically captured via surface electromyogram (sEMG) signals, which measure the electrical activity produced by muscles during contraction. However, translating these stochastic electrical signals into joint torque—the actual force required to move a limb—presents a significant technical challenge.
Historically, researchers have relied on "Hill-type" muscle models. While biologically accurate, these models are mathematically complex and difficult to tune for the unique anatomical characteristics of individual users. Conversely, modern machine learning (ML) approaches often treat the human body as a "black box," requiring massive datasets of kinematic and kinetic measurements to learn the underlying physics. Forward-MuscleNET seeks a middle ground: a computationally efficient model suitable for real-time control that leverages known physical constraints to reduce the burden on the learning algorithm.
Methodology
The Forward-MuscleNET architecture functions as a serial mapping model between sEMG signals and activation torque. The innovation lies in its hybrid structure, which splits the task into two distinct stages:
- Neural Network Component: A feedforward neural network processes filtered sEMG signals from 11 channels located on the shoulder. The researchers determined that a "shallow but wide" architecture—specifically two layers with 100 neurons each—provided the optimal balance of performance, training speed, and low latency required for robotic assistance.
- Physics-Based Component (MTG/COMT): The predicted activation torque from the neural network is passed into a Muscle Torque Generator (MTG) or a Control-Oriented Muscle Torque (COMT) model. This component handles the "known" variables of human movement, including passive torque, torque-velocity scaling, and torque-angle scaling.
By using the MTG model to compensate for the position and velocity dependence of the musculoskeletal system, the neural network is freed from learning the laws of physics and can focus exclusively on the complex mapping of electrical signals to muscle activation.
Experimental Setup and Datasets
The study utilized a comprehensive dataset involving 17 subjects. Data collection focused on the shoulder complex and included:
- 11-channel sEMG signals to capture muscle activity.
- Kinematic data (position, velocity, and acceleration) via motion tracking.
- Kinetic data to measure external loads.
To train the network, "ground truth" activation torque was calculated using an inverse dynamic model of the human skeletal system combined with the MTG model. This allowed the researchers to isolate the specific torque generated by muscle activation from the total joint torque.
Results and Discussion
The performance of Forward-MuscleNET revealed critical insights into the nature of sEMG-based control:
- Subject-Specific Precision: When the model was trained and tested on data from a single individual, it achieved a remarkably low RMSE of 0.1% of the activation torque. This suggests that the hybrid approach is highly effective at capturing the nuances of an individual's neuromuscular patterns.
- The Generalization Challenge: When the researchers attempted to create a "universal" model by training across all 17 subjects, the error increased significantly to 23%. This "anatomical gap" highlights the sensitivity of sEMG signals to individual physiology and the precise placement of electrodes.
- Architectural Efficiency: The study confirmed that for real-time assistive devices, deeper is not necessarily better. Shallow networks with higher width outperformed deeper architectures in execution speed without sacrificing accuracy, making them ideal for low-latency exoskeleton control.
Conclusion
Forward-MuscleNET demonstrates that hybrid intelligence—combining the flexibility of neural networks with the reliability of physics-based models—is a viable path forward for assistive robotics. While the high error rate in general models indicates that subject-specific calibration remains necessary, the efficiency and accuracy of the personalized models provide a robust foundation for real-time control loops. Future work will focus on integrating Forward-MuscleNET into physical robotic exoskeletons to move from theoretical torque estimation to active, real-world assistance.
Summary updated