Control-Oriented Muscle Torque (COMT) Modelfor EMG-Based Control of Assistive Robots

Ali Nasr, John McPheeProceedings of the International Conference of Control, Dynamic Systems, and Robotics 2020 · DOI

Abstract

The development of intuitive control systems for assistive exoskeletons remains a significant challenge in rehabilitation robotics. Traditional Electromyography (EMG)-based control methods often struggle with the trade-off between biological accuracy and computational efficiency. This paper presents the Control-Oriented Muscle Torque (COMT) model, a hybrid framework that integrates simplified muscle physics with machine learning. By shifting the modeling perspective from anatomical frames to actuator-centric frames and utilizing Jacobian transformations, the COMT model reduces the data requirements for training while maintaining real-time performance capabilities.

1. Introduction

Assistive robots, such as exoskeletons designed for rehabilitation or physical augmentation, require a seamless interface to interpret human movement intent. Electromyography (EMG) signals, which record the electrical activity of muscles, serve as the primary input for these systems. However, translating these stochastic electrical signals into mechanical torque is a complex biomechanical problem.

Current approaches generally fall into two categories: complex Hill-type models that are biologically accurate but computationally expensive and difficult to parameterize, or "black-box" machine learning (ML) models that require massive, diverse datasets to generalize effectively. The COMT model seeks to bridge this gap by providing a computationally efficient framework suitable for real-time control with minimal training data.

2. The COMT Framework

The Control-Oriented Muscle Torque (COMT) model, developed by researchers at the University of Waterloo, proposes a physics-informed structure to simplify the relationship between EMG signals and joint torque. The model decomposes the total torque at a joint into a four-part mathematical summation:

  1. Passive Torque: Represents the inherent mechanical resistance and elasticity of the joint tissues.
  2. Torque-Velocity Scaling: Adjusts the torque output based on the angular velocity of the joint.
  3. Torque-Angle Scaling: Accounts for the muscle's force-length relationship relative to the joint position.
  4. Activation Torque: The dynamic component derived from processed EMG signals representing user effort.

By pre-modeling the first three components—which are governed by predictable physical properties—the machine learning component is only required to solve for the "Activation Torque." This "head start" significantly reduces the complexity of the learning task and the volume of data required for training.

3. Actuator-Centric Modeling and Transformation

A key innovation of the COMT model is the shift from anatomical coordinate systems to an actuator-centric frame. Traditional biomechanical models are built around anatomical landmarks (e.g., how a shoulder moves relative to the torso), requiring complex, time-consuming coordinate transformations to be used by a robot in real-time.

The COMT model defines its functions directly in the robot's joint space. To populate the model with data, the researchers utilize two primary methods:

  • Direct Parameter Identification: Using a dynamometer (e.g., Biodex System 4 Pro) to measure specific user output.
  • Jacobian Transformation: Using a Jacobian matrix of velocities to translate existing anatomical Muscle Torque Generator (MTG) data into the exoskeleton’s specific coordinate system, such as the Levitate Airframe shoulder exoskeleton.

This geometric approach allows the system to leverage decades of established biomechanical research without the real-time computational overhead of traditional anatomical mapping.

4. Discussion and Key Insights

The COMT model addresses the "actuator redundancy" problem—where multiple muscles contribute to a single movement—by treating the entire joint as a single torque generator. This simplification is particularly beneficial for real-time robotic control, where speed and reliability are paramount.

Furthermore, the physics-informed nature of the model suggests a high degree of data efficiency. Because the kinematic dependencies (angle and velocity) are handled by the physical model, the ML component does not need to "re-learn" physics from scratch. This supports the hypothesis that the COMT model can achieve high accuracy with significantly smaller training datasets compared to pure ML approaches.

5. Conclusion and Future Work

The COMT model provides a robust theoretical framework for the next generation of EMG-controlled assistive devices. Initial proof-of-concept work has successfully demonstrated the transformation of anatomical torque data into exoskeleton-specific frames, specifically for flexion/extension and adduction/abduction torques.

Future research will focus on the live implementation of the ML component to calculate activation torque in real-time and clinical validation on live exoskeleton hardware to confirm the model's performance and data efficiency in practical, real-world applications.

Summary updated

BibTeX
@inproceedings{nasr2020control,
  author = {Ali Nasr and John McPhee},
  title = {{Control-Oriented Muscle Torque (COMT) Modelfor EMG-Based Control of Assistive Robots}},
  booktitle = {Proceedings of the International Conference of Control, Dynamic Systems, and Robotics},
  year = {2020},
  doi = {10.11159/cdsr20.144},
}
Figure