Real-time torque estimation using human and sensor data fusion for exoskeleton assistance

Abstract

Robotic assistive devices have emerged as a potential complement for repeti-tive and user-centered gait rehabilitation. In this field, the development of electromyography (EMG)-based torque controls has played a crucial role in improving the user experience with robotic assistive devices. However, most existing approaches for EMG-based joint torque estimation (i) are designed for upper limbs; (ii) often do not consider the complexity of the walking mo-tion, focusing only on the stance phase; and (iii) rely on complex mathemat-ical models that result in time-consuming estimations. This study aims to address these shortcomings by evaluating the generalization ability of a Deep Learning regressor (Convolutional Neural Network (CNN)) for estimating ankle torque trajectories, in real-time. Several inputs were incorporated, namely, EMG signals from Tibialis Anterior and Gastrocnemius Lateralis, hip kinematic data in the sagittal plane (angle, angular velocity, angular acceler-ation), walking speed (from 1.5 to 2.0 km/h), user's demographic (gender and age) and anthropometric information (height and mass, ranging from 1.50 to 1.90 m and 50.0 to 90.0 kg, respectively, and shank and foot lengths). Re-sults showed that a CNN model with two convolutional layers showed the highest generalization ability (Root Mean Square Error: 23.4±8.36, Normal-ized Mean Square Error: 0.494±0.299, and Spearman Correlation 0.754±0.105). CNN model’s time-effectiveness was tested in an active ankle orthosis, being able to estimate ankle joint torques in less than 2 millisec-onds. This study contributes to a more time-effective model for real-time EMG-based torque estimation, enabling a promising advancement in EMG-based torque control for lower limb robotic assistive devices.