Designing Deep Reinforcement Learning Systems for Musculoskeletal Modeling and Locomotion Analysis Using Wearable Sensor Feedback

Abstract

In this paper, a system for lower-limb amputees to collect and analyze their locomotion activities is developed. Wearable Gait Lab (WGL) system is a pair of smart insoles that can collect plantar pressure data and foot motion data of subjects. A reinforcement learning model is trained to imitate the walking pattern of the lower-limb amputee on a musculoskeletal model by introducing realistic velocity data into the training process. The plantar pressure data collected from our Wearable Gait Lab were used to recognize the locomotion modes of the amputees. Experiments showed that the outcome of the musculoskeletal model can be a reference for analyzing muscle activities of the amputee. The system also shows the promising and stable performance of recognizing locomotion modes in amputees' daily life. The accuracy of locomotion mode recognition reaches 98.02%. Monitoring muscle activities and locomotion modes of lower-limb amputees can help them prevent secondary impairments in rehabilitation.