Online ensemble deep random vector functional link for the assistive robots

Abstract

Active upper limb assistive robots have the potential to improve the quality of life for patients with limb disabilities and assist those who require rehabilitation. However, patients often have difficulty accepting these robots due to the lack of intuitive human-robot interaction. One of the key challenges is accurately predicting human motion intention throughout the movement trajectory. To address this issue, we propose a dynamic online ensemble deep random vector functional link (DOedRVFL) network that relies solely on data from wear-able inertial measurement units (IMU) for online joint angle prediction. The DOedRVFL employs multiple hidden layers to extract rich features from the IMU data. The random nature of these layers enables real-time applications. Additionally, we use recursive least squares to optimize each output layer's weights in real-time. Finally, we designed a dynamic ensemble module to aggregate all outputs while considering real-time performance. Comparative results demonstrate the superiority and suitability of DOedRVFL for predicting human joint angles. Furthermore, online learning and randomized feature extraction make it well-suited for real-time control of assistive robots.