Support Vector Machine classification of muscle cocontraction to improve physical human-robot interaction

Abstract

Current control strategies in physical human-robot interaction (pHRI) face some limitations: endpoint stiffness level is not directly measurable in typical haptic control situations and the bilateral nature of the interaction requires proper treatment of operator dynamics. These limitations have reduced current control approaches to estimation techniques based on correlated metrics, such as electromyographic (EMG) signals. The current study investigates a parameter known to represent the nullspace of the mapping from muscle forces to joint torques in the human musculoskeletal system, i.e., muscle co-activation. It is hypothesized that this parameter is a random variable that correlates with muscle activities, measured through EMG signals. A change in this variable directly affects endpoint stiffness, and therefore system performance and stability. This study additionally presents a methodology for processing EMG (cocontraction) data prior to training a Support Vector Machine (SVM) classifier, meant to be used as a decision tool for haptic impedance control. Results presented serve as a springboard for the incorporation of a human operator model suitable for real-time implementation and sufficient to support bilateral human-robot interaction. The long-term goals of this research are to understand the mechanisms governing neuromotor adaptation in pHRI, and ultimately design a novel haptic device that tunes its impedance gains in conjunction with changes in an operator's physical and cognitive state. Such implementation would improve the efficacy of robot co-workers operated in industrial settings, among other applications.