Implementation of an adaptive surface electromyography-based exoskeleton controller with k-nearest-neighbors classification

Controlling powered human exoskeletons presents challenges compared to other physical human-robot interaction (pHRI) contexts, such as constant human contact and constrained human motion. The fundamental interaction between the human and exoskeleton limits the ability to demonstrate proper robot behavior without wearing the system, a common method for learning algorithms. Exoskeletons commonly have a static set of reactions, tuned to specific human movements or manually controlled interactions. Anticipatory classification of movement intent via surface electromyography (sEMG) may alleviate fatigue caused by the constant force production often needed for the human to signal motion intent, and improve human-exoskeleton coordination by minimizing lag due to feedback loops. Previous implementations of sEMG-controlled devices required expert placement of sensors, and did not account for physical contact nor user fatigue, both of which alter sEMG profiles. We present a machine-learning-based exoskeleton system using anticipatory sEMG inputs that is robust to non-specific sensor placement and sEMG changes due to fatigue, and adapts to users over time.