Learning efficient control of robots using myoelectric interfaces

Abstract

Myoelectric controlled interfaces are a vital component for advancing applications in prostheses, exoskeletons, and robot teleoperation. Current methods search for optimal neural decoders for enhanced initial user performance. However, recent studies demonstrate learning an inverse model of abstract decoders to improve performance over time. This paper proposes a paradigm shift on myoelectric interfaces by embedding the human as controller of a system and allowing the human to learn how to control it via control tasks with similar mapping functions. The method is tested using two different control tasks and four different abstract mappings of upper limb myoelectric signals to control actions for those tasks. The results confirm that all subjects are able to learn the mappings and improve performance efficiency over time. A cross-trial evaluation reveals a significant learning transfer when a new control task is presented using the same mapping as a previous task, resulting in enhanced initial performance with the new task. Comparison of EMG signal evolution across subjects indicates a significant population-wide muscle synergy development that results from learning and implementing the inverse model of the mapping function to complete the tasks. This suggests that efficient performance may be achieved by learning a constant, arbitrary mapping function applied to multiple control tasks rather than dynamic subject- or task-specific functions. Moreover, this method can be used for the neural control of any device or robot, without restricting them to anthropomorphic or human-related counterparts.