Embedded human control of robots using myoelectric interfaces

Abstract

Myoelectric controlled interfaces have become a research interest for use in advanced prostheses, exoskeletons, and robot teleoperation. Current research focuses on improving a user's initial performance, either by training a decoding function for a specific user or implementing "intuitive" mapping functions as decoders. However, both approaches are limiting, with the former being subject specific, and the latter task specific. This paper proposes a paradigm shift on myoelectric interfaces by embedding the human as controller of the system to be operated. Using abstract mapping functions between myoelectric activity and control actions for a task, this study shows that human subjects are able to control an artificial system with increasing efficiency by just learning how to control it. The method efficacy is tested by using two different control tasks and four different abstract mappings relating upper limb muscle activity to control actions for those tasks. The results show that all subjects were able to learn the mappings and improve their performance over time. More interestingly, a chronological evaluation across trials reveals that the learning curves transfer across subsequent trials having the same mapping, independent of the tasks to be executed. This implies that new muscle synergies are developed and refined relative to the mapping used by the control task, suggesting that maximal performance may be achieved by learning a constant, arbitrary mapping function rather than dynamic subject- or task-specific functions. Moreover, the results indicate that the method may extend to the neural control of any device or robot, without limitations for anthropomorphism or human-related counterparts.