The Development and Utilization of Tandem Fluency in Human-Exoskeleton Interaction

Abstract

There is strong demand for portable technologies that enhance human power output while maintaining safety and range, not only in defense and industry but also in aerospace. Exoskeletons and other wearable powered devices have been proposed as solutions, but a major barrier to adoption is the issue of “fluency”: a combination of metrics representing the seamlessness of human-robot interaction. Most current exoskeleton systems, especially for non-cyclic motions, disrupt user intent and movement, often offering no benefit, or even causing harm by increasing discomfort and injury risk. This lack of fluency is frequently linked to poor intent recognition and absence of predictive control. To address this, we propose developing a human motion prediction system and studying its impact on fluency in exoskeleton-like devices and related human-centered technologies in real-world applications. We introduce an expanded metric “tandem fluency” based on conventional fluency, tailored for evaluating human-robot interaction (HRI) systems where human and robot agents are kinematically synchronized to perform functional tasks. We then develop a proof-of-concept and a functional deep neural network (DNN) capable of detecting human motion intent and predicting motion trajectories in advance using biosignals such as surface electromyography (sEMG). In parallel, we build and test prototype exoskeleton hardware with both single and multiple degrees of freedom. Finally, we conduct human trials with the full closed-loop tandem human-exoskeleton system to evaluate the impact of motion prediction-based control on tandem fluency. The results show that classification and regression prediction of human motion prior to initiation of physical motion is possible and can have performance necessary for practical application of this information, the prediction can be generated not only prior to the physical motion initiation, but often even before the full electrical activation of the primary agonist in many motions, the DNN is robust to variations in sensor hardware and input formatting, and furthermore the use of this prediction in the controls of a tandem robot system has potential to improve tandem fluency by positively affecting both subjective experience and objective/metabolic results.