Assistive HRI interface with perceptual feedback control: An approach to customizing assistance based on user dexterity.

摘要

In this dissertation, we propose a novel method for customizing an assistive interface on the basis of user dexterity for human-robot interaction (HRI). Customizing the assistive HRI interface, in practice, is a challenging problem due to the variety of user's task-performance and task-performing characteristics. For this reason, we develop a method to assist a user with strategies of a high-performer who has the most similar task-performing characteristics to the user. From the experiment evaluating user's task performance, we observed that our method indeed enhanced the user's performance in terms of task-completion time and the average velocity during the task.;The backbone of our approach is based on setting virtual fixture parameters to yield the assistive control and perceptual (haptic and visual) feedbacks which are customized for a specific user. First, we model a user as a cost function using the techniques from inverse optimal control (IOC). With the underlying assumption -- human users are optimizing a cost function while performing a given task -- we infer the unknown parameters of the cost function from observing the user demonstration. Next, we define three features that characterize the user's task-performing characteristics as the balances of the inferred parameter vectors, and classify the user based on the closest high-performer in feature space. Finally, we set the virtual fixture parameters according to the user's task-performance, class, and features to provide the user the customized guidance with the high-performer's strategies.;We carry out human subject experiments to evaluate the user performance in the presence of various assistance modes. The results from the experiments show that the customized virtual fixturing with haptic feedback outperforms the other types of "virtual fixturing and feedback" combinations, and depicts the enhancement of both high- and low-performing users' performance as well. Hence, we conclude that our approach is an effective way to improve the user task performance.