Abstract
Socially-Assistive Robotics (SAR) has been extensively used for a variety of applications, including educational assistants, exercise coaches and training task instructors. The main goal of such systems is to provide a personalized and tailored session that matches user abilities and needs. While objective measures (e.g., task performance) can be used to adjust task parameters (e.g., task difficulty), towards personalization, it is essential that such systems also monitor task engagement to personalize their training strategies and maximize the effects of the training session. We propose an Interactive Reinforcement Learning (IRL) framework that combines explicit feedback (task performance) with implicit human-generated feedback (task engagement) to achieve efficient personalization. We illustrate the framework with a cognitive training task, describing our data-driven methodology (data collection and analysis, user simulation) towards designing our proposed real-time system. Our data analysis and the reinforcement learning experiments on real user data indicate that the integration of task engagement as human-generated feedback in the RL mechanism can facilitate robot personalization, towards a real-time personalized robot-assisted training system.
Highlights
Socially-Assistive Robotics (SAR) is a research area that studies how robots can be deployed to assist users through social interaction, as users perform a cognitive or physical task [1]
We focus on the personalization procedure of a SAR system, formulating it as a reinforcement learning problem
Based on the current state, the robot selects one of the available system actions, and the system perceives the state, receiving a reward based on task performance and task engagement, as we describe in our experimental procedure
Summary
Socially-Assistive Robotics (SAR) is a research area that studies how robots can be deployed to assist users through social interaction, as users perform a cognitive or physical task [1]. Socially-assistive robotics has been developed to improve user performance through the use of physiological signals [3], considering the Yerkes–Dodson law, which links human arousal and task performance [10] From another perspective, recent works define interactive personalization for socially-assistive robotics as “the process by which an intelligent agent adapts to the needs and preferences of an individual user through eliciting information directly from that user about their state” [11,12]. We discuss how Interactive Reinforcement Learning (IRL) methods can be used to facilitate personalization for different types of users, in a SAR-based cognitive training scenario. We present related work on SAR systems and reinforcement learning approaches for robot personalization, as well as methods for measuring and using task engagement through EEG sensors for adaptive and personalized interactions (Section 2). We present the experimental procedure, including the user simulation and the interactive reinforcement learning experiments (Section 5), and we conclude with a discussion on possible improvements and future steps towards a real-time personalized SAR system (Section 6)
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
