Application of Reinforcement Learning to a Robotic Drinking Assistant

摘要

Meal assistant robots form a very important part of the assistive robotics sector since self-feeding is a priority activity of daily living (ADL) for people suffering from physical disabilities like tetraplegia. A quick survey of the current trends in this domain reveals that, while tremendous progress has been made in the development of assistive robots for the feeding of solid foods, the task of feeding liquids from a cup remains largely underdeveloped. Therefore, this paper describes an assistive robot that focuses specifically on the feeding of liquids from a cup using tactile feedback through force sensors with direct human–robot interaction (HRI). The main focus of this paper is the application of reinforcement learning (RL) to learn what the best robotic actions are, based on the force applied by the user. A model of the application environment is developed based on the Markov decision process and a software training procedure is designed for quick development and testing. Five of the commonly used RL algorithms are investigated, with the intention of finding the best fit for training, and the system is tested in an experimental study. The preliminary results show a high degree of acceptance by the participants. Feedback from the users indicates that the assistive robot functions intuitively and effectively.