Active robot learning for building up high-order beliefs

Abstract

High-order beliefs of service robots regard the robots' thought about their users' intention and preference. The existing approaches to the development of such beliefs through machine learning rely on particular social cues or specifically defined award functions. Their applications can, therefore, be limited. This paper presents an active robot learning approach to facilitate the robots to develop the beliefs by actively collecting/discovering evidence they need. The emphasis is on active learning. Hence social cues and award functions are not necessary. Simulations show that the presented approach successfully enabled a robot to discover evidences it needs.