Recursive process of motion recognition and generation for action-based interaction

Abstract

To integrate humanoid robots into daily life, it is necessary to develop the robots that can understand human behaviors and generate human-like behaviors adaptive to their human partners. In previous robotics research, frameworks of encoding human behaviors into motion primitive models, recognizing observed human actions as the motion primitives, and synthesizing motions from the motion primitive models have been developed. The bidirectional process of motion recognition and motion synthesis can be helpful to establish intuitive behavioral interaction between humans and robots. This paper proposes a novel approach to realizing a robot intelligence for the human-robot interaction. This approach symbolizes perception and actions using hidden Markov models. The recursive process of motion recognition and synthesis based on the hidden Markov models enables a robot to understand that the human partner recognizes robots actions and to perform actions accordingly. The proposed framework was applied to an example of human-robot interactions where robot walks through a group of several partners. This application was tested on captured human behaviors, and the validity of this framework is demonstrated through the experiments.