Abstract
Recently, autonomous robots working in human symbiotic environment have been studied and developed. A robot is required to adapt to the environment by itself. However, it is difficult to design in advance all robot behavior needed in the daily life environment. Therefore, imitation learning from human behavior observation has been studied so far. Conventional studies of imitation learning focus on recognition of demonstrated behavior or learning performance of the imitated behavior. It is not adequate enough to imitate the demonstrated motion by following joint angle trajectory of demonstrated motion precisely because link structure and dynamics of the robot is different from the one of the human demonstrator. Therefore, the demonstrated motion trajectory is often used as an initial one for the robot learning of the demonstrated motion. However, they use the demonstrated motion only for the initial parameter of learning but not for learning bias or evaluation during trajectory learning with via-point-representation. This paper proposes a trajectory-based imitation learning method that introduces observed human-demonstrated motion data as a bias during the learning. It shows case studies of humanoid walking motion learning with via-point representation and analyses the effects of the observed data to the learning and the validity of the proposed method.
