Acquisition of a page turning skill for a multifingered hand using reinforcement learning

This paper proposes a method of acquisition of a page turning skill for a multifingered robotic hand using reinforcement learning. The goal of this paper is generation of the manipulation skill of a flexible object without its explicit model and tactile sensation during its control. In this paper, a page turning task is considered as an example of such tasks, where a sheet of paper, generally used in books, is a flexible object. The fingertip trajectories are obtained by reinforcement learning based on simulation. The reward considering a friction condition is given, so that a page turning skill without slip between the finger and the paper is obtained. The validity is confirmed by an experimental system which consists of a 2-d.o.f. manipulator and two 2-d.o.f. fingers.