Robotic action/observation planning for playing yamakuzushi based on human motion

Abstract

In this paper, we propose a novel robotic action/observation planning method for playing Yamakuzushi game which is a Japanese board game selecting one Shogi piece from the randomly stacked pile and sliding it off the board. In our proposed method, we first select one of the pieces from the pile by referring a human operation characteristics. We use the CNN trained by using the depth image of the pile and the result of human operation. To successfully slide the selected piece off the board without making collision with the neighboring pieces, we formulate the action/observation planning based on the partially observable Markov decision process (POMDP). We select the view pose to obtain enough visibility of the pile around the selected piece. After obtaining the visibility, we determine the direction of piece's motion to avoid the contact with the neighboring objects. The effectiveness of the proposed method is confirmed by experiments.