Robot Learning to Play Drums with an Open-Ended Internal Model

The issue of how a robot playing instruments is one of the core challenges in the area of robotic musician and has been studied for many years. Most previous research mainly relied either on the close-loop adjusting manually or on the predesigned characteristic function with the expertise, while ignoring to learn the mechanism of human in some ways. Enlightened by recent works (humanoid walking, robot fall protection, object grasping, etc.) in robotics community, where inspirations from human being offer a promising alternative in robots acquiring their skills, the issue of how a robot learning the human-inspired drum-playing skill is addressed in this research. With the investigation on the learning process of human, a new approach for robots autonomously developing their drum-playing skill is proposed. Firstly, the robot developmentally learns its internal model through repeated sensorimotor experiences between its motor system and the sounds. Then it constantly adjusts its action according to auditory feedback under an open-ended learning style. Experimental results show that the proposed approach is effective for a robot to achieve the drum-playing skill autonomously. The robot can predict the likely sound of a planned or executed motor action with an average similarity of 81.19%, and the motor action required to achieve a desired sound with an average similarity of 81.28%. Furthermore, with our open-ended algorithm, the robot can achieve a desired sound with an average similarity of 96.93% via average 1.16 times adjustment. After adjusting, the internal model will be updated according to the sensorimotor experiences produced in previous adjustment process. Utilizing the new model, the robot can predict the motor action required to achieve a desired sound with an average similarity of 97.46%.