Motion Generation for Violin-playing Robot Using Reinforcement Learning
Kenzo Horigome, Koji SHIBUYA · The Proceedings of JSME annual Conference on Robotics and Mechatronics (Robomec) · 2023
We are building a system that can determine the bowing motion of our violin-playing robot. In this report, we describe effects of the parameters in the reinforcement learning such as the exploring rate, the number of the hidden layer, and the leaning rate on the bowing motion of the robot. In the violin playing, it is difficult to automatically determine the bowing motion (bow direction, bow speed, etc.) only from the musical score. Thus, we focus on the reinforcement learning as one method to solve the problem. In our system, Q-learning is adopted as a leaning strategy, and neural network a value function. We built a simulation system to evaluate our method. As a result of the simulation, we found that we can improve the success rate of the learning by changing the parameters.