Motion Generation Using Reinforcement Learning for Violinist Robot

-Effect of Reward Value on Learning and Increasing the Number of Bar-

Abstract

This study aims at building a system capable of automatically determining the bowing and fingering parameters for an anthropomorphic violinist robot. We adopt the reinforcement leaning for determining the bow speed and the bowing-direction under the constant bow-force condition, in which a neural network is embedded as a value function. In the previous report, the learning success rate was small, which might be caused by the reward values. Thus, the reword values were changed in this report. In addition, we applied the simulation method to a short musical piece of four bars. As a result, we found that more stable leaning results were obtained by using the new rewards values, and our proposed method can be applied to the four-bar musical piece.