Practice Makes Perfect
Lamtharn Hantrakul, Zachary Kondak, Gil Weinberg · 2018
When a pianist effortlessly glides across the keyboard during an improvised solo, the musician is executing a series of movements informed by years of practice ingrained with musical knowledge. This paper proposes an analogous approach that enables Robotic Musicians to learn about its degrees of freedom and physical constraints through "practice" in the form of Deep Reinforcement Learning. We use a Deep Q Network (DQN) to train a virtual agent representing a real 4-armed robotic musician, to motion-plan the optimal sequence of movements given a musical sequence through a learned strategy instead of a search strategy. Early results from our proof-of-concept system demonstrate that DRL can achieve optimal control of a musical agent, learning a form of bi-manual coordination in the process.