Embodied intelligence for drumming; a reinforcement learning approach to drumming robots

This paper investigates the potential of the intrinsically motivated reinforcement learning (IMRL) approach for robotic drumming. For this purpose, we implemented an IMRL-based algorithm for a drumming robot called ZRob, an underactuated two-DoF robotic arm with flexible grippers. Two ZRob robots were instructed to play rhythmic patterns derived from MIDI files. The RL algorithm is based on the deep deterministic policy gradient (DDPG) method, but instead of relying solely on extrinsic rewards, the robots are trained using a combination of both extrinsic and intrinsic reward signals. The results of the training experiments show that the utilization of intrinsic reward can lead to meaningful novel rhythmic patterns, while using only extrinsic reward would lead to predictable patterns identical to the MIDI inputs. Additionally, the observed drumming patterns are influenced not only by the learning algorithm but also by the robots' physical dynamics and the drum's constraints. This work suggests new insights into the potential of embodied intelligence for musical performance.