ZRob: Robot drummer listens to what he plays and learns as he goes
The robot drummer ZRob wants to be able to play the drums better than other musical robots. Using a flexible "wrist", he finds his own style.
Several ZRobs are installed on a drum kit.
(Image: Mojtaba Karbasi)
A team of scientists from the Center for Interdisciplinary Studies in Rhythm, Time and Motion at the University of Oslo has developed a drum robot called ZRob, which has a flexible "wrist" that can guide the drumsticks more loosely, similar to a human wrist. The robot drummer can listen to itself while playing the drums and uses reinforcement learning ( – RL) to improve its playing.
The drum robot is not particularly large. It essentially consists of a 3D-printed housing that accommodates a 48-volt DC motor and controls an arm with two degrees of freedom. A flexible holder for a drumstick is attached to it, suspended on two springs, as can be seen from the study "Embodied intelligence for drumming; a reinforcement learning approach to drumming robots", which was published in Frontiers in Robotics and AI. The suspension is intended to give the robotic arm the necessary flexibility that a human wrist has and is required for drumming.
The researchers argue that people often use their own bodies via movements to give the playing of an instrument a special expression. The scientists have something similar in mind with ZRob. By giving the robot a flexible suspension for the drumstick, it becomes more adaptable. Replacing the springs with ones of different hardness can influence the drumming and the sound. The robot body thus contributes to the playing behavior.
Playing optimization through reinforcement learning
ZRob is controlled at the lowest level by an Arduino microcomputer. It is fed with MIDI data that fundamentally determines the drumming. In order to optimize the robot 's playing, the researchers rely on reinforcement learning. The acoustic information of the drumming is recorded via a microphone, and the artificial intelligence (AI) is trained via RL to improve the drumming.
Videos by heise
The RL algorithm is based on the Deep Deterministic Policy Gradient (DDPG) method. The robot is trained with a combination of intrinsic and extrinsic rewards instead of relying solely on extrinsic rewards. The intrinsic rewards allow new rhythmic patterns to emerge, while the extrinsic rewards lead to predictable patterns that are identical to the MIDI data.
However, the scientists do not believe that this type of robotic music can touch people in the same way as music made by humans. However, ZRob can create new types of rhythms and play drums much faster than a human. So far, several ZRobs are needed to play a drum set.
(olb)
