Robots have a reputation for being, well, robotic. For the most part, this reputation has been well earned. Robots typically move with stiff, preprogrammed motions that get the job done, but they are not exactly what one might call graceful. So the idea of a robot taking on some type of creative pursuit (like playing music, for instance) seems very unnatural. If you are not convinced of that, consider the nightmare fuel that the Chuck E. Cheese animatronic band was for children of the 1980s.
But maybe it is time for a change. After all, humanoid robots (like the Unitree R1 we recently covered) are becoming more capable and affordable by the year. Pair that with the modern advances in generative artificial intelligence that give machines the ability to do everything from converse like a human to create beautiful works of art, and we might be on to something. A trio of researchers at the Polytechnic University of Milan recognized where this is all heading, and they have developed a robotic drummer in response.
An overview of the system (📷: A. Shahid et al.)
Called Robot Drummer (where do they come up with these names?), the system combines a humanoid robot’s physical dexterity with reinforcement learning (RL) to produce complex, human-like drumming performances. But drumming presents a particularly tough challenge for robots. It requires split-second timing, rapid multi-limb coordination, and the ability to maintain high-precision movements for several minutes at a time. Unlike typical robotic manipulation tasks, music is a process-driven activity where the goal is to sustain a dynamic sequence over an extended period. Even small timing errors can ruin the rhythm and emotional quality of the performance.
To meet the challenges of robot drumming, the researchers came up with an approach called Rhythmic Contact Chain in which music is represented as a sequence of precisely timed drum strikes. Instead of thinking in terms of continuous audio, the researchers work from MIDI files, which encode the exact timing and type of each percussion hit. By stripping away other instruments and focusing only on the percussion channel, they translate “note-on” events into mapped drum strikes for the robot’s drum kit. Each strike is encoded in a time-indexed one-hot vector, making it straightforward for the robot to know exactly what to hit and when.
Because songs can be long and complex, the researchers break them into fixed-length segments. These are trained in parallel using a unified RL policy, allowing the robot to learn across an entire repertoire more efficiently. The reward system is based on hitting the right drums at the right times, encouraging not just accuracy, but also rhythmic fluency.
Testing was done in simulation on a Unitree G1 humanoid robot, and the music was not limited to simple beats. The system took on genres from jazz to rock to metal, playing well-known tracks like “In the End” by Linkin Park and “Living on a Prayer” by Bon Jovi. The robot achieved rhythmic precision rates above 90%, and began to exhibit emergent human-like drumming behaviors such as cross-arm strikes, adaptive stick assignments, and optimized motion planning.
The team’s next step is to transfer these skills from simulation to real hardware and eventually teach the robot to improvise, adjusting its playing style to live musical cues. In the future, we might not only see humanoid robots playing alongside human musicians, but also contributing their own unique style to the performance. But by the looks of things, that future may still be a long way off.
