Researchers at Google have posed themselves an interesting problem to solve: mastering the piano. However, they’re not trying to teach themselves, but a pair of simulated anthropomorphic robotic hands instead. Enter RoboPianist.
The hope is that the RoboPianist platform can help benchmark “high-dimensional control, targeted at testing high spatial and temporal precision, coordination, and planning, all with an underactuated system frequently making-and-breaking contacts.”
If that all sounds like a bit much to follow, the basic gist is that playing the piano takes a ton of coordination and control. Doing it in a musical way requires both high speed and perfect timing, further upping the challenge. The team hopes that by developing control strategies that can master the piano, they will more broadly learn about techniques useful for two-handed, multi-fingered control. To that end, RoboPianist models a pair of robot hands with 22 actuators each, or 44 in total. Much like human hands, the robot hands are underactuated by design, meaning they have less actuators than their total degrees of freedom.
If you’ve ever seen a human-like robot hand pick something up, you’ve probably noticed how slow and ungainly they are. With so many joints to control, making a hand work is a tough task. Learning to play the piano is thus an excellent benchmarking test to develop these techniques. The project paper highlights a variety of methods the researchers have used to train RoboPianist to play the instrument. It’s not great yet; you’d certainly make excuses if RoboPianist asked you to come to a gig. But it can definitely play the piano to a basic degree, that’s for sure.
The work is available on Github for those that wish to dive deeper, while there’s also a live demo you can play with in a browser. In the meantime, you might like to explore other roboticized efforts to play the piano. Video after the break.