RoboPianist: Dexterous Piano Playing with Deep Reinforcement Learning (2023)
kzakka.com
kzakka.com
For context I have been working on playing multiple lines of music with right hand lately (Chopin etudes, much struggle...) and saw a video from a professional recommending playing the main melody "with the arm". This has helped a lot with the visualisation and now I can really have a cleaner technique playing the accompaniment above the escapement (using nice flat fingers) and the melody with a much richer tone (using the arm typically for the pinky).
I sort of understood in an intuitive nonverbal sense what I am doing but after your comment, the mechanics now makes a lot more sense of what I'm actually doing - thanks!
There's entire schools of thought around proper ergonomics with respect to piano playing - I took lessons in the Taubman technique which (very simplistically) tries to encourage movement farther upstream on the body.
https://www.ednagolandsky.com/the-taubman-approach-basic-pri...
https://ericye16.com/stanford-cs224r
We were able to make some improvements by tuning how the reward is distributed and also by first pretraining the agent on scales before fine-tuning them on the final pieces.
Thanks to Kevin Zakka for helping us get started with the RL environment!
I've personally experienced how research around this time was being shut down because of AI doomerism. People were getting laid off because of it. It's clear to me that these institutions actively spread AI doomerism so that they have full control over it. They actively called for a stop in AI research so that their personal labs can leap forward ahead. It was a little too on the nose for Big Tech, but they don't understand nuance.
I'd like to also hear how loud the mechanical noise of the machine playing the piano would be. Does the left hand work harder with the heavier keys? What would the hands be mounted to?
Each key (on a good piano) is weighted to have the same force requirements to be played. Of course varying forces can be used to achieve different decibels, but the curve is the same across the piano.
> Every single key on a grand piano keyboard is weighted differently. This is because the strings for each note are slightly thinner and shorter in the treble register, becoming thicker and longer towards the bass register. As a result, there is greater resistance when playing low notes than when playing high notes. In other words, a heavier touch is required in the left hand and a lighter touch in the right hand. To emulate this in a digital piano, the keys are often individually weighted, with the lower keys heavier than the higher ones — something that’s called graded weighting.
https://hub.yamaha.com/pianos/p-digital/a-quick-guide-to-wei...
Yes. The robot used in this study is Dexterous Hand from Shadow Robot. It's a real product, and costs around $200-300k (for a single hand).
Controlling a real robot using a RL policy (model) trained in a simulation environment is also doable. It's called "Sim2Real" and has been widely experimented in the last decade (with mixed success, though).
There's a video that shows the hand moving and it doesn't look like it would be fast enough for anything but relatively slow songs:
I play the piano and I think MIDI does not have enough parameters to describe an acoustic piano. It's not just a single strike velocity that determines each sound. Where you play within the double escapement, how far down you hit the key, they all change the sound a bit.
That said though there are better MIDI synthesizers, e.g. https://www.modartt.com/ They still don't match an acoustic. I can tell the difference quite easily. I think a neural TTS retrained on piano data could do better.
I understand it's a bad argument because 1) people may need assistance to do art/writing and 2) the advancements gained from teaching AI to do art can be applied to other non-artistic endeavors (e.g. piano AI could be really good to operate machines with lots of buttons and no computerized interface).
However, the cost cutting side of the argument is the one that bothers me because companies/people WILL use AI like that in place of actual humans because they're likely to be cheaper in the future. So that pianist or musician playing in a local restaurant can be sure their job will be automated away by a subpar AI and real humans will be relegated to very expensive locations (an extension of replacing humans with recorded music, in a way).
My pessismist side thinks greed will be the downfall of humanity.
I saw a stage performance of Clue a few weeks ago. It was lovely. Nothing was AI generated, and couldn't be, because it was people physically making art live. How nice.
I will continue to pay for those experiences and to prioritize them over recorded ones.
The player piano technology comes from the barrel organ, which seems to have been on the first pipe organs.
I share your pessimism partially, but I also wonder: if the system keeps optimizing for profit, at some point there just won't be enough people in the world to sell products and services to. I don't know what the answer is.
Edit: typo
If for example you find manual art more fun to practice than AI art, you'd be less inclined to give AI companies money to generate art, since you could perhaps spend hours and hours practicing and have fun.
And of course if you don't find manual art enjoyable, but still don't want to give up on the "art" idea entirely, AI art crosses the threshold into enjoyability, so the end product is at least in front of your eyes.
As for how to learn how to make one type of art more fun than the other, if ones preferences are set in stone, it's pretty tough. The AI/non-art-inclined crowd might invert your quote so the laundry and dishes get prioritized, because different people have different priorities.
this is really bad. It might be a breakthrough of what you are doing, but when I listen to the output all of the timing and phrasing is aweful.
The reinforcement learning algorithm takes in the state of the fingers and piano keys, and outputs the required motor actions to perform well.
Also, as you are too lazy to read an article but still feel entitled to comment on it, I herby give you -1 reward. Let's see if that helps to improve your character.
If you know the precise state of each motor in a robot hand, you can compute the absolute position of its fingertip, right?
To make a robot play piano, you need to solve the reverse problem of it i.e. to compute the motor states from the target robot position. This problem is called "Inverse Kinematics" and not fully resolved to this day.
Reinforcement Learning (RL) people claim that, if you let a robot move randomly, and give it apporopriate rewards depending on the resulting state, eventually it learns to solve the problem by itself (just like a human baby eventually learns how to walk through trial & error).
Now, you are looking at the state of the art of RL in 2023.
IK to my knowledge is well known in every setting I am aware of.
No, a state-based RL research like this is essentially an IK problem. Given a goal position in the world frame, you need to find out the motor configuration to move your EEF to that goal position.
> IK to my knowledge is well known in every setting I am aware of.
Really? When I was working in this field, I've actually never seen anyone who used numerical/analytical IK methods on real robots.
Granted that I was not a Robotics engineer (I was a Deep Learning engineer in the team), but my impression at the time was that no practical IK solution was available for a robot with 8 DOF, let alone a 20 DOF robot like Shadow Hand.
If you provide the 6DoF trajectory (+ gripper joints), a lot of robotics (manipulation in particular) is basically solved. The problem is, we don’t have these good trajectories.
Sure, joint space is commonly used for learned policies, but cartesian space isn’t uncommon either.
IK is really just not a major focus on the learning side of robotics because it’s not the problem. The problem is we don’t know what to do even at the slightly higher level.