An open-source library for GPU-accelerated robot learning and sim-to-real transfer.
172 karma · joined February 11, 2020
We ended up picking a minimal subset of forearm DoFs that wouldn't impact training speed too much.
From their whitepaper: "Fingerings in the dataset were provided by experienced pianists who graduated from a music college or who had played the piano for more than twenty years. The pianists were asked to choose pieces that they could play and provided the fingering that they had actually used for the performance."
The keys are actually implemented using a spring mechanism, but springs in MuJoCo are currently linear, which isn't the case in the real world.
The demo you are watching is an agent trained from scratch with reinforcement learning. It has roughly 6 days of experience (10M steps at 20 Hz). The Javascript demo is replaying the policy open loop which is why it's not super robust to disturbances.
Re:fingering: we actually use fingering information to create a dense reward for the agent (otherwise it makes exploration super hard). It would be an exciting future direction to have the agent discover and optimize for fingering that best suits its kinematics :) And beyond that, having RL inform pianists about the difficulty of a piece or even more optimal fingering would be amazing.
We trained a bunch of these policies on roughly 150 songs (baroque, romantic, classical) and we did some analysis in the paper if you're interested: https://kzakka.com/robopianist/robopianist.pdf