The gauging process in the domains of control and reinforcement learning advance is quite challenging. One area that hasn’t received much attention is the development of robust benchmarks for high-dimensional control tasks, particularly in bi-manual (two-handed) multi-fingered control. Despite years of research, high-dimensional control in robots remains a major difficulty.
A team of researchers from UC Berkeley, Google, DeepMind, Stanford University, and Simon Fraser University has introduced a new benchmark suite called ROBOPIANIST for high-dimensional control. This benchmark focuses on the task of playing songs on bi-manual simulated anthropomorphic robot hands, using sheet music in a Musical Instrument Digital Interface (MIDI) transcription. The robot hands used in the benchmark have 44 actuators in total, with 22 actuators per hand, similar to human hands.
Playing a song well requires precise spatial and temporal coordination, as well as strategic planning of key pushes to make other key presses easier. The ROBOPIANIST-repertoire-150 benchmark consists of 150 songs, each serving as a standalone virtual work. The researchers conducted comprehensive experiments using model-free (RL) and model-based (MPC) methods to study the performance of different control algorithms. The results indicate that the proposed policies have the potential for strong performances, although there is still room for improvement.
One notable aspect of ROBOPIANIST is its ability to assess the difficulty of learning a song, which can be used to categorize songs based on their complexity. This feature can facilitate further research in various areas of robot learning, such as curriculum and transfer learning. The benchmark also provides opportunities for studying different approaches like imitation learning, multi-task learning, zero-shot generalization, and multimodal (sound, vision, and touch) learning. ROBOPIANIST offers a simple goal, a reproducible environment, clear evaluation criteria, and the potential for future extensions.
For more details, you can check out the research paper, project website, and Github repository. Credit for this research goes to the researchers involved in the project. Don’t forget to join our ML SubReddit with over 26k members, Discord Channel, and Email Newsletter to stay updated on the latest AI research news and cool projects.