Abstract
Piano playing is not only a compelling art form that has fascinated mankind for over three centuries, but its mastery requires hard work and challenges the limits of human dexterity. This makes it an intriguing domain to test the current limits of Deep Reinforcement Learning, which encounters significant challenges when tackling complex tasks in high-dimensional action spaces. At the outset of this work, the RoboPianist framework, a simulated system of floating piano-playing anthropomorphic hands, was restricted to a fixed set of 150 pieces of music. We enabled it to train on arbitrary songs from MIDI scores to develop a method for estimating the difficulty of any piano piece based on the agent's learning process. Furthermore, in order to enhance the realism and to incorporate the effects and limitations of having a humanoid body, we introduce the "Full-body RoboPianist", a whole-body extension of the original framework.
| Original language | English |
|---|---|
| Supervisors/Reviewers |
|
| Publication status | Published - 2025 |
Fields of science
- 102001 Artificial intelligence
- 102003 Image processing
- 202002 Audiovisual media
- 102015 Information systems
- 102 Computer Sciences
JKU Focus areas
- Digital Transformation
Cite this
- APA
- Author
- BIBTEX
- Harvard
- Standard
- RIS
- Vancouver