PianoX Dataset

该数据集包含了来自一个YouTube频道的60首钢琴曲的笔记、相应的演示视频以及指尖轨迹。每首歌曲被划分为时长为30秒的片段，其中包含了12个未见过的片段，用于研究通用策略的泛化能力。该数据集的规模为431个片段，包含258,000个状态-动作对，其任务是通过互联网演示来训练一个钢琴演奏代理。

This dataset contains the note data, accompanying demonstration videos, and fingertip trajectories of 60 piano pieces sourced from a YouTube channel. Each piece is divided into 30-second segments, and 12 of these segments are reserved as unseen samples for investigating the generalization capability of general policies. The dataset comprises 431 segments in total, containing 258,000 state-action pairs. Its targeted task is to train a piano-playing AI Agent via internet-based demonstrations.