From the paper: "To generate sufficient amount of training data for our model in simulation, we use KanjiVG [5], a database of Japanese Kanji characters. This database consists of both images of the character as well as an ordered list of strokes in the SVG format."