MapTrace is a large-scale synthetic map path tracing dataset released by Google in collaboration with the University of Pennsylvania in 2025. The related research findings were published in a paper. MapTrace: Scalable Data Generation for Route Tracing on Maps This dataset aims to improve the fine-grained spatial reasoning and path planning capabilities of multimodal large language models (MLLMs) in map scenarios. The core objective is to train models to generate pixel-accurate, continuous, and walkable paths from the origin to the destination.

This dataset contains approximately 2 million path samples, covering maps of varying complexity and style. It is mainly divided into two subsets: complex maps (such as stylized maps of shopping malls and park guides) and simple maps (such as office building, apartment, or campus floor plans).

Data field descriptions:

• image_bytes / image: Raw byte data of the map image (RGB image), with start and end points marked in some versions;
• label_text / label: The string representation of the coordinate sequence of the target path, with the coordinates uniformly normalized to the range [0, 1];
• input_text / input: A path tracing problem in natural language form, requiring the model to find the corresponding path based on a map;
• map_description: A natural language description of the map, used to support text-to-image models in generating synthetic maps;
• (In some versions) Traversability mask: indicates whether each pixel is passable in binary form.