Google researchers have developed MapTrace, a synthetic data generation system that trains AI models to trace routes on maps with human-like precision. The system addresses a critical gap in multimodal large language models, which excel at recognizing objects but struggle with spatial reasoning and navigation tasks.

Traditional approaches to teaching AI navigation would require millions of manually annotated maps - an impractical task. Instead, MapTrace uses Gemini 2.5 Pro and Imagen-4 models in a four-stage pipeline that generates diverse maps, identifies traversable paths, builds navigable graphs, and validates routes. The system produces 2 million question-answer pairs of annotated map images with valid paths.

When fine-tuned on just 23,000 paths from the synthetic dataset, models like Gemma 3 27B and Gemini 2.5 Flash showed significant improvement on real-world navigation benchmarks. The researchers measured success using normalized dynamic time warping (NDTW), which compares predicted paths to ground truth routes. While generated images occasionally have typographic errors, the team focused on path fidelity, anticipating that future improvements in image generation will resolve these artifacts naturally.