You could actually add mazes and paths through them to the training corpus, or make a model for just solving mazes. I wonder how effective it would be, I’m sure someone has tried it. I doubt it would generalize enough to give the AI new visual reasoning capabilities beyond just solving mazes.