Maze Runner AI
Generate a maze, then race it yourself, hand it to a search algorithm, script your own runner, or train a small neural network to escape it.
A playground for pathfinding, human or otherwise. Every maze is generated fresh and guaranteed solvable, from a compact 15×15 up to a colossal 51×51, with optional twists: keys and gates, black holes that end the run, multi-trip exits and fog of war. Then pick a runner — yourself on the arrow keys, a classic algorithm (A*, breadth-first, depth-first, Dijkstra), your own program in the built-in MazeScript language, or a neural network you train on the spot. The network is deliberately small and inspectable: eight sensors into a 12-neuron hidden layer and four actions, with state memory for cells it has seen, tunable reward shaping and a live view of its activations while it runs. The page is honest about what that means — a high training accuracy says the route was learned, not that the thing generalises to unseen mazes.