The Crewe projectAn AI agent learns to cross Crewe town centre
Turn OpenStreetMap into a graph, let a Q-learning agent explore it, and count what that exploration costs.
The learner downloads a small rectangle of central Crewe from the OpenStreetMap API and keeps every road, path and footway. Junctions become points and the stretches between them become links with a length in metres: 940 junctions and 1,230 links. The two landmarks are attached to their nearest junctions, 26.1 m and 32.7 m away. In a straight line those junctions are 352 m apart, but nobody walks through buildings, and a planning algorithm called Dijkstra's, which reads the whole map, finds the shortest walking route: 498.4 m through 14 junctions, after examining 82 junctions.
The Q-learning agent gets no map in advance. At each junction it can take any link, and every step costs it the metres walked. It keeps a table of how good each move from each junction has turned out to be, and updates it after every step. Most of the time it picks the move that looks cheapest so far; one time in five it tries a random move, to explore. Each attempt, called an episode, ends when it reaches the library. The experiment is repeated with 20 different random seeds.
Learning the route by trial and error against planning with a map, our Python run on OpenStreetMap data, 29 September 2026| Method | Result |
|---|
| Dijkstra planning with the full map | 498.4 m route, 82 junctions examined |
| Q-learning: attempts until its route is the shortest | Median 54 (range 46 to 62) |
| Q-learning: distance walked in training to get there | Median 516.3 km (range 477.8 to 573.1) |
| Random wandering with no learning, one arrival | Median 50.8 km |
The agent does learn. After a median of 54 attempts, its preferred route is exactly the 498.4 m one the planner found, and it got there without ever being shown a map. But look at the price. To learn a walk of about half a kilometre it covered a median of 516.3 km of simulated streets, roughly a thousand times the length of the route, most of it in the early attempts when it knew nothing. Even so, that is far better than not learning: a purely random walker needs a median of 50.8 km just to reach the library once, and never improves.
The lesson for anyone building agents is clear. When a reliable map exists, planning with it is enormously cheaper than trial and error. Reinforcement learning earns its place where no such map is available, such as games with too many positions to list, or where mistakes are cheap because they happen in simulation. A robot or an AI agent acting in the real world usually cannot afford 500 km of wrong turns.
Ages 8 to 11
Solve a paper maze by trial, then again with the map, and count the wrong turns.
Ages 11 to 15
Turn a small street map into a graph in Python and find the shortest route.
Ages 15 and up
Code Q-learning, change the exploration rate and measure the cost of learning.
OpenStreetMap data, our agent
Map data is from OpenStreetMap and its contributors, available under the Open Database Licence. The graph, the routes, the agent and every distance in the tables are our own calculations.
