The Clacton-on-Sea projectA potential field agent on a grid of 6,454 Clacton buildings
Pull towards the goal, push away from walls, step downhill, and see how far that gets.
A single Overpass request fetches the outline of every building OpenStreetMap holds in a box over central Clacton-on-Sea: 6,454 of them. The learner draws the outlines onto a grid of 4-metre cells, 550 across and 497 down, and finds that 20.9% of the cells are covered by a building. In this toy world everything else counts as open ground, so the agent may cross roads, gardens and even water. Each cell gets a height: its straight-line distance to the goal, plus a steep extra hill within 12 metres of any building. The agent looks at its eight neighbouring cells, steps to the lowest, and repeats. If no neighbour is lower than where it stands, it has reached either the goal or a local minimum, and it stops.
300 random trips across the Clacton grid, potential field agent, our Python run on OpenStreetMap data| Straight-line distance | Trips | Reached the goal | Share |
|---|
| 50 m to 250 m | 10 | 4 | 40.0% |
| 250 m to 500 m | 33 | 11 | 33.3% |
| 500 m to 1 km | 97 | 28 | 28.9% |
| Over 1 km | 160 | 17 | 10.6% |
| All trips | 300 | 60 | 20.0% |
The agent arrived on 60 of 300 trips. The rest ended in a local minimum, typically a building face or a courtyard-like pocket that lies directly between the agent and its goal: the pull says forward, the wall says no, and stepping sideways looks like climbing. Stuck agents had covered a median of only 143 metres, on trips whose median straight-line length was 1,135 metres. Longer trips failed more often simply because they pass more buildings. When the agent did get through, its path was about as short as the shortest possible route on the grid, so the method is efficient when it works.
Changing the repulsion, same 300 trips, our calculation| Agent setting | Goals reached of 300 | Share |
|---|
| Attraction only, no repulsion | 68 | 22.7% |
| Repulsion within 12 m of buildings | 60 | 20.0% |
| Repulsion within 24 m of buildings | 58 | 19.3% |
Tuning did not rescue it. Removing the push or doubling its reach moved the success rate by only a few points, because the trap comes from the shape of the landscape, not from one badly chosen number. A planner rescues it. Dijkstra's shortest-path search on the same grid reaches the goal every time, since all the starts and goals lie in one connected open region of 215,161 cells. The planner pays for that certainty by examining a large part of the grid before it moves, while the reactive agent looks at eight cells per step. Real robots often combine the two: a planner sets waypoints and a reactive layer handles what is immediately in front.
Ages 8 to 11
Roll a marble across a tilted tray of obstacles and mark the pockets where it comes to rest.
Ages 11 to 15
Code a downhill-stepping agent in Python on a small grid and draw the spots where it stalls.
Ages 15 and up
Build the Clacton grid, run 300 trips, vary the repulsion and compare with a shortest-path planner.
Whose data, whose agent
Building outlines are © OpenStreetMap contributors and are used under the Open Database Licence. The grid, the agent and all counts are our own. The grid ignores fences, water and private land, so the results say something about the algorithm and nothing about walking around Clacton.
