Problem
An autonomous race car competing at Silverstone has to find its own way around a circuit it has never seen, marked only by cones, while moving.
Approach
The racing line is generated from the midpoints of a Delaunay triangulation over detected cone positions — the triangulation gives a principled way to find the drivable corridor between left and right cone sets without hand-tuned heuristics.
I owned normalisation and persistence of the completed line as a global racing line. After the first exploratory lap the car reuses the stored path rather than re-planning from perception on every subsequent lap, which removes per-lap perception noise from the control loop.
Constraints
Everything ran against real-time perception and odometry data inside a live control loop, so update rate and stability mattered as much as correctness. A planner that produces a better line too slowly is not a better planner.