I have been working almost exclusively on AI – look-ahead and collision avoidance routines; trying to weight the importance of things like wheels being on the ground, angular momentum, versus avoiding an impact. Spent some time recording racing ‘trails’ (approx 2000 nodes laid down when a good lap is recorded, approx 100 laid down following a crash). This turns out to need big data requirements (perhaps 70MB per track per type of vehicle), and more importantly very CPU intensive too (testing many potential steering/accelarion scenarios to see if any landed on a good trail node) — although it did work reasonably well, I have shelved this approach in order to find something more efficient. The simpler waypoint+collision avoidance works quite well; each waypoint records each time a car gets through/crashes at a certain speed, so safe/lowrisk/maximum speeds are known for each waypoint.
[This is an historical blog post which used to live on my Darkwind website]