HaberDashers

SplineCoords includes SplineID, SplineIndex, InputKey and Offset. SplineID refers to the unique ID each spline would be given on game starts (equals to their index in splines array), SplineIndex refers to a single blueprint spline's index (position) for a track, InputKey refers to the "position" on a spline that Unreal uses internally, and Offset refers to the displacement from a point on the spline to the exact point in the Cartesian coordinates. Splines could be given the same SplineIndex, so that short cuts could work properly for the game.

So from a point in Cartesian coordinates, the most direct way is to for loop all the splines, find the nearest, and fill out the SplineCoords info to return. But to improve performance, we did a specific progress tracking for AI: AI remembering its progress and spline and being offered spline choices when approaching a shortcut.

From SplineCoords, Cartesian world coordinates could be directly calculated from Get Location at Distance Along Spline. Also, the track width is utilized to clamp Offset, so that the returned position is always on track.

Tracks are not always flat in XY space and of the same width in HaberDashers and track layouts are constantly changing for a long time. The forward direction could be directly obtained from Get Direction at Distance Along Spline. So it is necessary to retrieve the right direction at the start of a game, while track width could be referenced in level design documents.

Therefore, I used data table in UE4 to record track width data, and ray cast at the start of a game to obtain direction information of the track. For every 1/10 segment between two spline points, ray cast once towards the ground, cross product to get the right direction. Translate ray cast hit actor's name into data table name to retrieve the track width.

We were encountering bugs related to the wrong execution order of components, like AI drives to the wrong place because of outdated advice. So we started to use Add Tick Prerequisite Component to force the dependency relation to be right.

But we soon got warnings from Unreal Profilers that AI computation became a concern, because every AI did all processing every game tick. However, simply change tick interval time was not satisfying: all AI karts were less responsive and sharp turns couldn't be handled anymore.

Finally, I mitigated the performance issue while guaranteeing AI responsiveness by updating only one AI's decision on the destination in one frame in turn (also choose spline), while the Steering component ticks every frame with AI controller, and use items/coins advisors ticks every several seconds.