If you go that route, here are some pointers (and mistakes I made, this is in the context of a Wipeout style AI).
Rotation might cause problems: in 2.56a I used a small torque force to turn the AI, this causes the AI to spin wildly when it hits unusual angles in the track mesh. If you use pure Drot this works really well, but it looks strange as turns of +/- 5 look juddery as they turn. This was using dynamic / rigid body physics.
With sensor ‘whiskers’, certain track shapes like U turns and narrow turns cause problems as the further ahead you look, the AI will turn into the bend too early.If the look ahead is too short, you will need a greater force to turn, which can again cause problems.
In addition, when a whisker is triggered, ensure the others are disabled momentarily (like through states) as otherwise the opposite sensor will eventually brush the track and the AI will get confused.
I did try helper nodes that instruct the AI to do things like brake or acellerate when collided with.
Sensor style AI will never have a good racing line, whilst waypoint AI will, although the latter will be less random during a race. I did try to mix both, but I realised that when the AI switched from ‘seek’ mode (in that the AI follows waypoints) to ‘evasive’ (using rays/radar) the AI would miss nodes. When the AI switched back to ‘seek’, it would have passed several nodes and would track back to them (going the wrong way round the track).
I hope this helps a little, and avoids the frustration I went through.