Draw your racing line on today's road. Pick a speed and how far ahead the steering looks. The car has a lag it does not tell you about, and gates that the centre line misses.
Cost is in seconds: lap time, plus three for every second off the road, plus two per missed gate. The steering is a feedback controller on your line plus a feedforward term from the line's curvature a distance L ahead. With L = 0 the feedforward arrives late by exactly the lag; the faint ghost drives your line that way. Lateral grip is finite, so a tight line at high speed slides wide. The second rung is the centre line with the two knobs at their best for today, so beating it takes a line of your own. Your hand-drawn line is the controller here.
Tyres have a limit. The lateral acceleration v²κ must stay under it or the car slides, and here it slides wide and loses time off the road. The centre line has the road's curvature; a good line uses the width of the road to make the curvature smaller, cutting each apex and running wide on the exit, which is why racing drivers do not drive down the middle and why your hand can beat the geometry. Then look ahead: the faint ghost drives your line with L = 0, and every corner it enters a beat late is the hidden lag made visible. A smooth line is a feasible line, and the right look-ahead is about speed times lag.
A controller that reacts to error can act only once the error exists; that is the price of feedback, worth paying because feedback needs to know little about the world. Put a lag between command and effect and the reaction lands later still. The lag has unit gain, so a magnitude plot never shows it; only the phase does, and a loop whose crossover sits near 1/τ oscillates however you tune it.
Numbers from this series: the steering lag here is 100–250 ms, and at speed a beat is a metre of gravel; in Stable is Still, 10–40 ms of actuator delay is what makes Riccati beatable; in the lab the balancing LQR falls at about 40 ms and no stiffness saves it. Two things beat a delay: knowing the future (preview, this game) or predicting the present from the past (a Smith predictor; Time Machine).
Two degrees of freedom: feedback for what you do not know, feedforward for what you do. A curve of curvature κ at speed v needs lateral acceleration v²κ, so command it before the error appears. Preview control reads the reference a little into the future, so a command that takes τ to act is issued τ early; in distance that is speed times lag, which is why the slider's sweet spot moves with the speed.
Feedforward is only as good as the model it is computed from: with the wrong lag the preview arrives early or late, and with a wrong plant the inverse is wrong. When the plant cannot be inverted at all, because a zero sits in the wrong half-plane, the remedy is to cancel its phase instead of its gain, which is Plotter. Walking robots preview footsteps, machine tools preview toolpaths, cars preview the map.