← driving demos

You are the planner

Same scene, same car. You choose what it values.

A boar steps into the road 38 m ahead. You are doing 58 km/h. A truck is 18 m behind you; a car is coming the other way. The planner samples 99 possible manoeuvres, scores every one, and drives the cheapest. (Thin lines = rejected candidates, coloured by cost. Orange = chosen.)

The cost function

How dearly it holds a wide berth and avoids any contact.
Penalty on hard braking and violent swerving.
Penalty on losing time.
How sure the detector is that the blob is an animal.
—
Danger to the boar—
Danger to the oncoming car—
Danger from the truck behind—
Danger of leaving the road—
Time lost—
Peak deceleration—
“CONTACT” means this plan actually hits something.
What the winning plan's cost is made of
risk
0%
discomfort
0%
lost time
0%
There is no neutral setting. Every slider position is a value judgement about whose safety, whose time, whose comfort. A car ships with some setting — chosen by someone, usually not in public.
Continuous weights, discrete outcomes. Nudge one slider a few percent and the plan flips from braking to swerving. A small engineering tweak becomes a categorically different act on the road.
And this is the transparent case. You can see these sliders because this is a modular planner. An end-to-end network has no sliders — its values are implicit in its training data. So where would you audit them?