Watch neural-net racing cars learn to drive

Originally published at:


Lol, in early trials, always the one car that goes the opposite way. That’d be me.


I’d like to see one those Gen10 or later cars put on a different course. Is it learning to drive, or learning the course?


Sort’a looked like food making its way through an intestine.


There’s no memory as such, but it does become optimized for the curves it sees during its run … make it see left hand curves instead and it’ll probably fail spectacularly. By the same token, it drives pretty well, so it doesn’t usually see itself facing a wall perpendicularly - if it ever does, it’ll probably get in trouble, even on the course it was trained with. They should be varying the initial conditions by starting it at different locations and directions.


It might be a little of both, but it’s mostly learning how to drive. The neural network that’s powering this will just have a fairly short connection between inputs and outputs. It has a small number of hidden nodes, which won’t be nearly enough to really “memorize” the track. Given the paucity of information it has access to, it needs to learn how to drive.

I say “a little of both” because it’s going to optimize some of its properties based on the generic features of this track: the width of the road and the tightness of the curves. Given another track that’s similar to this in those dimensions, these cars will do well. But if you have a new track with tight, narrow curves a car that learned to speed through this might find itself crashing.


This looks familiar.

1 Like

That’s your life after they upload your consciousness.

This topic was automatically closed after 5 days. New replies are no longer allowed.