It might be that in the process of learning, algorithm decided, that it can pass a constant value at the speed output and use this one node just to steer.

But I'm testing this scenario right now, and all I can get are cars going on different curves. Are you sure, that you have restarted the process after making changes to the script?

After 100 generations thay actually have learned to move correctly, that's impresive. But they move very slowly. But for 100 nodes in a single hidden layer it takes them only around 10-20 generations to learn for me (not 80), and they race really fast.

That actually sounds right. For me, they took 80 or so to race what I subjectively thought of as "well", but they only took 10 or 20 - as you said - to roughly figure out what they were doing.

And I figured out what the behavior difference was: before they finally "figure it out", they spend most of their time spinning in place or moving in "epicycles", for lack of a better term.

Thanks for the project; it's been great fun!