Post
Ethan Mollick
@emollick.bsky.social
Apple team shows self-driving AI can learn entirely by practicing against itself - no human driving data needed
In testing, their system averages 17.5 years of continuous driving between incidents, far surpassing humans. All through self-play, not imitation. arxiv.org/abs/2502.03349
February 6, 2025 at 5:20 PM
10 reposts
77 likes
haven't read it yet, but if I understand self-play correctly, that doesn't amount to much. if your opposite player is a copy of yourself, you could develop meta rules which reduce accidents. but real traffic participants wont follow those rules
Interesting, it looks like a similar concept to what eventually turned into Nvidia’s cosmos / omniverse?
No human driving data needed for simulated driving, right? This is simulated, not real world?
I guess I’m confused, with that success, why they would shut down the program entirely?
Anything is possible in a simulation of sunny daytime and perfect knowledge of the road conditions. Irl driving is a non-ergodic system that regularly trips up AVs in surprising ways and which DL systems cannot reliably, accurately model. Limitation of modeling in general.
Alas, nothing will come from Apple. They shut their self-driving (and iCar) efforts down last year.