Using EfficientNet instead of ResNet, depth detection algorithms, better object classifiers are available than just a few years ago.
I believe that there's less than 5 years between the first company achieving self driving and many other companies.
Compared to 2012, it now takes 44 times less compute to train a neural network to the level of AlexNet.
How that affect self-driven car is still to be determine, but progress is very real
https://openai.com/blog/ai-and-efficiency/
However this gain also directly translates to a real-world result improvement: You can now train a lot more at the same cost.
Using EfficientNet instead of ResNet, depth detection algorithms, better object classifiers are available than just a few years ago.
I believe that there's less than 5 years between the first company achieving self driving and many other companies.