The whole problem seems to me that to get a 90% solution for a self driving vehicle is now no longer hard. The remaining 10% is
super hard and the degree to which the present day solutions tend to get these wrong is worrisome enough that if the quality doesn't jump up very quickly the whole self driving car thing could end up in an AI winter all of its own making. And that would be a real loss.
This is not a field where 'move fast and break stuff' is a mantra worth anything at all, the stuff that is broken are people's lives and as such you'd expect a far more conservative attitude, akin to what we see in Aerospace and medical equipment software development practices.
But instead barely tested software gets released via over-the-air updates and it's anybody's guess what you'll be running the next time you turn the key.
I agree with you that the software should have been able to detect that something was wrong either way, either it was already halfway in the wrong lane or it was heading to be halfway in the wrong lane, a situation that should not have passed without the car at least alerting the driver.
And from what we have seen in the video it just gave one inferred situation priority over the rest and that particular one happened to be dead wrong about it's interpretation of the model.
This is not a field where 'move fast and break stuff' is a mantra worth anything at all, the stuff that is broken are people's lives and as such you'd expect a far more conservative attitude, akin to what we see in Aerospace and medical equipment software development practices.
But instead barely tested software gets released via over-the-air updates and it's anybody's guess what you'll be running the next time you turn the key.
I agree with you that the software should have been able to detect that something was wrong either way, either it was already halfway in the wrong lane or it was heading to be halfway in the wrong lane, a situation that should not have passed without the car at least alerting the driver.
And from what we have seen in the video it just gave one inferred situation priority over the rest and that particular one happened to be dead wrong about it's interpretation of the model.