What I got from speaking to one driverless perception engineer is that the best traffic light recognition is at 99% now. That doesn't seem reliable to me, but if a human is suppose to augment it, then I guess it is passable. Still seems very dangerous to me, because people will tend to tune out and not watch the traffic lights closely.
Is that 99% internal at Tesla? Or were they referring to the industry as a whole?
In any case I'm firmly of the belief that self-driving cars need to be equipped with V2E (vehicle to everything) technology so they can be told not only the status of the light, but when it will change.
That way it accomplishes multiple things.
It augments the >99% accuracy of the cars ability to visually identify the light
It can help the car determine the correct speed to go to hit the green light
It can eventually be used to optimize traffic through a light
It's going to take awhile to get there so I see true self driving being a slowly expanding white list of areas that are modernized for it.
For L2 driving I share your concerns that it's going to be dangerous especially at that 99% level.