Stopping at red lights corner case

[Keith909man]

With the upcoming stop at traffic light FSD update coming out soon, there are unexpected corner cases that could pose a risk. Found this on Instagram where the system thinks lights the back of a semi are red traffic lights. I hope stuff like this get iron out before the official update before our cars slam on the breaks

video found here

@teslamax.dk on Instagram: “FSD I guess it's not possible to overtake before the traffic light turns green . Credit: SavedTesla (twitter) . . Don't forget to…”

 

[MrBadger]

Sigh.

There are a number of occasions where situational awareness appears not to be taken into account and this is one of them. What must the car be 'thinking' to believe that a traffic light junction moving at 70mph is a likely real scenario? Another example where continual situational awareness could rule out highly improbable scenarios.

Another example is the car braking for an overhead junction/roundabout at end of the offramp when the car never turned onto the offramp and was and continues to navigating a set route under the junction.

It strikes me that the car may be purely reacting to events on a frame by frame basis with little or no input from contextual awareness.

 

[Saghost]

It strikes me that the car may be purely reacting to events on a frame by frame basis with little or no input from contextual awareness.

There’s actually quite a bit of evidence in the car’s behavior that the current versions of Autopilot are only looking at a few frames at a time.

I think having the car build a persistent worldview and adjust the evaluation of the current frames based on that persistent understanding will be one of the keys of making FSD work - and likely on of the most challenging parts, especially the weighting of “what I think I see now” vs “what I thought my surroundings were.”

Tesla is well situated to do that, though, with all the cars running around to generate edge cases and shadow mode to trial different weightings.