-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Well I've posted on this the number of times, pausing in the median is common practice but it's illegal, certainly in my state. It's not a good thing to do because it blocks the view down the road for everyone else.
Granted this median is a little wider than .most of the ones we have but the comment stands.
I do like that Chuck tests this turn over and over with each build, but truly the correct behavior for FSD to recognize the view blockage, turn right and make a U-turn.
The most disturbing thing about these cases is the car doesn't always seem to recognize that it can't see. On top of that it's inconsistent about getting as far forward as possible.
Some improvement certainly, I think the goal is for it to become reliable unconfident about these situations when it can see, and then reliably rerouting* when it can't.
*And some mechanism for tagging the map locations where this occurs, so that other Teslas don't try the same thing and block traffic while deciding whether to bail out.
I look forward to seeing more unprotected lefts that are not vision-blocked. Anyway, this was a great and eagerly awaited benchmark for beta 10!
Daniel in SD
(supervised)
It doesn't always recognize vehicles coming from the left so that wouldn't help much.
The point is that can and should recognize that it can't see the road because of the fence blockage in Chuck's example. Is exacerbated by the camera placement / angle of view issue, but that's a slightly different topic.
If it can't see, it can't safely pull into the lane. If it doesn't know that it can't see, then that's a bigger problem.(The creeping behavior suggests that it does know, at some NN level, that it can't see yet, though various aspects of this are still disturbingly inconsistent.)
As a follow-up comment, I think it would be great to have a very similar set of tests from an entrance road that is not blocked by a structure like this. Of course the one Chuck has been using is a tougher test, but we're having some debate about the core problem - is it a problem of the blocked sight-line, or is it something the car couldn't do anyway?
I feel guilty suggesting more tests when Chuck is already doing so much, but I think it would be an interesting comparison to enter the very same roadway another block or two over, from an intersection that has no difficult blockage of the corner.
stopcrazypp
Well-Known Member
Sure, a lot of things where it fails now are not edge cases, but that doesn't change the Seattle monorail system from being an edge case. The Seattle one (out of less than the 10 total in the US) is the only one I see where there are road level curbless gaps between the pillars that go along public roads, and not only that, it is legal for cars to go between the pillars to reach the other side (so the move the Tesla attempts to do is not illegal, if it was able to perform it successfully).
List of monorail systems - Wikipedia
As for a generalized solution, that's the L5 aspiration part, but something Tesla is not aiming for in the short term (as per DMV doc). City Streets will be fully L2, so it does not need to be a generalized solution. So even if it completely fails on the monorail test, that doesn't affect its release in any significant way. There are however other things that would, that I will point out later.
Well you don't have to point to the manual at all, the fact it's an L2 system pretty much means by default it's not Tesla's fault, the driver bears the legal responsibility. There may be rare cases where the car may do a move that the driver can't possibly override (in which case might be considered car manufacturers' responsibility), but so far no evidence that's what is happening.
We don't have safety related stats that Tesla may have, so no visibility on that question, but an obvious reason why Tesla is waiting for wide release is the current version is simply not good enough to be a useful feature.
You asked about the difference with release of AP previously up thread. Well the difference is that even under the most basic AP1, it was able to do the primary job of keeping the car in the lane 99.9% of the time, which is what won so much people over (vs other lane keeping systems that fall apart with even slight curves or ping pong between lanes). For City Streets, the primary function that is added is finally the car being able to make left and right turns on intersections. And on that front, for unprotected lefts, thanks to Chuck's videos we see v8.x well under half successful, v9.x roughly half or very slightly over, and the latest v10 is still not close to reliable in doing them. That's simply not good enough to be useful yet.
Last edited:
Then it need to add future visibility into its path planner optimization. So it ends up in an angle where the rear facing cameras can see the coming traffic. I think this will be a few versions into the future where they have done this. Fortunately with Tesla’s approach the can just add another term to their cost function, validate and it should work…
So are you guys excited about getting to do your own videos?
So I'm just wondering if you guys are going to hit the button next week?
diplomat33
Average guy who loves autonomous vehicles
You are going to have a blast! If they feel this comfortable I think we'll see quite the progression from here. I also wonder how far along are they with Dojo build out?
When do we see some videos from the UK and continental Europe?
powertoold
Active Member
powertoold
Active Member
Knightshade
Well-Known Member
...wut?
Teslas insurance algorithms explicitly include using autopilot.... (for example forced disengagements from ignored AP alerts count against you)
Also since you seem continually unclear, FSDBeta (really city streets) is an L2 system, not L4.
It's interesting that they are blocking the release from their newest, most expensive cars too.
Similar threads
