I wonder what will happen when they go to single stack. Will NoA continue to be basically a L3 system or will it be obviously L2 like FSD Beta?
I always figured FSD Beta was superset of NoA.
I think that depends on where you think the shortcomings of city streets are and how much an issue they'd be in a different situation...
From my use so far it's been 2 main areas:
1) Perception issues seem to mainly stem from:
a) The location of the side facing forward cams in the b-pillars limiting visibility at some occluded intersections (and why i think they eventually need more forward mounted side cams to ever get city streets above L2)
b) Lack of persistent memory of objects- which allegedly is an issue they already have some code to address, we'll see.
2) Planning/policy especially for turns- this could be related to 1a, it can't always see well enough where the future path of the turn is so it's hesitant or just gives up partway through-- this is one of those things (for city streets) NN training SHOULD get much better at over time, but AFAIK a lot of this is only now starting to transition to that from traditional code
So how does that relate to single stack?
1a is basically a non-issue for NoA ODD on highways. You don't need to see around blind corners and check for cross traffic to make decisions- highways don't typically HAVE blind corners and they don't [B[ever[/B] have cross traffic.
1b would probably help a LITTLE bit on NoA for lane changing consistency- but it'd be anywhere from a neutral to tiny improvement
2) Again you don't really make turns like this on highways... I guess it might improve tight turn ramp/interchange behavior a little?
But overall, the major issues city streets have that make it far less reliable than NoA are issues that simply don't happen in NoAs ODD... so moving to single stack should be overall a non-issue or maybe a tiny improvement to what is (for me anyway) already a REALLY good system in real life use.
I'm impressed. You are extraordinarily lucky. It's hard for me to imagine using Autopilot for that distance without having to disengage. Though it seems like an L3 system would have to go many multiples of that to achieve human level safety.
My daily drive (when I had one pre covid, I still do it 1 day a week now though) was ~75 miles... ~70 of that interstate highways.
On ramp to off ramp just like it says on the tin, no issues ever with one exception I can actually think of.... and now that you bring it up it's one way single stack might IMPROVE noa
There was an overpass bridge on the route for a while that in the far left lane was missing the line seperating the lane and the shoulder.
So for the period that was true- AND if you happened to be in the far left lane for that 100 feet or whatever of the route, it'd do that thing where it thinks your lane got SUPER wide and being to drift a little to the left for a second before "seeing" the paint post-overpass and snapping back into the lane center.
It wasn't really a big safety concern because it was only for a second or two--- but city streets seems to do FAR better staying where it should for a second or two when markings vanish... so presumably that WOULD help in that weird situation.
Anyway- overall the route is fairly easy (mostly on one interstate, then an interchange to a second for a few miles, then the reverse the other way) and well marked.... so the system works great.
I've also taken a number of road trips, but none more than maybe 300-400 miles in a day, and usually primarily interstates there too, where again it's been pretty flawless....all have been in the US southeast (most driving in NC specifically-- but some in SC, GA, VA, and TN).
I've been plenty of human-driven wrecks in that time, but NoA has never come remotely close to being in an accident.
https://www.google.com/maps/place/33°39'49.1%22N+117°47'50.1%22W/@33.66364,-117.7977996,241m/data=!3m2!1e3!4b1!4m14!1m7!3m6!1s0x80dcde7313cf89a3:0xa594fd03d9382684!2sUniversity+Dr,+California!3b1!8m2!3d33.6574767!4d-117.8304621!3m5!1s0x0:0x3c4e8f92d8de506b!7e2!8m2!3d33.66364!4d-117.7972449
Oh that's weird as hell... it looks like an existing lane STAYS an existing lane with double yellows on either side at a glance
Obviously it MEANS for you to go left or right depending if you want the HOV lane or not- but that's not obvious with JUST the lines. Great example of an edge case to train an NN on though- once it understands that specific marking it should work perfectly from then on.
But to be fair- this is a
marking error in violation of Federal highway marking standards-
HOV lanes are supposed to use a double WHITE, not a double YELLOW.