-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Software Update 2019.5.x
- Thread starter Penfold
- Start date
-
- Tags
- firmware update
squarepeg77
Member
I got 2019.5.14 on Friday and drive about 100 miles yesterday in my 2018 75D Model S.
Lots of phantom braking on both Autopilot and TACC, to the point I switched it off. This is the worst Autopilot I have experienced.
I’ve also had two media computer freezes/reboots in two days.
A couple of things I noticed that seen new to me, upgrading from 2018.50.6:
1. On the Autopilot display on the dashboard it now seems like it has locked-on to two cars - the one in front in my lane and the one in the next lane to the right.
2. When cars are turning/road curving, the dash display renders the cars at an angle, instead of just translating a forward-facing car. Is that new too?
Lots of phantom braking on both Autopilot and TACC, to the point I switched it off. This is the worst Autopilot I have experienced.
I’ve also had two media computer freezes/reboots in two days.
A couple of things I noticed that seen new to me, upgrading from 2018.50.6:
1. On the Autopilot display on the dashboard it now seems like it has locked-on to two cars - the one in front in my lane and the one in the next lane to the right.
2. When cars are turning/road curving, the dash display renders the cars at an angle, instead of just translating a forward-facing car. Is that new too?
Decibelle
Member
Does this version take care of the problem with the iPhone starting to automatically play music silently when entering the car? I’m on 2018.50.6 and still have this problem.
hybridbear
Active Member
diplomat33
Average guy who loves autonomous vehicles
These changes make me think that the software is slowly being updated to prepare for FSD. These changes seem like they are hinting at things that would be needed for FSD.
Locking onto two cars was introduced around the same time they introduced NoA and enabled the side cameras. It happens when AP thinks the neighboring car is slightly in your lane, or perhaps moving into your lane. This, I believe, is the cause of much of the increase in phantom braking events since NoA/side cameras landed. The upshot is that TACC is less likely to barrel into a vehicle that actually is entering your lane... but of course it gets it wrong sometimes and locks onto a neighboring vehicle unnecessarily because it has very limited precision in 3D locations of cars around it.
squarepeg77
Member
Interesting. We don’t have Navigate on Autopilot yet in the UK.
The car in the left lane is more likely to be going slower than you when it pulls into your lane (in the UK), so that’s the one it should be monitoring, IMO.
I also noticed that the display showed cars near to or crossing the lane lines, when they were in fact nicely in their lanes.
Yeah, it only gets very rough depth information from the cameras, and very rough bounding boxes around the cars. So it cannot reliably determine the actual position of the cars. You really need precision of a few centimeters, but from what the display shows, it's clear they are lucky to get it within a couple of feet. And furthermore it bounces around a lot.
Yeah, their NN isn't very accurate at judging distances for surrounding cars.
I suspect this is why Navigate on Autopilot is going to FSD. Hardware 2.5 isn't going to cut it even for NOA because of how poorly it estimates distances. Which means everyone that purchased EAP is probably going to eventually get HW3.
I wonder why it has such rough granularity. If you notice, it jumps back and forth at least 2 feet. You would think it would be inaccurate, but smoothly moving around. It seems to be quantized at a very granular level.
So there are two components -- angular position and distance. The NN gives you angular position directly from bounding boxes, but if you look at any of the videos from rooted Teslas, you can see that those bounding boxes sort of jitter and wobble a lot, which translates to angular position wobble/jitter.
So then there's depth. Cameras don't measure depth directly so the NN must be sort of guessing based on apparent size. Cars in neighboring lanes are REALLY close to the side cameras, so the cameras don't even see the whole vehicle. They only see a snippet, and somehow the NN is going to accurately gauge distance? Not a chance.
So take the sine of the error in angle, multiply it by the distance, and you get the absolute position error due to angular error. Now add in the uncertainty in distance, and you get a really big "error box" (actually looking more like a swept section, but close enough to a box). The car could be anywhere within that error box.
It's this magnification of angular error by distance, combined with the uncertainty in distance, that makes this so problematic. This is why you would generally rather have lidar for this sort of thing.
squarepeg77
Member
Does the NN not use the lane lines to help determine the position of the other cars?
I can see how judging distance from an (incomplete) view of a vehicle would be hard, but the relative position of the lines and the tyres/wheels seems a strong cue.
I can see how judging distance from an (incomplete) view of a vehicle would be hard, but the relative position of the lines and the tyres/wheels seems a strong cue.
The system now locks indeed on the car in front of you, as well as the car in the (faster) adjacent lane, and that's because AP2+ now does the same thing as AP1: it won't overpass a car going slower than you on the highway, if that car is in the faster lane (left here, right in the UK).
It however does that a bit better than AP1 as it leaves quite a distance from the left (right in the UK) car, such as it's easy to initiate a lane change.
If the car is quite close to the vehicle, the cameras likely can't see its tires. Or in some circumstances maybe they can, and maybe the problem is that the AP2 NN is still terrible. I think both might in fact be true at the same time.
2019.5.14.1 also... which is weird. Both on 3's.... now it's getting confusing. I first thought one would be the final one for us here in North America, but with 25, who knows... (14.1 is USA, 25 is (again) Europe...
Similar threads
