Agreed. I sat in a new Model S today- nicely screwed together, I like the centre screen.
No luck getting one downunder....
-
Want to remove ads? Register an account and login to see fewer ads, and become a Supporting Member to remove almost all ads.
Tesla Software updates - Australia
- Thread starter Grenadine
- Start date
No luck getting one downunder....
Autopilot in Australia has in essence remained unimproved in about 2 years. They totally dropped the ball for us, all for the silly pipe dream of FSD. Even the comparatively trivial problem of speed sign detection is so unreliable it's unusable. Just yesterday (on 2023.32.4) driving behind a bus, the car set the speed limit to 40, again, or still, I should say. And the lane open indicators in tunnels and on the Sydney Harbour Bridge are still considered traffic lights worth slowing down for. Which is totally wrong behaviour even if the indicator is on red, it never means stop in a tunnel or on a bridge, it means switch lanes to the ones not having a red light. Another massive irritation is the incorrect use of indicators on a freeway onramp (sigh. motorway slip road): two lanes on the slip road, they merge before entering the motorway, when driving in the right lane and on AP, the car indicates LEFT when the left lane merges into the right lane I'm in. It's embarrassing.
What's worse, the latest iteration of the FSD stack reportedly doesn't allow them to cater to exceptions anymore - the entirety of the drive behaviour is drawn from the AI model derived from millions of hours of other drivers. So at best, AP is going to drive as good as the average driver. Which is... not very good, at least in my view.
What's worse, the latest iteration of the FSD stack reportedly doesn't allow them to cater to exceptions anymore - the entirety of the drive behaviour is drawn from the AI model derived from millions of hours of other drivers. So at best, AP is going to drive as good as the average driver. Which is... not very good, at least in my view.
Apparently they filter out poor driving behaviour. How they do that I don't know
Do you report bug each time it is wrong. I believe that is the only way for them to know, apart from the fact Tesla have FSD testers in Australia
I've stopped reporting bugs after a) Tesla Alexandria told me they only look at the bug reports when the user reports a problem with the service centre, and b) it takes up to 20 attempts for the voice recognition to actually understand I'm saying "bug report". My wife helped a little by suggesting I emulate a heavy southern US drawl for a one out of 10 success rate. "mmmmmbuuug report". No joke.
If they have FSD testers in Oz, I am flabbergasted why they have not approached me. I've had much closer contact than I care for with the Tesla service centre about many a bug over the past 3 years. I drive 30'000km a year, in about as many varied scenarios as you can here in Oz. I'm a former test pilot, then airline pilot, and now physicist/scientist and working as such, extremely methodical in my testing, I can't think of anyone more suitable to be frank... unless they want an average driver to say average things about an average product of course. In which case I am patently unsuitable.
If they have FSD testers in Oz, I am flabbergasted why they have not approached me. I've had much closer contact than I care for with the Tesla service centre about many a bug over the past 3 years. I drive 30'000km a year, in about as many varied scenarios as you can here in Oz. I'm a former test pilot, then airline pilot, and now physicist/scientist and working as such, extremely methodical in my testing, I can't think of anyone more suitable to be frank... unless they want an average driver to say average things about an average product of course. In which case I am patently unsuitable.
The employed full time drivers to do FSD testing in Australia and a number of other RHD countries. Those advertisements went out several months ago now from memory. I belive it is also for V12 so I wouldn't expect to see results for a little while yet.
You should have applied when the jobs were advertised if you feel that strongly about it. With that experience I'm sure you would have got a gig. One of my mates is doing it. 12 hour shifts behind the wheel on V12 in Victoria. Multiple cars doing it running around 24 hours a day basically. He can't say too much about it due to the agreement but it is keeping him pretty busy.
Even if I'd known about it, I already do have a job that keeps me busy 12+h a day... good to hear something is happening though. All the more confounding then why we haven't seen any measurable improvement with AP over the past couple of years. That's a long time to be stuck with such basic problems as I described above.
So stop complaining about Tesla not personally approaching you? 30k kms is hardly much driving. I did 45k kms in the last 12 months in my M3 and another 15k kms in my ford ranger, and I hardly feel like that is many Kms compared to many many other people.
Back to software, weird TACC problem in 2023.32.4. Just slows out of nowhere, seemingly where there are other roads entering to make a T intersection, slows rapidly approaching roundabouts and intersections but like 70m beforehand, well before you would press the brakes. I click TACC on often due to the amount of speed cameras around my area. Usually just continues along normally but not now. Hopefully reverts back again.
Tested it on a dual lane road and had no one behind me so I just let it do its thing to see what happened and it slowed from 70kmh all the way down to 30kmh before I got sick of it and accelerated. There was a road that joined on the right hand side is all I could put it down to.
Tesla tacc has not been fit-for-purpose since the day they ditched Mobileye for their own quarter baked system
@Skurfer The problem you describe is well known - and has many different causes and incarnations. I've provided Tesla with video and detailed logs for many of those over the past years only to get shrugged off. As @paulp says, it's not fit for purpose.
The only AP component I have no complaints about is autosteer. By and large works very well.
The only AP component I have no complaints about is autosteer. By and large works very well.
Autosteer would be massively better if it adopted the mercedes concept that allows the driver to temporarily take over without dropping out. Makes avoiding pothole etc much easier
Seeing that Tesla is so bullish on Tesla Vision, why not use it to detect potholes, and steer around them?
It would then have to decide what to do if going around a pothole had other consequences
A pothole is indistinguishable from a singular hard shadow, a flat piece of bitumen, or a little mound of concrete a truck dumped (an all too common occurrence on aussie roads) using the 1280x960 resolution cameras currently in use. Only half of those obstacles need to be avoided.
Developing a deterministic avoidance plan for multiple consequential conflicts arising from a non-standard action such as pothole avoidance is well beyond an AI system trained by driver behaviour alone. There are simply too many possible scenarios for a large enough training set to be available. So the only solution to this is a non-deterministic method, and that's in essence how Autosteer works today. It simply evaluates the best option for steering at t=now, and exactly now.
Compare this to a deterministic system like TCAS [1] in aviation. Every airliner is equipped with this system, monitoring surrounding air traffic and giving very simple aural and visual instructions to the pilots if a conflict arises. All actions given to the crew are in a single guidance axis (vertical) only, making it a relatively simple (and deterministic) system, because several actors in a conflict can be given instructions with predictable outcomes. It took many decades to develop, certify, and refine these systems. There are no cyclists suddenly appearing, or children running into the path. TCAS is orders (plural) less complex than what's encountered on the road.
[1] Traffic collision avoidance system - Wikipedia
Developing a deterministic avoidance plan for multiple consequential conflicts arising from a non-standard action such as pothole avoidance is well beyond an AI system trained by driver behaviour alone. There are simply too many possible scenarios for a large enough training set to be available. So the only solution to this is a non-deterministic method, and that's in essence how Autosteer works today. It simply evaluates the best option for steering at t=now, and exactly now.
Compare this to a deterministic system like TCAS [1] in aviation. Every airliner is equipped with this system, monitoring surrounding air traffic and giving very simple aural and visual instructions to the pilots if a conflict arises. All actions given to the crew are in a single guidance axis (vertical) only, making it a relatively simple (and deterministic) system, because several actors in a conflict can be given instructions with predictable outcomes. It took many decades to develop, certify, and refine these systems. There are no cyclists suddenly appearing, or children running into the path. TCAS is orders (plural) less complex than what's encountered on the road.
[1] Traffic collision avoidance system - Wikipedia
I would not want it to be open to public testing. That just opens it to knuckleheads and boof heads doing crazy things. Paid drivers are the way to go
Who knows, that is why I have not bought FSD. FSD will come soon enough when the car comes without a steering wheel. Until then anything apart from TACC is beta. There is a reason why its called beta.
Looking at it the wrong way. Im just parroting what Tesla says:
The Tesla fleet now in the millions and increasing. Tesla uploading lots of driving scenarios and apparently curating the good ones. Deterministic coding (If A then B, C, D, etc but if F then don't do B) is not possible. There are, as you say too may permutations and nuances. It is impossible to label objects in every possible scenario in a world view. Better for the system to learn from millions upon millions of good driving examples and scenarios. The rate limiting step is computer power - step aside bitcoin mining. So much better for a car to recognise to stop at a stop sign or a roundabout within the nuances of stopping at a stop sign and roundabout driving without explicitly programming all the nuances and permutations. "It just knows because its seen it done before".
A human driver gets better with experience - but with the experience of one driver driving in one lifetime
A FSD system gets better with experience - but with the experience of millions of other Teslas driving in multiple generations.
Lidar? Look at Waymo which is deterministically and explicitly coded. It can only operate in a geofenced location. The whole point of tesla is that it wants to be able to operate in a multitude of locations, all under different scenarios without explicitly coding for each scenario.
Similar threads
