Auto-Pilot - I give up, it's downright lethal

[Teddy69]

I've had the car 3 weeks, have done 1200 miles so far

Had the braking with car on left junction circa 6 times now, sometimes it was a warning (loud beeping), a few times it was full emergency braking. last time I throttled through it to prevent tail gate crash. Quite upsetting for passengers in the car.

This is just on cruise control, not lane control

I also have a 2 year old BMW X3 with active cruise control and 12,000 miles. It's never phantom braked on me once and I use that cruise control everywhere I can.

Now aware of the situation my foot hovers over the throttle any time a car on left appears looking to pull out.

Also had it trying to steering correct when I was in manual few days ago, it read the long repair running parallel in the road as an edge of the road thinking I'd crossed the edge.

So far a nice enough car but the "intelligent" features to me all appear very American centric, this car doesn't have a clue what a round about is.

My wife who initially was keen to get the car now avoids driving it calling it "scary" and sticks to the X3 (even though I've explained she doesn't use CC so it wouldn't happen)

 

[Cloggie]

I have to say I'm surprised by the number of apologies made for Tesla's 'suboptimal' implementation of TACC, which many cheaper cars and, especially cars that don't market their AI or other advanced software driving assistance capabilities, get right.

The fact that you have to keep your foot over the go-faster pedal in case the car has a brain fart is just compensating. Yes, the driver should always be in control, I fully subscribe to that, 100%. That doesn't however absolve Tesla from getting TACC wrong, creating dangerous situations and undermining one's confidence in the car. Not to mention the confidence and comfort of passengers. We all should expect a car manufacturer to do better, Tesla or another brand.


End of rant, back to my wine

 

[MrT3]

I also have a 2 year old BMW X3 with active cruise control and 12,000 miles. It's never phantom braked on me once and I use that cruise control everywhere I can.

and it's not just BMW. Any Audi, MB, Lexus, Jag, Volvo etc would all work correctly with zero phantom brakes.

More frustrating is that the car is sending data back to Tesla so they must know how often this occurs yet it is never improved/fixed.

 

[Teddy69]

and it's not just BMW. Any Audi, MB, Lexus, Jag, Volvo etc would all work correctly with zero phantom brakes.

More frustrating is that the car is sending data back to Tesla so they must know how often this occurs yet it is never improved/fixed.

does feel a bit like we're beta drivers for it, ACC is well established and reliable, I'd rather they fix this first before "full self driving"

Another issue I find is the system is truly terrible at figuring out what speed I'm currently driving in.

Turn on to the A9 at Aviemore heading south, and there's signs saying 50 with a lorry on them (basically the lorries can go 50 instead of 40 on this road) The cars sees the 50 and sets that for 10's of miles at a time, occasionally sees a proper sign, then back to 50. What's the point if it's so wrong a lot of the time (and it interferes with turning on cruise control etc)

Again the BMW seems much better in this regard with it's "dumb" camera

Hopefully tesla will improve on their regionalisation, but they seem well behind the established car brands in this regard.

 

[Type2]

does feel a bit like we're beta drivers for it, ACC is well established and reliable, I'd rather they fix this first before "full self driving"

Another issue I find is the system is truly terrible at figuring out what speed I'm currently driving in.

Turn on to the A9 at Aviemore heading south, and there's signs saying 50 with a lorry on them (basically the lorries can go 50 instead of 40 on this road) The cars sees the 50 and sets that for 10's of miles at a time, occasionally sees a proper sign, then back to 50. What's the point if it's so wrong a lot of the time (and it interferes with turning on cruise control etc)

Again the BMW seems much better in this regard with it's "dumb" camera

View attachment 687476

Hopefully tesla will improve on their regionalisation, but they seem well behind the established car brands in this regard.

We are beta testers. TACC/Autopilot/FSD are in seemingly permanent beta. I honestly don’t know how Tesla can charge £6800 for beta software that doesn’t work.

People are pinning their hopes on the big and (supposedly) imminent software update which, as with most things to do with Tesla, is way behind schedule. Personally I prefer to look at their track record, and their record is one of continually broken promises, missed schedules and the Musk spouting complete crap. Despite this last point, there are still plenty of people who hang on to his every word as if he’s the second Messiah . . . . . and even still believe his empty promises.

 

[gangzoom]

People are pinning their hopes on the big and (supposedly) imminent software update which, as with most things to do with Tesla,

I persume you have seen the FSD v9 Beta videos? Whilst its clearly not 'Full Self Driving' what Tesla have decent so far is pretty damn amazing.

The problem is some people are idiots, as we saw with AP1 from a few years ago. Unfortunately until Tesla can get the disengagement ratio to near 0 they will be very reluctant for wider fleet release because the minute it hits the wider fleet someone is going be climbing into the back seat whilst letting the car 'drive it self' past a school at 3pm whilst live streaming from their phone!! So the safest thing Tesla can do is what they are doing at present, and not give the general public access to their latest software.

Even the current software is so far advanced to anything you can buy its not even funny. Show me another car manufacturer that has implemented software that can 'see' lane marking on the road!

Unfortunately the more data you gather the harder processing gets, and until we see a wider fleet push out of the current vision only based software stack phantom braking will remain a major issue.

So yes software updates will be the solution, but the software is no where near complete and I doubt it will be ready for wider fleet roll out for another 12 months.

 

[pgkevet]

I persume you have seen the FSD v9 Beta videos? Whilst its clearly not 'Full Self Driving' what Tesla have decent so far is pretty damn amazing.

The problem is some people are idiots, as we saw with AP1 from a few years ago. Unfortunately until Tesla can get the disengagement ratio to near 0 they will be very reluctant for wider fleet release because the minute it hits the wider fleet someone is going be climbing into the back seat whilst letting the car 'drive it self' past a school at 3pm whilst live streaming from their phone!! So the safest thing Tesla can do is what they are doing at present, and not give the general public access to their latest software.

Even the current software is so far advanced to anything you can buy its not even funny. Show me another car manufacturer that has implemented software that can 'see' lane marking on the road!

Unfortunately the more data you gather the harder processing gets, and until we see a wider fleet push out of the current vision only based software stack phantom braking will remain a major issue.

So yes software updates will be the solution, but the software is no where near complete and I doubt it will be ready for wider fleet roll out for another 12 months.

New Level 3 Autonomous Vehicles Hitting the Road in 2020 - IEEE Innovation at Work

Of particular note is that Audi A8 with level3 is available Japan and a comment that many automakers don't plan to release until they have level 4 function working.
Crashing multiple rocket attempts to sort a problem is fine when they're unmanned but passenger vehicles really should be close to foolproof....

Assisted Driving Gradings | Euro NCAP

Note the position of the Model 3 in this evaluation..

You want another manufacturer that can see lane markings? But not for wide release until properly proven.

Mobileye demos self-driving car that uses cameras to get around

 

[gangzoom]

Crashing multiple rocket attempts to sort a problem is fine when they're unmanned but passenger vehicles really should be close to foolproof....

But how do you get to 'foolproof' without real world testing like Tesla is doing with FSD beta 9?

Rockets are governed by the laws of physics, you are in control of all the variables, get it right and it will be right everytime.

Driving unfortunately involves human interactions and trying to understand how to navigate human roads. There no real rules to this, the fact nearly every driver breaks the law by driving over 30mph in 30 zones every day shows how difficult it is to 'solve' FSD.

Tech demos are one thing real world usage is quite another, even Waymo currently you can argue is a tech demo given its geo fenced, monitored constantly and even followed by human teams, whilst only using route planning that avoids complex junctions.

Tesla FSD beta 9 is the closest thing to autonomous driving around at present, the fact it actually runs on hardware that is in every Tesla made since 2017 is nothing short of amazing. In comparison the Mobile eye demo cars have literally a boot full of computers.

I don't get how anyone can moan about Tesla's development FSD software development. Comparison to any existing lane keep assist really is a joke.

 

[pgkevet]

The point remains that autopilot has some major issues and it's amazing phantom braking hasn't led to some serious rear endings and likely motorway pileup. I was in one many years ago in one where I stopped short of the car in front and got hit by 6 vehicles in turn from behind.

Mobileye demo is unarguably impressive A boot-full of computers means nothing - mobile phones were brick sized not so long ago. You asked who else could visualise road markings..

 

[gangzoom]

Mobileye demo is unarguably impressive A boot-full of computers means nothing - mobile phones were brick sized not so long ago. You asked who else could visualise road markings..

So you can buy a car today that has Mobileye with road marking visualisation?

Let also not forget FSD Beta 9 is doing a whole lot MORE than that, and is running on the same hardware as most of the cars in the UK.

Come 2022, how much more advanced is the Mobileye equipped car going to be versus any Tesla running the latest code?

The point is Tesla have turned the world of automotive car development on its head, am currently using a laptop from 2013 that can still run the latest OS from Apple, yes I suspect it wouldn't get the next major Apple OS update, but its software that determines functionality not hardware.

No other car manufacture have got their head round this yet, out sourcing your software development to a third party like Mobileye leaves you not in control of your own product as software is quickly becoming the key differential in EVs. After all anyone can make an EV that will do 300 miles and crazy 0-60 times, just look at Hyundai/Kia. So whats going to differentiate manufactures apart from the badge in the future?

 

[Type2]

I persume you have seen the FSD v9 Beta videos? Whilst its clearly not 'Full Self Driving' what Tesla have decent so far is pretty damn amazing.

The problem is some people are idiots, as we saw with AP1 from a few years ago. Unfortunately until Tesla can get the disengagement ratio to near 0 they will be very reluctant for wider fleet release because the minute it hits the wider fleet someone is going be climbing into the back seat whilst letting the car 'drive it self' past a school at 3pm whilst live streaming from their phone!! So the safest thing Tesla can do is what they are doing at present, and not give the general public access to their latest software.

Even the current software is so far advanced to anything you can buy its not even funny. Show me another car manufacturer that has implemented software that can 'see' lane marking on the road!

Unfortunately the more data you gather the harder processing gets, and until we see a wider fleet push out of the current vision only based software stack phantom braking will remain a major issue.

So yes software updates will be the solution, but the software is no where near complete and I doubt it will be ready for wider fleet roll out for another 12 months.

Yes, I’ve seen some videos of people driving with the v9 beta on streets in the US described as “busy” or “challenging” but which look like a Sunday afternoon over here. None of the ones I’ve seen can be described as an in depth, technical analysis of v9 capability. And we all know that what might work in the States doesn’t necessarily translate over here.

I certainly hope vision is better in v9. In the current version the way other vehicles, traffic cone, wheelie bins and traffic lights dance all over the screen is ridiculous.

I maintain my point that the Musk has been disastrously wrong in his predictions in the past and I have absolutely no faith in his predictions for the future. Remember what he said in March 2015?:

"I don't think we have to worry about autonomous cars because it's a sort of a narrow form of AI. It's not something I think is very difficult. To do autonomous driving that is to a degree much safer than a person, is much easier than people think.... I almost view it like a solved problem."

And in February 2919:
“I think we will be feature complete, full self-driving, this year. Meaning the car will be able to find you in a parking lot, pick you up and take you all the way to your destination without an intervention. This year. I would say I am of certain of that, that is not a question mark.”

Yeah right, Elon! He clearly thought the software at the time was capable of Level 5 autonomy. He was completely wrong. Now he’s telling us v9 is the answer to everything. Given his track record I, for one, find he has no credibility whatsoever.

 

[Battpower]

The only time I use it is in nose to tail traffic

Yes. Even then I am slightly conscious that some unexpected and inexplicable glitch response might trigger a random misbehavior. But so far, it hasn't. I wonder if control switches to ultrasonic data for close / slow queuing situations?

Phantom braking struck as I was passing a large truck which was to the left of me

Despite all the posts that phantom braking 'hasn't effected me', this is a very common scenario for many (passing a truck). IMO when you see visualisations of adjacent vehicles twitching from side to side, it's not so unlikely that sometimes that twitching will be linked to a erroneous reaction like sudden slowing.

Those that pooh-pooh phantom braking and claim it can be avoided by hovering over the brake pedal haven't experienced

It's real! It is a thing! Of course it's just the result of a machine's 'best attemp' view of how it should navigate. There are so many nuances that anicdotal 'evidence' posted by owners can be misleading.

Theoretically a 3D approach to fsd sorts it - but there'll be a whole slew of new corner cases.

That's my concern too. Radar limitations seem fairly well understood and quite specific. Vision based anomalies will be based on very complex processing and NN's that seem to implicitly have many edge cases.

'phantom' braking occurs maybe 1-4 times per round trip.

..... which you are suggesting is not too bad? I would tend to agree with your experience which matches mine. However I would say more likely 0-1 real phantom braking event, 1-2 slowing with fairly obvious cause that driver can respond to plus may be 1 more marginal braking event that I could focus on once I'm having a bad braking drive. For me, anything like this level of incorrect braking instigated by the car is way off acceptable. I would say one modest incorrect slowing event every couple of thousand miles is not too unreasonable.

an unpredictable situation ahead (eg a car trying to cross a dual carriageway), I will disengage AP, until the hazard is passed,

I struggle to take 'unpredictable' situations into account when I'm driving!

Be interesting to see how vision only improves things as it gets used globally in varying situations.

Yes, road layouts and markings as well as setbacks and common occurrences in proximity of drivable space do vary considerably between countries. Eg: box markings for bus stops in UK.

all appear very American centric,

And that is my biggest concern. Given how tough progress seems to be on fixing basic TACC / AP which is probably less country specific, and FSD City for US, how can other countries reasonably anticipate viable FSD in the foreseeable future.....?

frustrating is that the car is sending data back to Tesla so they must know how often this occurs

...... which ties in with the above point. I don't see any evidence that the many situations my car has 'experienced' on UK and French roads have contributed to any improvements in response.

What's the point if it's so wrong a lot of the time

Yes. It does keep the driver on their toes though, but why would Tesla allow these behaviors to persist and beta or not, it seems debatable to 'sell' software that costs so much when you are more realistically paying for membership of a somewhat unequitable testers' club.

We are beta testers. TACC/Autopilot/FSD are in seemingly permanent beta. I honestly don’t know how Tesla can charge £6800 for beta software that doesn’t work.

Yes. As per previous comment. And is 'beta test' just a handy get-out to avoid sale-of-goods rules?

it's amazing phantom braking hasn't led to some serious rear endings

It really is. Unfortunately, this is a possible argument supporting the problem not really being significant. But it is very hard to know just how many accidents have been avoided by swift driver intervention.

 

[pgkevet]

The other car makers are still providing practical and economic (to buy) vehicles for the mass market that can't afford a Tesla as well as quietly creating safe and mature autonomous systems and in time will offer them out
Lane keeping, TACC,, ergonomic driver assists are all mature technologies Tesla is still selling vapourware, albeit their promises and demos are getting more ambitious.
To differentiate themselves from other marques will be as before - essentially all marques offer hatchbacks, small family saloons, SUV’s and higher end options with various warranties and build qualities, dealership chains and prices and their loyal or not so loyal followers for German or Japanese, Korean etc brands. They don't bother filling vehicles with unnecessary sensors until the software works. There’s a good chance that by the time level 4 autonomy actually gets licensed UK my 2018 FSD S will be past it's use-by date your car will be out of warranty and repairable only at extreme cost and we'll both be looking at market alternatives.

Come 2022 Tesla software will still be unlicensed apart from the US, and still be beta. Mobileye and other developers will be quietly working on their versions ready to pounce with their safe product once ready.

What is the point of road marking visualisation in a car when you can't trust the car to drive by it There is a section of the A5 southbound where the single carriageway splits to two lanes for overtaking uphill. If there’s no traffic then the car sees this and stays left. If there's a slow queue then it doesn't take notice of the widening road..keeps you dead centre and then jinks hard left when it finally sees the new middle dashes. To counter argue that one shouldn't use the software there - well it shouldn't allow one to. It smacks of early computer days - remember the "Keyboard not present. Press F1 to continue" days?

 

[Battpower]

"Keyboard not present. Press F1 to continue"

Oh yes.

That caught me unawares (your post). How time flies.

 

[Battpower]

practical and economic (to buy) vehicles

I just got a Kona to replace my Zoe. I can easily hit 6 miles per kwh, and the 64kwh battery seems well able to manage 300 miles with a mix of normal driving.

Loads of gimmicks like hud with speed, lane keeping and nav info. Ventilated seats, Apple & Android integration and auto cruise control. 70kw dc charging.

And sub £30k for low miles demonstrator.

It makes it a tough call to justify the MS, but the Tesla is obviously still in a different class and charging network is a big deal still for long trips..... but becoming less so.

No question that Tesla needs to deliver on many of its promises or seriously upset a big proportion of its owner base just as other manufacturers are bringing more mature products to market.

 

[pgkevet]

Oh yes.

That caught me unawares (your post). How time flies.

If your memory is long enough then names such as Osborne, Tulip, Apricot, Dragon and others are known. Arguably, Apricot had the best looking PC's with memory and processors to suit. Everything got 'destroyed' (in the sense of innovation) by the format release of the IBM compatible. Only Apple survived that due to their niche publishing software and portrait format. Back then it was a C> prompt for everyone. Windows GUI only happened when MS spotted and bought out the guys who came up with it (cant remember their name).
I’m sure Tesla will prevail in software but to think that others aren't biding their time is foolish.

 

[Fursty Ferret]

Note the position of the Model 3 in this evaluation..

The appalling EuroNCAP score is due to the way Autopilot interacts with the driver. I'm also not convinced about the impartiality of EuroNCAP given that the videos on their tests frequently show the vehicles twatting cyclists and pedestrians but still getting high scores overall.

The Tesla Model 3 excels in the level of Vehicle Assistance but fails to balance that high level of support with a similar level of Driver Engagement leading to possible overreliance.

 

[Battpower]

Windows GUI only happened when MS spotted and bought

iirc Xerox had a big development team working on mouse control and GUI's originally linked to printing and publishing using their copier technology.

Sorry, OT.

But technology moves on, and it serves to remind that all players have their contributions to make.

 

[TestPilot]

I just don’t use TACC at all, it’s too unpredictable on A roads.

I would much prefer a dumb cruise control option.

And before anyone chips in and says “It’s no worse than TACC on any other car.”, I wouldn’t use theirs either if they’re as bad or worse.

 

[jpm777]

I find on motorway sections with narrowed lanes, roadworks with a reduced speed limit, this triggers phantom braking lots. It's a real shame as this is a prime situation to use autopilot, so unfortunately I avoid using it now in situations I know it will trigger.

I use mine in this situation every day and thanks to operation Brock on the M20 have had lots of practice. I may be lucky, but I use TACC every day for the 16 mile section each way without having any phantom braking incidents.
I have had them around a year ago, but when passing a truck I now gently depress the throttle, not enough to change speed, yet this prevents any phantom braking for me. I know this isn’t technically the correct way to use the system, but I mostly cover the throttle when using TACC after a big phantom braking event previously. Using this method and only using it on motorways, which is about 95% of my journeys, I haven’t had a single incident in over a year and more than 11,000 miles.