FSD / AP Rewrite - turning the corner?

[Drew57]

I hope the rewrite does turn the corner because unfortunately it's not relaxing having to anticipate something happening randomly when the car is in control - unexpected braking, swerving, excess speed into some corners, displaying but not responding to speed signs etc. I'm a great admirer of what Tesla have achieved but the car shouldn't be causing us occasional embarrassment in front of our passengers or other road users - but it does.

Don't get me wrong, driving a Tesla manually makes it the most relaxing and enjoyable vehicle I've experienced & I expect they will eventually achieve a stress free level of autonomous driving but sadly it's a long way from that in its present form, some of it caused by the quirks of our particular road network.

As I'm retired, I rarely use motorways these days & don't expect to keep the current car for more than 3-4 years so I've held off from purchasing FSD but may well add it to the next under a subscription model. I'm in the category of 'want' FSD but don't 'need' it so will follow progress with hopeful anticipation that both the software and UK regulatory approval will bring it to a level where it becomes truly worthwhile.

 

[Glan gluaisne]

@Drew57 , I'm in a similar position, I've been retired for ten years now (retired a bit early at 57) and my view of my driving and car requirements has changed radically since I stopped work.

I have EXACTLY the same view as you about driving my Model 3, and what sounds like a similar usage, as I don't often drive on motorways (nothing against them, it's just there are none on route to the places we tend to visit). Driving manually is usually fantastic fun (except I have had a few false emergency interventions by the car that detract from that a bit). Driving using either TACC or AP can be damned stressful at time, waiting for the next surprise from the car. I view FSD as being another gadget that might be very useful a few times a year for me, if it doesn't throw the same surprises that AP does at the moment.

 

[firetrax]

@firetrax , I get your enthusiasm and desire to never see anything critical written about Tesla, but can you honestly say that the Model 3 doesn't have major failings when it comes to the way the driver assistance functions behave?

I honestly cannot believe that any Model 3 owner can think that the driver assist functions work well, as far as helping, rather than hindering, the driver in everyday driving conditions (i.e. not using FSD). This forum, like every other Tesla forum, is full of reports from owners about potentially serious safety issues, like phantom braking and random steering inputs, things that happen for no discernible reason. The fact that the only visible indication is a brief warning, flashed up well to the left of the driver's sight line, doesn't help, as when something like this happens the last thing you're able to do is look down and left to try and see what the car's doing.

The car is fantastic to drive, but there's never a moment when I'm not thinking about the need for me to take rapid corrective action when the car decides to do something unpredictable. There's never any prior warning when these things happen, either, and that seems to be one of the concerns that NCAP had, the disconnect between the car systems and the driver. The first the driver knows about something going awry is when the car takes action, often quite aggressively, and simultaneously sounds the alarm gongs. I've driven other cars with driver assist functions, and none have behaved in this way, and certainly none have ever done things like the aggressive swerve for no reason that my Model 3 seem to do from time to time (even now I'm not sure what causes it to do this).

We are talking about the NCAP report and I'm correcting your factual inaccuracies. There is currently no widely available system that can drive well automously outside of motorways and dual carriageways. That's why I only use autopilot as currently intended on these roads and not on inside roads and have had no material issues with it in that use case. As per the NCAP report autopilot is far superior to other systems for its current intended use case i.e. motorways and dual carriageways.

 

[Medved_77]

Nice video of FSD handling a roundabout pretty well:

https://twitter.com/kimpaquette/status/1319663469860356096?s=09

 

[Glan gluaisne]

We are talking about the NCAP report and I'm correcting your factual inaccuracies. There is currently no widely available system that can drive well automously outside of motorways and dual carriageways. That's why I only use autopilot as currently intended on these roads and not on inside roads and have had no material issues with it in that use case. As per the NCAP report autopilot is far superior to other systems for its current intended use case i.e. motorways and dual carriageways.

Frankly, citing opinion as "fact" doesn't seem morally right to me, and I can find zero evidence that NCAP somehow made up stuff just to diss Tesla. Apart from anything else, why would NCAP wish to penalise Tesla unfairly? I can't see any possible motive, especially given the high rating they gave to the safety backup aspects of their assessment.

The evidence from ordinary Tesla owners for the Model 3 failing to work cooperatively with the driver, is overwhelmingly compelling - the Model 3 has very definitely got some serious issues with random stuff that just happens for no apparent reason. Perhaps your Model 3 driving experience has been different to most of us, and you've never experienced the issues that many here have, if so, it might be useful to pass this on to Tesla, as a counter to the many reports they've had of the Model 3 behaving erratically at times.

I can understand why NCAP suggested a HUD as a possible benefit, as, although the central screen is an outstanding feature of the car, it is far from ideal when it comes to being able to spot things that have caused the car to throw a wobbly, or as a way of giving any early warning information to the driver. Out of many dozens of false braking/steering interventions, I've only been able to spot something on the screen a couple of times (mainly because having the steering wheel yanked hard tends to increase the focus on looking where I'm going). Maybe that would be less of an issue (both for me and NCAP) if the screen was in the driver's eyeline, better still if the driver assistance feature gave early warning of something it interpreted as being a hazard.

 

[m3gt2]

This is very impressive, if only they'd transfer FSD to be locked to account rather than vehicle! https://twitter.com/wholemarsblog/status/1319713084613652480?s=21

 

[kavyboy]

This is very impressive, if only they'd transfer FSD to be locked to account rather than vehicle! https://twitter.com/wholemarsblog/status/1319713084613652480?s=21

That's the most impressive video yet.
I am used to seeing, on AP1, the car perfectly detecting the road lines and then just driving over them - lane departure warnings included - for no reason. This clip makes me a lot less cynical!

 

[LongRanger]

One of the things I would really like to see is a user-configurable “envelope” of the safety systems to a greater degree than is currently offered.

Late, Medium and Early don’t make a difference to the car trying to apply corrective avoidance when on a country lane (not on TACC, or AP), just driving along and suddenly it tries to steer you into the opposite lane because a leaf blew out of a hedge or a large weed/fern/shrub had decided to grow on a verge.

If I could adjust the car’s sensitivity or it could learn based on areas of focus that I choose, I would have much more respect for the software developers who manage the sensor/safety part of the software product. I still think very few of them drive, and even fewer of them have ever driven on roads that we frequently encounter in rural UK.

Once they have got this level of detail sorted out then I would start to become more of a supporter and trusting of the fancy stuff - but it’s more sensitive to everyday non-hazards than it needs to be and that makes it annoying.

 

[Wol747]

>>If I could adjust the car’s sensitivity or it could learn based on areas of focus that I choose, I would have much more respect for the software developers who manage the sensor/safety part of the software product. I still think very few of them drive, and even fewer of them have ever driven on roads that we frequently encounter in rural UK.<<

I've said it (many times!) before, but I can't see FSD in it's true sense EVER being realistic outside restricted areas, let alone in "rural England", until/unless it incorporates REAL artificial intelligence.
Experts can't even define what intelligence is and despite all the hubristic noises coming from the labs I don't think we are anywhere close to anything more than simulations of AI.

And true AI might allow true FSD - but all sorts of worm-cans would be opened then!

 

[gangzoom]

I've said it (many times!) before, but I can't see FSD in it's true sense EVER being realistic outside restricted areas, let alone in "rural England", until/unless it incorporates REAL artificial intelligence.

It depends on how you classify driving. Is it a process or does it require consciousness?

AI neural nets are far far superior to human brains at pattern recognition and predictive powers. The realisation by Google 3 years ago neural networks learn best WITHOUT human interference is whats driven the AP rewrite, and I don't think anyone can argue how impressive it is versus the old AP software which was written with human guidance.

Does playing Go requires more or less attention from the human brain versus driving? If the answer is less than humans are already beaten as demonstrated by AlphaGo Zero. Is driving really that hard if pretty much ever human being on this planet can do it?

It really is an amazing time we live in, this stuff is bleeding edge AI tech which soon all of us will have access to.

The limitation of AI neural network tech maybe our own brains inability to see its possibilities. Remember these AI networks can 'learn' non stop all day 24/7, whilst having 'real life' data to study from a database bigger than any human brain can even comprehend. They can also learn in parallel, ten/twenty/thousand networks all generating better and better predictive algorithm non stop 24/7, versus our brains......are you sure your brain really is that much better at learning?

Personally I don't think we (humans) should be so arrogant about how 'special' our brains are. At the end of the day each neuron is acts virtually the same as a transistor, but we are limited by physiology and time when it comes to learning. Even if you wanted to learn 24/7 you cannot, sleep/eatting/gathering learning materials all get in the way. Our maximum active learning attention span in less than 1 hour, 1 hour versus 24/7 non stop.

Does driving really require intelligence? I guess we are on the cusp of finding out .

 

[GeorgeSymonds]

Hands up those dismissing the NCAP report who have actually spent any time personally driving more than one of these systems reviewed (not a previous generation version or had a conversation with a man in the pub whose best friends daughters class mates father said...)
- I recon it will be none

And hands up those who agree with NCAP who have spent time with more than one system
- me

 

[Js1977]

Hands up those dismissing the NCAP report who have actually spent any time personally driving more than one of these systems reviewed (not a previous generation version or had a conversation with a man in the pub whose best friends daughters class mates father said...)
- I recon it will be none

And hands up those who agree with NCAP who have spent time with more than one system
- me

Hands up if your agree to disagree

 

[Sparkeur]

Nice video of FSD handling a roundabout pretty well:

We're going to be seeing a few more of these videos coming out in the next weeks. If they are a precursor to what is about to come, I think a number of genuine complaints about the current FSD will just fade away.

 

[Wol747]

>>Does playing Go requires more or less attention from the human brain versus driving? If the answer is less than humans are already beaten as demonstrated by AlphaGo Zero. Is driving really that hard if pretty much ever human being on this planet can do it?<<

I would argue that playing GO is fundamentally easier that driving in that it plays according to well defined rules. A computer can riffle through trillions of moves and use its neural net to come out with a move that has such and such of a positive outcome.
Driving is different - there are many rules but also many other factors such as those who don't obey them, those who mistake road markings, animals that don't care anyway, objects in the road, signs that point the wrong way, splashes of mud that suddenly obscure cameras, burst tyres, well you get the picture.

Intelligence is more than just massive computation ability. I accept that what the industry has achieved is most impressive, but I think that last - critical - 10% is going to be much more challenging that is recognised. Much more.

 

[MrBadger]

A decent length non city drive. Not good just before 8 minutes (still watching so may be more) with truck position in visualisation - thought that glitches like these were meant to be a thing of the past? So maybe reinforcing that 'the (fundamental) rewrite' is not in action at this point?

 

[Dilly]

Yes the truck was a bit of nonsense but that part of the journey wasn’t on the new beta.
The red lines denoted when it was being used.
Bit scary on some of the turns, but I guess no worse than driving yourself. It’s a bit like a passenger!
Don’t know why not all the journey was on the rewrite. Perhaps it’s a mapping thing.

 

[Dilly]

A decent length non city drive. Not good just before 8 minutes (still watching so may be more) with truck position in visualisation - thought that glitches like these were meant to be a thing of the past? So maybe reinforcing that 'the (fundamental) rewrite' is not in action at this point?

One thing I noticed in the video is that under the speedo, mph is replaced with the word’ hold’ when stationary.
Not sure if that’s a region thing or a change.

 

[Strategium]

A decent length non city drive. Not good just before 8 minutes (still watching so may be more) with truck position in visualisation - thought that glitches like these were meant to be a thing of the past? So maybe reinforcing that 'the (fundamental) rewrite' is not in action at this point?

These recent videos showing examples of the new 4D FSD beta are definitely impressive. My understanding however was that at present the 'rewrite' 4D system was only being activated when the car entered the sort of city streets in which NoA currently does not operate. Therefore, what we are seeing on these videos on a longer journey like this is a hybrid between the car using the same NoA functionality most of us have on 2020.40.8 (but here with the US version of this that is not restricted by EU regulations) plus the new 4D rewrite Autopilot system when it gets off the motorway and onto city streets - switching automatically and seamlessly between these modes.

I may have misunderstood but that was how I interpreted it. So that means the performance of the car on the motorway (highway) sections of the drive is not behaving differently to what we have seen before and that is also why the display shifts to the familiar one we are used to when the car drives along the motorway.

Please correct me if anyone knows better.

 

[Dilly]

These recent videos showing examples of the new 4D FSD beta are definitely impressive. My understanding however was that at present the 'rewrite' 4D system was only being activated when the car entered the sort of city streets in which NoA currently does not operate. Therefore, what we are seeing on these videos on a longer journey like this is a hybrid between the car using the same NoA functionality most of us have on 2020.40.8 (but here with the US version of this that is not restricted by EU regulations) plus the new 4D rewrite Autopilot system when it gets off the motorway and onto city streets - switching automatically and seamlessly between these modes.

I may have misunderstood but that was how I interpreted it. So that means the performance of the car on the motorway (highway) sections of the drive is not behaving differently to what we have seen before and that is also why the display shifts to the familiar one we are used to when the car drives along the motorway.

Please correct me if anyone knows better.

No. I agree, your interpretation was mine too. I kind of assume that the guy had recently updated and was probably seeing general improvement. I’d say 40.8 is better than 36.12. Not perfect, but better.

 

[Strategium]

The table of results has been posted above - it’s comes LAST on assistance competence.

You are cherry picking the odd metric as if it’s the only one that matters in true fanboy style.

Have you tried any of the other actual system being tested?

First off, may I respectfully suggest that we all avoid using derogatory labels like 'fanboy' on this forum? I appreciate reading contrasting views and opinions - and especially from experienced and knowledgeable people such as yourself and others who regularly post here, regardless of whether I may, or may not, agree with you on a specific point.

In this specific case I don't have a strong opinion about how NCAP chose to evaluate different driver assist features on different makes and models and I don't really care much about their assessment either. I think most of these reports and assessments that are comparing new technologies inevitably have a degree of 'apples vs oranges' comparison built in to them. They definitely have a very arbitrary weighting based upon how you define 'categories' and what relative score you assign to separate categories. You can in this way get radically different outcomes from many types of evaluation just by altering the categories you arbitrarily define and the scoring system (also arbitrary) you choose.

I would say that at present we are still very much in an intermediate phase of evolving technology for autonomous driving and with many parallel systems developing quickly and changing rapidly, we don't know which one - or maybe many - will end up dominating the market in future. It could be that one system 'wins' - think VHS wiping out betamax for video tapes for example - and that does not even have to mean that the 'winner' is truly technically superior in all respects (I am not at all sure VHS was superior technically for example). Maybe no single solution will win, and the consumer then 'wins' as they get a wider choice of successful, great products that in reality all do a really good job, but with some differences - think Nikon vs Canon DSLRs for example. Which you choose will then depend upon personal preferences between a range of features and factors, so there is no need to conclude that one is objectively better than the other. Depends upon what matters most to you.

As of today, my personal preference is for the Tesla system and in my opinion it is the one most likely to rapidly evolve into a comprehensive autonomous system. I may be proved wrong about that. Also, my preference and opinion does NOT make me blind to current limitations with the Tesla system (I also hate phantom braking!) but I guarantee all of the other competing systems currently have limitations also and are not at all 'feature complete' in a way some posts might imply. Regardless, if anyone simply prefers driving an alternative solution from another manufacturer to Tesla and finds that works better for them - fine with me! I am also genuinely interested to hear why you prefer it and what features and experiences most affect your evaluation.