Enhanced Autopilot 2 may never happen before Full Self Driving is ready!

[lunitiks]

I fully expect that each feature enabled on HW2 will exceed the performance of it equivalent on MobileEye..

No doubt

 

[3Victoria]

No doubt

Agree. Such worry-warts here. GPS is accurate enough, along with viual reckoning and HD maps, go determine location. HD mapping was already underway: Tesla is mapping out every lane on Earth to guide self-driving cars

 

[Dstrohl]

This. It's a tail-wagging-the-dog argument to assume that "4 cameras" is set in stone, particularly if Tesla finds that using the additional cameras would add safety to AP2. In fact, Tesla would be remiss (if not borderline negligent) NOT to use the extra cameras in that case. I presume that all new cars will ship with all eight cameras physically installed and operational, if for no other purpose than to gather driving data for Tesla.

I find it much more likely that AP2 will primarly use the four cameras for most typical tasks, with occasional added input from the additional cameras when dealing with edge cases. And once full autonomy is achieved, it makes far more sense to use all eight cameras exactly the same whether or not "full autonomous" is enabled, and simply restrict the high-level feature set in software. There's no sense having to write (and test and validate and debug) the onramp-to-offramp subroutines twice: once for 4 cameras, once for 8 cameras.

I agree, see the end of my post! I dont know this to be true, but it makes more sense to me that they would use whatever cameras make the most sense per feature, and then sell the features in the packages... rather than saying you have 4 cameras here, and 8 there, *given that they have installed them all on all cars... if they sold the cars with only 4 physical cameras, it would be different.

HOWEVER, we really dont know that for sure, just because it makes more sense, and it seems to be a better approach (to us) doesn't mean that the engineers at Tesla agree.

 

[Dstrohl]

I hate to break it to you but hands on ready to take over will still exist and actually more than ever. You will need to be even more vigilant. Why? Because EAP is actually still L2 automation because there doesn't seem to be any safeguards, plus there are blindspots based on the 4 camera configurations. One in the left and right front and one at the back.

Think about those freeways where u come from the on ramp and immediately you have to get over to the left side, 3-4 lanes away so you can catch your exit or transition to another freeway. Now your car has to negotiate with wreckless speeding human drivers to get over three lanes.

Or even negotiating to enter the freeway from an on ramp so you don't get hit by the oncoming exit on the lane you are in.

Let's just say your hands will be alot more sweaty than simply driving straight. Again the car has to do with with numerous blindspots and relying on its ultrasonic. That's just unsafe.
I see the accident rate increasing from AP1 not decreasing unless tesla focus on making it safe

I think the point is that this version will be less likely to just stop driving at the drop of a hat and be much less confused in as many situations. AP1 had 1 camera and lower quality ultrasonic sensors... AP2 has (at least) 3 additional cameras and better sensors that reach out to 12' (so, about 2 lanes away). I don't think that Electricfan was trying to say that this next version would allow "hands off" driving (at least not right away) I could be wrong here, my mind reading skills don't work for him any more than they work for reading the Tesla Engineers minds (sadly!)

Those ultrasonic sensors should be able to help as well in terms of watching for cars in your immediate blind spot for merging left and right.

Note, in the end, we still don't know what they will be able to do, they might really suck! They might even totally fail to deliver. But I don't believe that what they are saying is "impossible", and I prefer to wait and see where they actually end up.

 

[Dstrohl]

Thanks for your detailed feedback.
I believe there are two teams for AP. One team working on EAP and training the NN using 4 cameras and another team working on FSD and training the NN using 8 cameras.

Why? (not disagreeing, I haven't seen anything that seems to suggest this, and it seems like it would make more sense for them to have one team for all "self driving" features.)

So it won't be easy to abruptly add a camera. And the only camera they can add without throwing off the NN for EAP is the forward main camera. But they could be lying and already have it on.

They did say they are all always on and sending data back to Tesla, so it's really a matter if deciding to use the data verses not using it. I would also not be surprised if the whole 4 cameras v. 8 cameras is more of a marketing concept, and the engineers are using whatever cameras they need.

However, I like your freeway exit transition scenario. It fits better with the rest of the industry. But that requires a level of FSD to be complete and be turned on during exits.

True, but generally a small one, and one that should already exist (based on the video's they have released. Even if it is not "perfect", or approved, it's really an emergency action anyway.

If tesla put a timer 10 seconds before exit as someone suggested we will simply get the same type of accidents as AP1. And isn't AP2 purpose to eliminate those quirks? It will introduce more dangerous quirks. I think I speak for everyone when I say. We don't want more parlor trick that we have to heavily supervise. We can something relatively safe.

If Tesla pulls a 10 seconds before exit and disengages. What could they pull under FSD?

The key here is that no one knows yet since they haven't spoken about how they are handling it. I agree that 10 seconds seems like a pretty short time if the driver is not paying attention, I woudl personally do at least 30-60 seconds... but they might disagree. and, if they do have something they can do after exit instead of just blindly disengaging, then 10 seconds might be fine.

I think this is the time to figure out how to make a 100% safe system in a particular scenario like an official highway.

Nothing can be 100% safe, %99.999 perhaps, but never 100%... there can always be things that no one could predict or fix... (a brick dropping off of an overpass? oncoming traffic suddenly hitting the divider and flipping into your lane? whatever... in those cases, you will be in an accident, and nothing that Tesla,or anyone else, can do will stop it. )

Having said that, I agree that this is the time to figure out how to make the safest system possible in a particular scenario... but a big part of that is to get the system on the road being used and discover what doesn't work. No amount of company testing will EVER be able to find all of the weird situations that humans get ourselves into.

If you want a perfect system, don't use AP for a while, those of us who use it immediately are beta testers and researchers. (not being snide here, it will get better over time, but I really believe that early users WILL find problems.)

Or there will be even more accidents and ppl being blamed for not being vigilant.

Yup, this is gonna happen... in fact, I predict that as soon as AP2 is released, EVERY accident involving a Tesla will be reported as "Tesla's new autopilot car involved in accident". even if the car is a 2012 model without any AP hardware, or if the car did not have it engaged. And people will have to be vigilant (at least at the beginning), which means that some will not be, and some of those WILL get into accidents, causing problems for the rest of us. That sucks, but nothing that Tesla could do would stop it. Even if the system was 100% accurate, most news outlets would report it incorrectly anyway,

Why not also have a L4 highway autonomy mode that acouple other company will be releasing 2017 onwards?

EAP is basically still L2 autonomy.

AP1 is definitely still L2 Autonomy, and at initial release EAP is likely to be as well. I think the goal for EAP is exactly what you said though, L4 highway autonomy in limited environments, I don't expect it to be there right away, but I suspect that in a few years we WILL see it, and probably (almost definitely) we will see it faster in Tesla's than in other cars, BECAUSE they are releasing it now and adding the features as they make them better.

Someone like Volvo will have the same problem, if they release a full L4 highway mode, how will they get it tested to the level that that anyone would be sure about how well it works in all situations without getting it in the hands of thousands of drivers in monitored (L2) driving.

 

[malcolm]

Part of the problem is that some owners' experience (not to mention, underoos) have been colored by "truck love" and "speed limit fascism".

Whilst I feel that it is far too early to call on the whole possible/impossible thing, I do understand a bad guided driving experience makes owners very nervous.

Both the abilities of AP and the learning curve it requires are not locked down and won't be for a good few months. For older versions, maybe longer/never.

Yes you paid for it, but it's not done baking and repeatedly sticking your head into the oven doesn't speed things up.

It just gives you a red face and weird hair.

Am I a fanboi for pointing out that nobody buys a car from Tesla? You buy an evolving platform, not a car.

With all the good and bad that implies.

 

[Bladerskb]

Enhanced Autopilot adds these new capabilities to the Tesla Autopilot driving experience. Your Tesla will match speed to traffic conditions, keep within a lane, automatically change lanes without requiring driver input, transition from one freeway to another, exit the freeway when your destination is near, self-park when near a parking spot and be summoned to and from your garage.

Tesla’s Enhanced Autopilot software is expected to complete validation and be rolled out to your car via an over-the-air update in December 2016, subject to regulatory approval.

So EAP update was released yesterday and its subpar and only contained
Traffic Aware Cruise Control feature, Forward Collision Warning, and (low speed) Autosteer.

Infact the update doesn't even match AP1 and there is no mention of summon or autopark.

Based on Tesla page we were supposed to get autosteer+, self-autolanechange, autoexit, smart summon and self-autopark.
We got none of those. All we got was cruise control and an auto steer that works below 35 mph

What am i saying? with Tesla development release timeline being 2-3 months as said from Elon during his AP2 hardware conference. It will take another 3-4 months for an update for EAP to bring it within parity with AP1. Then it will take another 2-3 for another update that will introduce some samples of the true EAP features listed above. Then another 2-3 for a update will bring the cars close to feature-complete of the promises of EAP.

So the estimate is atleast 9 months before Tesla is able to deliver anything close to what it promised. That's Sept 2017.

I can't say i didn't tell you so. Now you say 9 months isn't "NEVER".
Well i said "close feature complete" not feature complete. Based on elon, the end of 2017 is the time for their FSD demonstration.
Meaning that it will be in the final stages, although i don't believe they will be able to showcase anything independently verifiable in 2017

 

[ZeroDarkSilver]

I believe there are two teams for AP. One team working on EAP and training the NN using 4 cameras and another team working on FSD and training the NN using 8 cameras.

Hilarious. You are just straight up making stuff up to entertain yourself. Have fun.

 

[Bladerskb]

Hilarious. You are just straight up making stuff up to entertain yourself. Have fun.

Not at all, you should read nvidia paper on how they train their neural nets.

https://images.nvidia.com/content/t...2016/solutions/pdf/end-to-end-dl-using-px.pdf

 

[Canuck]

I can't say i didn't tell you so.

Leaving aside the fact that patting yourself on the back is gauche, I'd say you're doing it far too early given your prediction.

 

[ZeroDarkSilver]

Not at all, you should read nvidia paper on how they train their neural nets.

https://images.nvidia.com/content/t...2016/solutions/pdf/end-to-end-dl-using-px.pdf

Yeah that's not what I meant...

What I mean is that mashing together a bunch of marketing bullet points, blog posts, tweets, and white papers is not a suitable foundation on which to build a technical understanding of a system and a development process this complex. Much less predict the what, how, and when of its progress and completion.

Unless you're an automotive autonomy engineer actively working in the field right now, you're just engaging in pseudo-technical speculation for entertainment purposes and your guesses as to what is going on are almost certainly wrong.

Also, until you've put thousands of actual AutoPilot miles under your belt, your intuitions about what constitutes acceptable levels of utility are wrong too. Most people who've never tried AutoPilot, I've found, tend to think about it in very binary terms, either "I'm driving" or "I can go to sleep". In reality it's a continuum and, without exaggeration, something that's a completely new experience for human beings. You can't just think about AutoPilot based on what you've read and truly understand how you will really feel about it. Not until you've used it in the real world for extended periods of time will you really get it.

 

[schonelucht]

To me, it is comical to see all this negativity now about HW2 and EAP, when the doubters (myself included) were so wrong about AP1. No it will not be perfect, especially on day 1 (and day 1 is very soon), but I fully expect that each feature enabled on HW2 will exceed the performance of it equivalent on MobileEye..

Depends on what you call 'release' and 'soon'. Whatever was pushed to 1000 cars yesterday for sure is much worse than the MobileEye equivalent when comparing equivalent features.

 

[MXWing]

Depends on what you call 'release' and 'soon'. Whatever was pushed to 1000 cars yesterday for sure is much worse than the MobileEye equivalent when comparing equivalent features.

What came out yesterday could have killed me or others if I trusted the software blindly. I did not.

The worst offender is not the lack of features but the false positives from Collision Avoidance that made my MX break hard randomly from phantom cars in front of it.

I documented it here.

8.0 (2.50.185) unsafe to use.