Tesla autopilot HW3

[ZeApelido]

The animosity in this thread is amazing. I mean, I think I get why, it's not unreasonable. I just think people totally don't understand intentions.

I don't necessarily believe that Level 5 will be achieved with cameras only HW 3.0, but I don't get angry at Elon.

Why?

Because it has been shown time after time that Elon Musk really believes this things can be achieved. You know, starting a rocket company from scratch (being laughed at by Russians), making electric vehicles relatively affordable and desirable (laughed at by many). Did he really truly know these could be achieved when he proclaimed as such, or did he just have the (over) confidence that with enough drive it could be done? Probably both.

Elon likely truly believes they can get to FSD. He probably really believes in the timelines. And he's been wrong, and he'll be wrong again. I bet he stretches the truth a little, but not nearly as much as some of you imply.

Get over yourself. Falling asleep behind the wheel by end of 2020 is probably not going to happen. So what. Tesla does really want that to happen, they aren't trying to hoodwink you here.

In the meantime, they will leave you with some of the best L2/L3/L4 systems around. Seriously I don't know, who has the best in production cars right now? Not a bad consolation prize.

 

[diplomat33]

As I see it, the reason the experts go with the full hardware (cameras, radar and LIDAR) is because it is the most obvious and logical way to solve the problem of FSD. In other words, it's the hardware package that is most likely to make solving FSD the easiest and most reliable. The hardware covers all the bases. Tesla's camera vision approach is a lot harder. It will have a lot of weaknesses. but Tesla is hoping to compensate for these weaknesses with clever software. So technically, Tesla's current hardware might be good enough but it makes the problem a lot harder. Musk is probably so adamant that L4 is doable on the current hardware because he believes that the team can compensate for the weaknesses in the hardware with the right software approach. A good example of this might be the stationary vehicle problem. LIDAR can easily detect a stationary vehicle since it just measures the reflections of the beams and gets a 3D picture of the surrounding and knows the exact location of an object. You then easily tell your car to steer to avoid that location. The problem can be solved with camera vision only so the camera hardware would be good enough but it is a lot harder because you need more sophisticated vision algorithms to analyze the image, recognize what it is seeing and then do a lot of geometry calculations to determine distance. in fact, it's what humans do every day. So it is obviously doable. But it is certainly a lot easier to just bounce a light beam off the object and measure the time and get a precise location that way.

The advantage of AP3 is that it will dramatically increase the processing power, making it possible to process more data and thus make it possible for the software solution to work better. And we know that Tesla has been working on bigger NN. So my guess is Tesla is hoping AP3 will allow them to "brute force" their way to a software solution that overcomes the weaknesses of their hardware. The experts think that is not possible, Musk thinks it is.

 

[ZeApelido]

Now in terms of the details of how well could HW 3.0 plus better algorithms do, I find it hilarious anyone has any confidence in their predictions.

I work in machine learning, and do deep learning, and all I can say is deep learning progress was very fast over several years, with new methods discovered via trial and error (not theory). Most image detection nnets are at one frame at a time. I mean Jimmy_d was freaking out when the code indicated a new network that was going to input 2 temporal images at once! Since obviously video / aka multiple frames could yield much more perception benefits (but take a ton more to train), it just shows that the field is still growing with respect to video learning.

How anyone could say a team of some of the best deep learning talent won't achieve significant improvements, or confidently that they will achieve them, is funny. How the heck do you know if you aren't down in the level of details they are? No one has the data set they have. I literally would not be surprised by any outcome.

 

[J1mbo]

Conversely, nobody has ever succeeded at making an AV, at any cost.

Not true. Plenty of AVs out there. What do you think Waymo do?

 

[electronblue]

Robo-taxis only need to be Level 4...

Elon says HW3 is FSD computer + AP2.5 suite.

So you believe that to be Level 4?

 

[electronblue]

Now in terms of the details of how well could HW 3.0 plus better algorithms do, I find it hilarious anyone has any confidence in their predictions.

I work in machine learning, and do deep learning, and all I can say is deep learning progress was very fast over several years, with new methods discovered via trial and error (not theory). Most image detection nnets are at one frame at a time. I mean Jimmy_d was freaking out when the code indicated a new network that was going to input 2 temporal images at once! Since obviously video / aka multiple frames could yield much more perception benefits (but take a ton more to train), it just shows that the field is still growing with respect to video learning.

How anyone could say a team of some of the best deep learning talent won't achieve significant improvements, or confidently that they will achieve them, is funny. How the heck do you know if you aren't down in the level of details they are? No one has the data set they have. I literally would not be surprised by any outcome.

Woohoo! Level 5 capable AP2 hardware!

 

[J1mbo]

In case anyone was wondering...

("These tasks" = NoA)

 

[electronblue]

Tesla does really want that to happen, they aren't trying to hoodwink you here.

If Tesla is saying something they just want to be true, but is not true, then they are hoodwinking you.

 

[electronblue]

As I see it, the reason the experts go with the full hardware (cameras, radar and LIDAR) is because it is the most obvious and logical way to solve the problem of FSD. In other words, it's the hardware package that is most likely to make solving FSD the easiest and most reliable. The hardware covers all the bases. Tesla's camera vision approach is a lot harder. It will have a lot of weaknesses. but Tesla is hoping to compensate for these weaknesses with clever software. So technically, Tesla's current hardware might be good enough but it makes the problem a lot harder. Musk is probably so adamant that L4 is doable on the current hardware because he believes that the team can compensate for the weaknesses in the hardware with the right software approach. A good example of this might be the stationary vehicle problem. LIDAR can easily detect a stationary vehicle since it just measures the reflections of the beams and gets a 3D picture of the surrounding and knows the exact location of an object. You then easily tell your car to steer to avoid that location. The problem can be solved with camera vision only so the camera hardware would be good enough but it is a lot harder because you need more sophisticated vision algorithms to analyze the image, recognize what it is seeing and then do a lot of geometry calculations to determine distance. in fact, it's what humans do every day. So it is obviously doable. But it is certainly a lot easier to just bounce a light beam off the object and measure the time and get a precise location that way.

The advantage of AP3 is that it will dramatically increase the processing power, making it possible to process more data and thus make it possible for the software solution to work better. And we know that Tesla has been working on bigger NN. So my guess is Tesla is hoping AP3 will allow them to "brute force" their way to a software solution that overcomes the weaknesses of their hardware. The experts think that is not possible, Musk thinks it is.

You avoided my question.

Do you think current Teslas will do Level 4? (FSD computer + AP2.5 sensor suite.) After all Musk says they will do autonomous ride sharing that must be Level 4 or above by definition.

 

[mongo]

Elon says HW3 is FSD computer + AP2.5 suite.

So you believe that to be Level 4?

Yeah, I think they can level 4 with that.

 

[diplomat33]

You avoided my question.

Do you think current Teslas will do Level 4? (FSD computer + AP2.5 sensor suite.) After all Musk says they will do autonomous ride sharing that must be Level 4 or above by definition.

No, I tried to answer your question:

"To answer your question, I think AP3 will probably do L3 autonomy. I think L4 is a huge stretch on the current hardware. Although the increased processing power of AP3 should help somewhat."

In other words, my answer is "I am leaning towards no but maybe it will".

 

[electronblue]

OK, thanks @mongo and @diplomat33.

 

[diplomat33]

In case anyone was wondering...

View attachment 395077

("These tasks" = NoA)

This might explain why Musk is optimistic about FSD ride sharing. If AP3 is only at 10% compute load with full fail-over redundancy doing NOA then Musk must think that they have enough processing power to handle all the necessary operations for FSD. And if the new NN are really that good then I could see Musk being optimistic.

Come on April 22. We need to know!!!

 

[electronblue]

Sometimes I feel like @diplomat33 is the one gaslighting us all.

Bless the different perspectives I guess.

 

[ZeApelido]

If Tesla is saying something they just want to be true, but is not true, then they are hoodwinking you.

Well I believe they truly think it will happen.

 

[rnortman]

To answer your question, I think AP3 will probably do L3 autonomy. I think L4 is a huge stretch on the current hardware. Although the increased processing power of AP3 should help somewhat. Ideally, I would love to see Tesla's get a rear radar, 360 degree LIDAR and a better computer than AP3. That would complement the existing cameras nicely. If Tesla's got that added to the existing hardware, I would be a lot more confident in L4.

I feel like suddenly I'm arguing with a different person. Like, a person who mostly already agrees with me. I've always said that Tesla has the makings of a damned fine L2+ ADAS system (bemoaning lack of rear radar to make it even better) which might maybe make it to L3 on the highway in limited/ideal conditions and particular approved roads. You might also count parking lots at excruciatingly low speed as L3. But the posts you have made here suggest that you really believe in the L4/L5 vision... I'm finding this difficult to reconcile with the above statement.

 

[rnortman]

I don't necessarily believe that Level 5 will be achieved with cameras only HW 3.0, but I don't get angry at Elon.

Why?

Because it has been shown time after time that Elon Musk really believes this things can be achieved.

I've said many times that I think Elon believes that he is doing the right thing. Specifically he thinks that he is saving the human race by slowing global warming and giving us access to other planets to move to if/when we destroy this one. I think the fact that he puts himself at the center of all this is a narcissistic fantasy, but nevermind that -- he believes himself to be pursuing a noble goal. I have no doubt about that whatsoever.

The problem here is that he very much believes that the ends justify the means. He must - must - have consciously known that the 2016 demo video was a lie. Maybe he thought the lie wouldn't matter because within 6 months they'd have FSD figured out, but that doesn't change the fact that it could only possibly have been a knowing lie. By this I mean the accompanying text that the driver was only there "for legal reasons" and that the software was only pending "validation and regulatory approval". We all know now that the software did not exist when that video was made and that driver was absolutely critical for safety, not just legal reasons.

For that -- and the fact that he clearly manipulates Tesla customers in order to raise more money to pursue his noble goals, and the fact that people have died as a result of clearly misguided beliefs about Autopilot -- I believe I am justified in criticizing him and anybody else at Tesla who was involved in the making and release of that demo video.

 

[PaulJohn]

A lot of pessimism around here... I believe Tesla will eventually excel in all areas of driving, providing incomparably better safety than that of a human driver.

 

[diplomat33]

Sometimes I feel like @diplomat33 is the one gaslighting us all.

I feel like suddenly I'm arguing with a different person. Like, a person who mostly already agrees with me. I've always said that Tesla has the makings of a damned fine L2+ ADAS system (bemoaning lack of rear radar to make it even better) which might maybe make it to L3 on the highway in limited/ideal conditions and particular approved roads. You might also count parking lots at excruciatingly low speed as L3. But the posts you have made here suggest that you really believe in the L4/L5 vision... I'm finding this difficult to reconcile with the above statement.

I can assure you both that I am not gaslighting anyone. But I have evolved in my views. In the past, I used to be super gun ho about FSD, even starting threads that were basically "OMG!! Musk says we will have FSD!!!!". Now, I do remain very optimistic about FSD, that has not changed, and occasionally the "Tesla fanboy" inside me comes out, but at the same time, I also understand a bit better the difficulties of FSD and I know that L4 autonomy ride-sharing with the current hardware is probably not realistic. So I am trying to be more realistic in my positions while at the same still being true to myself and remaining optimistic about FSD.

 

[rnortman]

A lot of pessimism around here... I believe Tesla will eventually excel in all areas of driving,

Yes I have no doubt... but will the cars that they have sold and accepted payment for (AP2/2.5) ever excel in autonomous driving?