Andrej Karpathy - AI for Full-Self Driving (2020)

[diplomat33]

Then why wouldn’t it stop for the example you gave?

This is my non HD mapped private road. View attachment 535057
I doubt anyone will ever HD Map it. We added theses stop signs to keep folks from thinking our street is a shortcut. A Tesla will absolutely stop here. Will Waymo?

How do you know that a Tesla will absolutely stop there? have you tried it personally to know that it stops?

I ask because Tesla also uses map data for stop sign detection and the manual warns that it might not reliably stop for unmapped signs. And you admit that you added the sign so it is probably not in Tesla's map data.

The fact is that Waymo uses camera vision and HD maps for stop sign detection. Waymo would most likely stop for that sign.

But frankly. I think we are both guessing. Have you been in a Waymo and can say for sure that it would not stop for that stop sign?

Seems to me you are playing hypothetical games because you want to argue that Tesla vision is better than Waymo's HD maps because you believe a Tesla would stop there but a waymo would not stop there. But neither of us know for sure.

 

[Bladerskb]

This is my non HD mapped private road. View attachment 535057
I doubt anyone will ever HD Map it. We added theses stop signs to keep folks from thinking our street is a shortcut. A Tesla will absolutely stop here. Will Waymo?

Waymo handles construction and traffic personnel holding slow and stop signs so i don't understand why waymo wouldn't handle a newly installed stop sign in a legitimate intersection. Nevertheless a car equipped with mobileye eyeq4 and crowd-sourced REM HD map with appropriate control algorithm will absolutely stop for that. Does that mean its better than an L4 waymo?

 

[scaesare]

The short answer is...

As long as you ignore the basic premise of my points, there's no sense repeating them.

 

[NHK X]

Seems to me you are playing hypothetical games because you want to argue that Tesla vision is better than Waymo's HD maps because you believe a Tesla would stop there but a waymo would not stop there. But neither of us know for sure.

Exactly this. But this is for both sides of the coin. We are all outside looking in. Got to be careful about presenting opinion as fact for all of us.

 

[Bladerskb]

Aren't you participating? Maybe if you don't want these things to go on, just let it go.

In fact, news alert to all who come on this forum. You don't have to keep bickering about your opinion until you get the other person to agree that your view is right. Civil discussion is great, and I love reading what people post. These topics are controversial. But even if the other person believes something you think is foolish, its okay to LET IT GO.

There's legitimate discussion and then there's disingenuous. Saying other SDC systems are just a glorified train because they use HD map is just that. Dishonest. Bad faith. illogical. Its like saying Tic Tac Toe is the same level of complexity as the game Go. You can't let statements like that stand. You have to call it out as it is. Maybe i have less patience because i have been responding to the same thing on this forum since basically 2016 and its still rinse and repeat. Nothing has changed. Its the same argument with lip gloss on it.

 

[NHK X]

There's legitimate discussion and then there's disingenuous. Saying other SDC systems are just a glorified train is just that. Dishonest.
Bad faith. illogical. You can't let statements like that stand. You have to call it out as it is. Maybe i have less patience because i have been responding to the same thing on this forum since basically 2016 and its still rinse and repeat. Nothing has changed. Its the same argument with lip gloss on it.

But you do realize the whole tesla fan boy argument doesn't do much for what you are trying to help people understand right? Up to that point I was understanding what you were presenting and what he was presenting. Found the discussion interesting. After it escalates to this it's just meaningless.

 

[diamond.g]

How do you know that a Tesla will absolutely stop there? have you tried it personally to know that it stops?

I ask because Tesla also uses map data for stop sign detection and the manual warns that it might not reliably stop for unmapped signs. And you admit that you added the sign so it is probably not in Tesla's map data.

The fact is that Waymo uses camera vision and HD maps for stop sign detection. Waymo would most likely stop for that sign.

But frankly. I think we are both guessing. Have you been in a Waymo and can say for sure that it would not stop for that stop sign?

Seems to me you are playing hypothetical games because you want to argue that Tesla vision is better than Waymo's HD maps because you believe a Tesla would stop there but a waymo would not stop there. But neither of us know for sure.

I was granted access to a vehicle in the program and saw it in person. If you cover the sign the vehicle proceeds as if the sign was never there.

Waymo handles construction and traffic personnel holding slow and stop signs so i don't understand why waymo wouldn't handle a newly installed stop sign in a legitimate intersection. Nevertheless a car equipped with mobileye eyeq4 and crowd-sourced REM HD map with appropriate control algorithm will absolutely stop for that. Does that mean its better than an L4 waymo?

Does the Mobile Eye solution read the words on the sign. I am glad it would stop in my intersection, it just seems like you guys are saying you need to have it mapped to be legit. And on dirt roads I doubt that would ever happen.

 

[diplomat33]

As long as you ignore the basic premise of my points, there's no sense repeating them.

What I am missing? You argued that vision can detect stop lights so detection is not the problem rather it is:

So it's the ability to proceed at that point that is the issue here.

And you asked:

So tell me again how vision is insufficient?

So you believe vision can detect stop lights and detect color so you want to know how vision is not good enough to handle intersections? Is that right?

I answered you: Vision alone can have many false positives. And HD maps/LIDAR can also help with tracing an accurate path through the intersection more reliably than vision. So when it comes to actually proceeding through the intersection which is what you specifically asked about, HD maps/LIDAR can help to make it more reliable than with vision alone.

 

[diplomat33]

I was granted access to a vehicle in the program and saw it in person. If you cover the sign the vehicle proceeds as if the sign was never there.

Uh, if you cover the sign, a Tesla would blow through it too.

Plus, you did not ask about the case of covering a sign. You asked about a sign that is easy to see with vision but that is unmapped. That's a different case.

 

[DanCar]

... Saying other SDC systems are just a glorified train because they use HD map is just that. Dishonest. Bad faith. illogical. Its like saying Tic Tac Toe is the same level of complexity as the game Go. You can't let statements like that stand. You have to call it out as it is. ...

I found the comment interesting. People have different perspectives and different styles of communication and that doesn't make your's right and theirs wrong. This isn't a scientific forum or engineering forum, it a discussion forum for all public members to voice their views. You shouldn't be trying to silence them. If you want a more technical and scientific discussion create your own thread and state engineers or scientists only.

 

[diamond.g]

Uh, if you cover the sign, a Tesla would blow through it too.

Plus, you did not ask about the case of covering a sign. You asked about a sign that is easy to see with vision but that is unmapped. That's a different case.

I mentioned the sign covering behavior because there was some thought that the sign had to be mapped for Tesla to stop, they don’t. That is why I asked @Bladerskb if Mobileeye is reading the words on the sign to know if is an actual stop sign or not. As it stands I don’t believe Tesla cares about the words in the sign.

 

[diplomat33]

Here is an interesting case: let's say that you have a regular intersection with stop signs and the stop signs are mapped. But kids remove one of the stop signs so now there is a missing stop sign. The autonomous car should still stop since it is a legal stop at an intersection but there is no visible stop sign. An autonomous car relying purely on vision would not know that it needs to stop since there is no visible stop sign. But an autonomous car with HD maps would stop since the HD maps would tell it that it needs to stop.

 

[NHK X]

I found the comment interesting. People have different perspectives and different styles of communication and that doesn't make your's right and theirs wrong. This isn't a scientific forum or engineering forum, it a discussion forum for all public members to voice their views. You shouldn't be trying to silence them. If you want a more technical and scientific discussion create your own thread and state engineers or scientists only.

exactly this. I’m still trying to understand what makes people so defensive about this discussion. And what makes people start to think they are the end all truth to the discussion.

we are all outsiders looking in.

 

[diamond.g]

Here is an interesting case: let's say that you have a regular intersection with stop signs and the stop signs are mapped. But kids remove one of the stop signs so now there is a missing stop sign. The autonomous car should still stop since it is a legal stop at an intersection but there is no visible stop sign. An autonomous car relying purely on vision would not know that it needs to stop since there is no visible stop sign. But an autonomous car with HD maps would stop since the HD maps would tell it that it needs to stop.

Yes, what you describe should be the preferred reaction, but then it would never work right on private roads or really rural areas that are not state maintained.

Is it safe to say that Waymo should work great anywhere they have Google Street View?

 

[Bitdepth]

This is my non HD mapped private road. View attachment 535057
I doubt anyone will ever HD Map it. We added theses stop signs to keep folks from thinking our street is a shortcut. A Tesla will absolutely stop here. Will Waymo?

I've been lurking and reading this forum for a long time and some of the discussions going on in here is rather perplexing. Of course waymo would stop for a stop sign even though it is not on the map database. Waymo stops at a police controlled intersection and reads the hand signs.

It detects stop signs held by traffic control attendants.

It can detect a school bus with the stop sign out and stop for it.

The Map is just another point of reference which the neural network uses to make its decision. They use 29 cameras, 3 radars, 4 lidars in total to sense the surrounding, and also a microphone which they use to detect sirens and emergency sounds. They are not there for show.

 

[NHK X]

Here is an interesting case: let's say that you have a regular intersection with stop signs and the stop signs are mapped. But kids remove one of the stop signs so now there is a missing stop sign. The autonomous car should still stop since it is a legal stop at an intersection but there is no visible stop sign. An autonomous car relying purely on vision would not know that it needs to stop since there is no visible stop sign. But an autonomous car with HD maps would stop since the HD maps would tell it that it needs to stop.

interesting, wouldn’t there be other markings on road potentially to cue stop? If no other cues and the post is removed too, would a human stop if they were traveling in that area for the first time?

I think tough part about these hypotheticals regarding this argument is that it’s all circular.

 

[diamond.g]

interesting, wouldn’t there be other markings on road potentially to cue stop? If no other cues and the post is removed too, would a human stop if they were traveling in that area for the first time?

I think tough part about these hypotheticals regarding this argument is that it’s all circular.

In our case, coming from the opposite direction to the picture I took is basically a blind intersection. Where you should probably stop anyways.

 

[Bitdepth]

I found the comment interesting. People have different perspectives and different styles of communication and that doesn't make your's right and theirs wrong. This isn't a scientific forum or engineering forum, it a discussion forum for all public members to voice their views. You shouldn't be trying to silence them. If you want a more technical and scientific discussion create your own thread and state engineers or scientists only.

If you are having a discussion based on a scientific topic, your language should reflect as such. You can't say Autopilot is basically level 5 autonomous, you certainly cannot you say a Level 4 autonomous vehicle which uses every sensor modality afforded is just a glorified tram.

You are no longer discussing in good faith and calling out such cavalier and disingenuous language is not the same as silencing someone. I read this forum and I'm sure there are lots of people who read but do not post. I have seen many people use disrespectful language towards one another (a few posts up for example) because we are all human and we get heat sometimes.

 

[NHK X]

Devils advocate here. People take Tesla’s delay to achieve FSD as a sign that it’s nice advanced drivers assist, but it’s all smoke and mirrors beyond it. What makes Karpathy stay if it all truly was a non viable approach? Money can be made elsewhere.

 

[diplomat33]

Devils advocate here. People take Tesla’s delay to achieve FSD as a sign that it’s nice advanced drivers assist, but it’s all smoke and mirrors beyond it. What makes Karpathy stay if it all truly was a non viable approach? Money can be made elsewhere.

A few possibilities come to mind:

Maybe Karpathy believes the approach will lead to L5 autonomous driving.

Maybe Karpathy just likes working at Tesla.

Maybe Karpathy genuinely likes what Tesla is doing with machine learning and deep neural networks and it's a good match for his passion and skills.

Maybe Karpathy is just interested in getting the most out of machine learning and does not specifically care about whether it will achieve L5 any time soon.