Autonomous Car Progress

[powertoold]

This is wrong. We know quite a lot. Waymo gives regular updates on what they are doing. And unlike Tesla, Waymo has actually put their latest research out there for anyone to read. You can search by topics and read what Waymo has done in perception, prediction, planning and general ML:

I did watch some of those videos already, and they're general videos about using ML in AVs, just as the title suggests. The examples are very general and don't show any detail about what Waymo currently does.

Again, there's little detail about how Waymo approaches mapping, driving, etc. You have no clue about what it takes to get a city / area up and running.

 

[Doggydogworld]

This is wrong. We know quite a lot. Waymo gives regular updates on what they are doing. And unlike Tesla, Waymo has actually put their latest research out there for anyone to read. You can search by topics and read what Waymo has done in perception, prediction, planning and general ML:

The talks and papers I've read give interesting insights, but not nearly enough info to make the claims you make.

Consider lidar. Do they feed it into NNs for perception and such? Absolutely. Do they then throw away those millions of valuable, precise measurements? Well, maybe, but since that'd be criminally insane I kinda doubt it. I'm confident they also use the lidar data and (gasp) hand code to assign precise measurements to the "23-ish feet" guesses coming out of the perception NNs. And they almost certainly use hand code and lidar points to independently build a 3D mesh. Potential uses for this mesh include contradicting a NN that predicts the now-occluded front end of an articulated bus has continued on its merry way with the more concrete "2.3m wide by 3.0m high stationary flat surface directly ahead at a distance of 8.23 meters..... 7.94 meters..... 7.61 meters.... STOP STOP STOP".

This is completely false. Waymo does not do this. Waymo has even explained that their HD mapping is automated, not hand coded.

That doesn't even make sense. Almost all automation is hand coded.

This entire thread is odd, with a strong dose of NN mysticism. They all use a mix of NNs and hand code. Tesla is not special, they're not "ten years ahead" and they don't have magical NNs that you just dump "billions of miles" into. And Waymo isn't 98% hand code with a couple of primitive NNs tacked on, they don't blindly drive the "brittle" HD map like a "train on tracks". But they're also not 100% NN. Nor should they be.

 

[diplomat33]

The talks and papers I've read give interesting insights, but not nearly enough info to make the claims you make.

Consider lidar. Do they feed it into NNs for perception and such? Absolutely. Do they then throw away those millions of valuable, precise measurements? Well, maybe, but since that'd be criminally insane I kinda doubt it. I'm confident they also use the lidar data and (gasp) hand code to assign precise measurements to the "23-ish feet" guesses coming out of the perception NNs. And they almost certainly use hand code and lidar points to independently build a 3D mesh. Potential uses for this mesh include contradicting a NN that predicts the now-occluded front end of an articulated bus has continued on its merry way with the more concrete "2.3m wide by 3.0m high stationary flat surface directly ahead at a distance of 8.23 meters..... 7.94 meters..... 7.61 meters.... STOP STOP STOP".

Certainly, I am not suggesting that we know every single thing about Waymo's design. I am simply pushing back against Powertoold's claims that we know nothing or very little about their approach. I don't think that is a true statement. We know a lot about Waymo's approach even if we don't know everything.

That doesn't even make sense. Almost all automation is hand coded.

My point is that maps are not done "by hand". There aren't Waymo engineers manually creating HD maps like some try to claim.

This entire thread is odd, with a strong dose of NN mysticism. They all use a mix of NNs and hand code. Tesla is not special, they're not "ten years ahead" and they don't have magical NNs that you just dump "billions of miles" into. And Waymo isn't 98% hand code with a couple of primitive NNs tacked on, they don't blindly drive the "brittle" HD map like a "train on tracks". But they're also not 100% NN. Nor should they be.

Thank you for saying this. Again, I am not claiming that Waymo does end-to-end. We know they do not. I am simply stating the fact that Waymo heavily uses NN in all 3 parts of their stack (perception, prediction and planning). I am really just pushing back against the silly "Waymo does not use NN and just hand codes driving on virtual tracks" nonsense that Powertoold keeps pushing.

 

[powertoold]

The HD map defines the rails the car can take. The HD map takes a long time to refine and finetune. That's why you see multiple Waymos passing through the same areas time and time again. They need to refine and then test and refine and test again, to make sure all the traffic lights / speeds / parking times / obstructions / etc. are what they expected throughout the whole day.

It's a lot of human-involved finetuning. That's what I mean by hand-coded and feature engineering: all the directions of traffic, turn lanes, connectivities, one-way roads, cross traffic, bumps / dips, crosswalks, exact locations of traffic lights and associations to lanes, etc. etc.

I'll just give you one example in SF. During certain times of the day, cars are not allowed to park on some lanes. These lanes become turn lanes during these times. All this information needs to be on the HD map.

You're silly if you think Waymo can just generate all this information without human involvement. If they could do that, it wouldn't take them months or years to roll out a new area.

 

[diplomat33]

I did watch some of those videos already, and they're general videos about using ML in AVs, just as the title suggests. The examples are very general and don't show any detail about what Waymo currently does.

I know I will never convince you but I disagree. For example, Anguelov discusses Waymo's transformers architecture to do 3D object detection from point clouds and unsupervised learning for auto labelling moving objects. That is not "general videos about ML". That is specific and it gives us details about what Waymo is currently doing. And in the "long tail" video, Anguelov shares a specific example of how their unsupervised learning auto labelled a troller. That is a specific example.

It is remarkable how Tesla can talk about occupancy networks, you gush that it is amazing the level of detail but when Waymo talks about their occupancy networks for behavior prediction, somehow they are just giving "general videos" and not giving us any details.

Again, there's little detail about how Waymo approaches mapping, driving, etc. You have no clue about what it takes to get a city / area up and running.

If you actually paid attention to the Waymo videos, you would see there is a lot of detail on how Waymo approaches mapping, driving etc. You just can't admit it. That does not mean that we know everything about Waymo but we know a lot, if you cared to pay attention.

But if you really believe that we have no clue then you should not push your baseless "opinions" on how Waymo does things. You should just keep silent.

Case in point: you offer another "opinion" that is wrong. You don't know what you are talking about. You are just making guesses.

The HD map defines the rails the car can take. The HD map takes a long time to refine and finetune. That's why you see multiple Waymos passing through the same areas time and time again. They need to refine and then test and refine and test again, to make sure all the traffic lights / speeds / parking times / obstructions / etc. are what they expected throughout the whole day.

It's a lot of human-involved finetuning. That's what I mean by hand-coded.

The HD maps do not define the rails the car can take. And they do need constant fine tuning. The reason so many Waymos pass through the same area is for testing, not because they need to constantly fine tune the maps. Also, Waymo has said that the maps are used as priors and they don't need the maps to be perfectly accurate in order to drive. So no, they don't need to constantly remap every route all the time.

This is from Waymo blog on mapping:

"We’ve automated most of that process to ensure it’s efficient and scalable. Every time our cars detect changes on the road, they automatically upload the data, which gets shared with the rest of the fleet after, in some cases, being additionally checked by our mapping team."

Note that most of the process is automated which implies not a lot of hand coding. Also the cars automatically upload changes to the map. And they say "in some cases". So the cases where they do need humans to check or do fine tuning is only some cases, not a lot.

Please, stop offering your "opinions" that are factually wrong. You don't know what you are talking about. You are making incorrect guesses.

 

[powertoold]

Note that most of the process is automated which implies not a lot of hand coding. Also the cars automatically upload changes to the map. And they say "in some cases". So the cases where they do need humans to check or do fine tuning is only some cases, not a lot.

Please, stop offering your "opinions" that are factually wrong. You don't know what you are talking about. You are making incorrect guesses.

You need to read closer when Waymo creates marketing material.

There's no doubt Waymo is getting better at creating the HD maps, but it's very resource intensive to create and test and finetune the maps. The biggest problem is that it's the wrong approach to self-driving, in the long term.

 

[diplomat33]

You need to read closer when Waymo creates marketing material.

I do more than read waymo marketing material. I read their research papers that detail how they do their ML.

And you should not offer speculation based just on seeing a lot of Waymo cars drive the same route. At least, I am able to provide citations from actual Waymo material.

 

[powertoold]

And you should not offer speculation based just on seeing a lot of Waymo cars drive the same route. At least, I am able to provide citations from actual Waymo material.

Nothing you posted contradicts what I said though.

 

[stopcrazypp]

I do more than read waymo marketing material. I read their research papers that detail how they do their ML.

And you should not offer speculation based just on seeing a lot of Waymo cars drive the same route. At least, I am able to provide citations from actual Waymo material.

Links to some of these so people can make their own judgement? There's a whole lot of back and forth but no sources.

For example, the time based lanes example brought up (they are quite common in SF), I would be quite interested how Waymo can do it and similar scenarios with no human involved coding. It requires reading the street sign and interpreting it like a human does, as well as incorporating in time of day. A similar scenario was brought up in terms of school zones. That would seem to be a scenario that hand coding would work best.

 

[johnm]

This is completely false. Waymo does not do this. Waymo has even explained that their HD mapping is automated, not hand coded. You just keep pushing these lies about Waymo. I wish you would stop.

Waymo seems less confident of this than you are:

“The mapping process is a critical part of our operations,” Nick Smith, a Waymo spokesperson, told TechCrunch. “Here’s how it works: First, we drive our Waymo vehicles on public roads and collect information with our sensor suite. Next, we take that information, clean it up and automatically or manually annotate it with features such as crosswalks, road edges, curb heights, boundary paint, intersections, etc. Then we put our newly created map through quality control testing. This process is the same no matter where we go, and is also the same process we follow when updating our maps.”

 

[diplomat33]

Waymo seems less confident of this than you are:

Confidence has nothing to do with anything. That does not contradict what I said. That is an accurate description of the mapping process. Note that it says automatic or manual. So the annotation can be done automatically. The process is very automated but not completely per the blog I sourced. So there is some manual annotations for certain features.

 

[diplomat33]

Links to some of these so people can make their own judgement? There's a whole lot of back and forth but no sources.

I am happy to provide sources. I usually do but not always.

For example, the time based lanes example brought up (they are quite common in SF), I would be quite interested how Waymo can do it and similar scenarios with no human involved coding. It requires reading the street sign and interpreting it like a human does, as well as incorporating in time of day. A similar scenario was brought up in terms of school zones. That would seem to be a scenario that hand coding would work best.

Good question. I don't know the answer. Although, I see no reason ML could not handle that case with the right training and data. I know years ago, Waymo was already able to read address signs so I imagine it would be doable.

 

[stopcrazypp]

I am happy to provide sources. I usually do but not always.



Good question. I don't know the answer. Although, I see no reason ML could not handle that case with the right training and data. I know years ago, Waymo was already able to read address signs so I imagine it would be doable.

I'm not talking about reading the text. That part should be relatively trivial. I'm talking about understanding its connotations/context as it relates to driving. That part machines have a hard time doing and where hand coding would be much simpler for signs that involve more variables. It's not so simple to create an NN that can generalize such understanding. However others linked that Waymo still manually annotates some parts of their maps, so this part likely falls under manual annotation.

 

[diplomat33]

I know one short clip does not prove reliability but if you are curious how Mobileye's vision-only FSD handles an unprotected left, here is one example:

 

[EVNow]

I know one short clip does not prove reliability but if you are curious how Mobileye's vision-only FSD handles an unprotected left, here is one example:

This is just a video posted by the company PR. They can run this 100 times and 1 time it works they can post the video (or may be it is successful all 100 times - we just don't know).

That is why only official statistics they submit to regulators is what we can go by. Everything else can be smoke and mirrors.

 

[diplomat33]

This is just a video posted by the company PR. They can run this 100 times and 1 time it works they can post the video (or may be it is successful all 100 times - we just don't know).

That is why only official statistics they submit to regulators is what we can go by. Everything else can be smoke and mirrors.

I never claimed otherwise. That is why I prefaced the video by saying "I know one short clip does not prove reliability". I am acknowledging exactly what you just said, that one clip does not prove that it would work 100x times, it only proves that it worked this one time. Still, I thought people might be interested in one video example.

 

[drtimhill]

I know one short clip does not prove reliability but if you are curious how Mobileye's vision-only FSD handles an unprotected left, here is one example:

You see the problem I have here is you are showing a promotional video. How typical do you think that video is? I think it is naive to imagine a manufacturer would not cherry pick the best examples of their system (or even outright fake them).

 

[diplomat33]

You see the problem I have here is you are showing a promotional video. How typical do you think that video is? I think it is naive to imagine a manufacturer would not cherry pick the best examples of their system (or even outright fake them).

Yes, I know I am sharing a promo video. I know it is cherry picked. I am not claiming that it proves that it works every time. That is the whole reason why I put a disclaimer before the video!

Now, if your point is that we should not share cherry picked promo videos period. Well, like I said, I thought a cherry picked promo video would still be interesting to share since I put a disclaimer first.

 

[EVNow]

Yes, I know I am sharing a promo video. I know it is cherry picked. I am not claiming that it proves that it works every time. That is the whole reason why I put a disclaimer before the video!

Now, if your point is that we should not share cherry picked promo videos period. Well, like I said, I thought a cherry picked promo video would still be interesting to share since I put a disclaimer first.

I'm fine with anyone sharing promo videos - except when used as any kind of evidence (not saying you did).

Also you should look at Tesla the same way. Be sceptical of Waymo etc as much as you are of YT "shills".

I call YT testers shills because they are. For example, Whole Mars is a Tesla shareholder who puts out tweets and videos just to promote Tesla and pump TSLA stock. So yes, they are shills. Brad is a bonified AV expert who has worked in the AV field for years.

 

[diplomat33]

I'm fine with anyone sharing promo videos - except when used as any kind of evidence (not saying you did).

Ok. And just to be clear, I was not sharing that Mobileye promo video as proof of anything, merely as an interesting example.

Also you should look at Tesla the same way. Be sceptical of Waymo etc as much as you are of YT "shills".

Certainly, some skepticism is healthy but beware of too much skepticism. Automatically being skeptical of everything out of hand is not good. We should look at the source. If it is a reliable source then we can trust it. I think it is important to look at the type of material. A promo video is different from a consumer or rider video. An unedited video is more trustworthy than an edited video. A FSD Beta video is different from a Waymo driverless ride video since one has driver supervision and the other is fully driverless. Unedited driverless ride videos are more trustworthy than unedited videos with driver supervision since driverless is implied to be more reliable since the company is assuming liability when it is driverless. A peer reviewed research paper is authoritative. It is also important to look at the authority of the speaker. For example, a Tesla youtuber does not have the same authority as say Andrej Karpathy or Drago Anguelov who work with ML. When Karpathy or Anguelov speak on ML, they are speaking authoritatively on the subject so we can trust what they are saying more than a random youtuber.