Autonomous Car Progress

[Bladerskb]

This was discussed at several different times in the past.

I think this is the first confirmation from Elon that the mic will be used to listen to surroundings.
View attachment 683681

Elon still playing catch up.

Mobileye's Supervision on ZEEKR 001

"What's interesting is that Jikrypton 001 has "skeletal recognition" technology, which recognizes posture and predicts behavior by locating the characteristics of human bones. For example, at an intersection where there are no traffic lights, when a pedestrian suddenly raises his hand, bone recognition can quickly determine whether the person is waving his hand to stop you, or raising his mobile phone to make a call."

3 more months and its 'game over'

 

[Knightshade]

This is a myth. When Waymo wants to test in a city they spend a couple weeks driving around to build maps and learn local customs and quirks.

Months, not weeks, was the time I saw cited regarding their Los Angeles mapping project, seems to split the measures you and I were suggesting.

I could maybe buy weeks for a tiny bit of a Phoenix suburb though

Google mapped the whole world pretty cheaply, and updates most areas every year or two (and yes, those cars have had lidar for many years).

Given google is where Waymo came from, presumably that data wasn't precise enough for this purpose- otherwise why is Waymo doing work already done?

I've not come across anyone in the autonomous driving field who thinks maps are a bottleneck.

NVIDIA to buy autonomous vehicle mapping company DeepMap

NVIDIA has acquired DeepMap, a company that makes high-definition mapping technology for self-driving cars.

www.engadget.com

As recently as a month ago Nvidia was still acquiring companies to address the problem of being able to scale mapping for autonomous- I guess they didn't get the memo this is no longer an actual problem?

 

[diplomat33]

As recently as a month ago Nvidia was still acquiring companies to address the problem of being able to scale mapping for autonomous- I guess they didn't get the memo this is no longer an actual problem?

No, it is not a problem anymore. Mobileye has solved the problem of scaling maps for AVs.

 

[Knightshade]

No, it is not a problem anymore. Mobileye has solved the problem of scaling maps for AVs.

I'm not sure I'd call "we hope to have scaled solutions by 2024" solved.

If it is then I guess FSD has been solved for a while

 

[stopcrazypp]

No, it is not a problem anymore. Mobileye has solved the problem of scaling maps for AVs.

Are Mobileye maps the same as Waymo's? I imagine they are fairly different.

Also I imagine the maps Google have generated for Google maps (which do not have to detail all the road furniture) also are very different from the ones that Waymo is generating and using for their fleet currently (which do include all the road furniture).

 

[diplomat33]

I'm not sure I'd call "we hope to have scaled solutions by 2024" solved.

If it is then I guess FSD has been solved for a while

I say it is solved because their approach addresses every problem and works at scale:
1) They crowdsource the maps so they have a way of collecting the data at scale
2) The maps are detailed enough to work for AVs
3) They are creating and updating the maps at scale

Are Mobileye maps the same as Waymo's? I imagine they are fairly different.

Also I imagine the maps Google have generated for Google maps (which do not have to detail all the road furniture) also are very different from the ones that Waymo is generating and using for their fleet currently (which do include all the road furniture).

Yes, MB's maps and Google Map's are different from Waymo's maps.

Google Maps are not HD as far as I know. They are basically just navigation maps.

MB's maps are generated entirely with camera vision only. Waymo's maps are generated with lidar.

MB's maps are still accurate and detailed but with less features than Waymo's HD maps which are incredibly accurate and with a lot of features.

There is basically a trade off. The more detailed and accurate the map is, the more helpful it will be for AVs, but the harder it will be to scale since it will take longer and cost more. So the question is what level of accuracy and detail is still good enough for reliable AVs but still cost effective to scale. In other words, what is that "middle ground"? MB believes they have found that middle ground.

EDIT: Waymo is sticking with their super HD maps because they want to focus on accuracy and detail of the map in order to make the robotaxi safer. Since they are doing ride-hailing in 50-100 sq mi areas, they don't need super large scale maps. So they are prioritizing detail and accuracy of the maps over size of the maps. However, Waymo has greatly improved the speed and cost of creating and updating their HD maps by automating as much of the process as possible.

You can read more about Waymo's HD maps here:

The Waymo Driver Handbook: How our highly-detailed maps help unlock new locations for autonomous driving

For any emerging technology to be trusted, it helps to first be understood. In the past, people could see how their cars worked, looking under the hood and tinkering with them with the help of a user manual. In 2020, vehicles have so much technology that they’ve become difficult for the general...

blog.waymo.com

You can read more about the different types of maps here:

The Mapping Singularity Is Near

A high falutin’ title to convey how maps are converging and fidelities blending toward a new “medium definition” steady state.

medium.com

 

[Knightshade]

I say it is solved because their approach addresses every problem and works at scale:

Ah, cool.

So Teslas FSD is solved too.

I say this because since it's 2024 before those other guys prove it actually works at scale we can give Tesla same benefit of the doubt!

There is basically a trade off. The more detailed and accurate the map is, the more helpful it will be for AVs, but the harder it will be to scale since it will take longer and cost more. So the question is what level of accuracy and detail is still good enough for reliable AVs but still cost effective to scale. In other words, what is that "middle ground"? MB believes they have found that middle ground.

Wait... you said this was "solved"

Did you really mean "nobody knows the right answer because NOBODY HAS SOLVED IT yet"?

pro tip: Press release/marketing material that says "WE HAVE THE SOLUTION- YOU WILL SEE IT SOME YEARS FROM NOW!" is not a solution. It's a guess.

There's been a LOT of wrong guesses in this field so far.

 

[diplomat33]

Ah, cool.

So Teslas FSD is solved too.

No, because Tesla's approach has not addressed every FSD problem and does not work reliably. Tesla does not have FSD with no driver supervision that works in all conditions with a safety greater than humans. MB does have AV maps at scale.

Wait... you said this was "solved"


Did you really mean "nobody knows the right answer because NOBODY HAS SOLVED IT yet"?

pro tip: Press release/marketing material that says "WE HAVE THE SOLUTION- YOU WILL SEE IT SOME YEARS FROM NOW!" is not a solution. It's a guess.

There's been a LOT of wrong guesses in this field so far.

No, it is not marketing and guesses. MB has collected the data and created the maps and they are using their AV maps today. For example, they've been able to map routes in a few hours and start testing autonomous driving the very next day in cities all over the world from Jerusalem, to Munich to Detroit and more. So it is not guesses. They are using their solution today!

 

[Thp3]

So far the only comparison I understand is the extreme optimism of Mobileye and Tesla.

I have to wonder about Lidar and it’s HD mapping by Waymo. Isn’t this a huge data storage problem? the Lidar data is approaching fractal resolution in my opinion.

We will see how this becomes publically available. Right now, Tesla Autopilot is a huge safety improvement on expressways and making long trips far less taxing. I drive 8-10 hours and it ”feels” like I drove my old ICE car for about three hours.

 

[diplomat33]

I have to wonder about Lidar and it’s HD mapping by Waymo. Isn’t this a huge data storage problem? the Lidar data is approaching fractal resolution in my opinion.

Each Waymo car stores a copy of the HD map for its local geofenced area "offline". So it does not seem to be a problem.

We will see how this becomes publically available. Right now, Tesla Autopilot is a huge safety improvement on expressways and making long trips far less taxing. I drive 8-10 hours and it ”feels” like I drove my old ICE car for about three hours.

I think you are touching on what the public cares about. The public cares about safety and convenience more than they care about whether the car is technically autonomous or not. So AVs like Waymo or Cruise, have to show great safety and convenience in order for the public to pick them over an existing ADAS like Tesla's AP. But if the AVs do achieve greater safety and convenience, then they become much more attractive than ADAS because the AVs also have the extra benefit of not requiring the driver to pay attention.

Humans tend to assume something is safer when they are in control:

"From the safety-in-numbers viewpoint, the technology’s safety benefits are jeopardised by irrational public risk perceptions that mean we underestimate the safety of modes of transport that we presume to control, such as driving, while overestimating the risks of systems that are out of our control and seem uncanny, such as flying." From paper "How can we know a self-driving car is safe?" by Stilgoe.

So humans think driving is safer than flying (even though statistically it is not) because they are in control of driving but are not in control of flying. It is why some people have a fear of flying. Similarly with autonomous driving, people might assume the AV is less safe because they are not in control of the driving (like with flying). And some people might have a fear of riding in the back seat of a robotaxi (similar to fear of flying) because of that lack of control even though the robotaxi might be very safe. I think this can bias people to think that an ADAS is "better" because they are in control and so they think it is safer when it might not be. Of course, AV companies like Waymo need to prove that their technology is indeed safer than human driving.

 

[Knightshade]

No, because Tesla's approach has not addressed every FSD problem

Sure it has. Humans drive with vision, FSD will drive with vision.

It just won't be done for a while. Just like that 2024 mapping solution (that apparently requires help from a bunch of third parties too)

Tesla does not have FSD with no driver supervision that works in all conditions with a safety greater than humans.

Neither does MB. Otherwise you'd be able to buy it today.

MB does have AV maps at scale.

Nope.

If they did their 2024 aspirations wouldn't be over 100x greater than what they have today.

No, it is not marketing and guesses. MB has collected the data and created the maps and they are using their AV maps today.

So have multiple other companies- using entirely different methods as you pointed out.

None of them have scaled worldwide though. Again MB hopes with lots of help from others they will by 2024 actually scale their maps over 100x from what they're mapping today.

While Waymo and others continue to go down entirely different avenues in pursuit of THEIR HD map solutions.

Almost like none of them have actually solved this today or something but all them think THEIR solution, SOMEDAY IN THE FUTURE, will be a solution and everyone else is doing it wrong.

Because until someone ACTUALLY solves it-generally in worldwide and in consumer-safe production- they're all guessing at what is "right"

 

[Thp3]

@diplomat33 The Stilgoe article is very informative. Thanks.

I still disappointed that we seem to never learn from our past. We had new car technology in the early 1900s.

Our road infrastructure was not helpful for car driving. We did not require painted well paved roads to drive cars on.

We had at best limited safety features, a brake and a steering wheel… Some cities even regulated that a person waving a flag had to walk in front of a car to warn people and horses. It was quite a while before even side mirrors were required.

I suspect headlights were mounted on early cars to enable people to claim a car is safer than horse buggy at night.

Autonomous driving will gradually be developed and adopted. But it may take 5-10% of the cars on the road driving with computers before we learn what good regulations are. Until then, I am confident that all liabilities will remain with the human driver despite how many lawsuits try to blame Tesla or other car companies.

I agree with you that a human driver will be part of the autonomous driving system of Tesla for the near future, maybe years. However, I expect car safety to vastly improve and I expect that my Tesla will never hit a deer or pedestrian in 2022 and beyond. (Curbs and stationary objects may still be at risk in 2022, but…)

 

[diplomat33]

@diplomat33 The Stilgoe article is very informative. Thanks.

I still disappointed that we seem to never learn from our past. We had new car technology in the early 1900s.

Our road infrastructure was not helpful for car driving. We did not require painted well paved roads to drive cars on.

We had at best limited safety features, a brake and a steering wheel… Some cities even regulated that a person waving a flag had to walk in front of a car to warn people and horses. It was quite a while before even side mirrors were required.

I suspect headlights were mounted on early cars to enable people to claim a car is safer than horse buggy at night.

I am glad you found the Stilgoe article informative. I found it informative as well. I've been reading a lot about AV safety. It's a fascinating and complex topic.

And yes, regulations and road infrastructure have lagged behind tech innovation.

Autonomous driving will gradually be developed and adopted. But it may take 5-10% of the cars on the road driving with computers before we learn what good regulations are. Until then, I am confident that all liabilities will remain with the human driver despite how many lawsuits try to blame Tesla or other car companies.

I agree with you that a human driver will be part of the autonomous driving system of Tesla for the near future, maybe years. However, I expect car safety to vastly improve and I expect that my Tesla will never hit a deer or pedestrian in 2022 and beyond.

In fact, there are some experts who argue that augmenting human driving with automated safety features is a better approach than pursuing fully autonomous vehicles with no driver. They argue that it will get us to safety greater than humans much quicker than the full autonomy approach. We already have data to suggest that an attentive human driver aided by automated safety features can be safer than a human driver alone. Proving AV safety greater than humans is tricky. Furthermore, it is a cheaper approach than developing fully autonomous vehicles since it requires less sensors and uses established technology. Developing fully autonomous vehicles that don't require a human driver and are much safer than humans is a more daunting and expensive task since it requires teaching a machine to drive better than a human from scratch and requires a lot of sensors and powerful computers.

Long term, I think fully autonomous vehicles will eventually win out. But short term, I think the approach of combining an attentive driver with automated safety features will provide much improved safety.

 

[Thp3]

Long term, I think fully autonomous vehicles will eventually win out. But short term, I think the approach of combining an attentive driver with automated safety features will provide much improved safety.

Yes, that is the mostly likely future. The only caveat from my technology adoption days is that most people overestimate the amount of change that can be made in the short term (one to six months) and underestimate the change that can happen in one or two years.

 

[stopcrazypp]

No, it is not marketing and guesses. MB has collected the data and created the maps and they are using their AV maps today. For example, they've been able to map routes in a few hours and start testing autonomous driving the very next day in cities all over the world from Jerusalem, to Munich to Detroit and more. So it is not guesses. They are using their solution today!

But the core point is they have not demonstrated their maps are good enough for consumer facing use, given AFAIK they don't have a driverless L4 service that services real riders (like Waymo does), nor do they have L3 (Honda's the only one so far that does in a very limited fashion), nor do they have even end-to-end L2 in general release. Seems a lot like counting eggs before they hatch (like what some did in regards to the Audi L3 system that got eventually cancelled).

 

[diplomat33]

But the core point is they have not demonstrated their maps are good enough for consumer facing use, given AFAIK they don't have a driverless L4 service that services real riders (like Waymo does), nor do they have L3 (Honda's the only one so far that does in a very limited fashion), nor do they have even end-to-end L2 in general release. Seems a lot like counting eggs before they hatch (like what some did in regards to the Audi L3 system that got eventually cancelled).

Fair enough. Certainly, MB has not proven their FSD approach yet. But I do like MB's AV mapping approach a lot. And they are using their maps in the real world so I think they are beyond the "marketing and guessing" phase as @Knightshade said.

 

[rxlawdude]

I say it is solved because their approach addresses every problem and works at scale:
1) They crowdsource the maps so they have a way of collecting the data at scale
2) The maps are detailed enough to work for AVs
3) They are creating and updating the maps at scale



Yes, MB's maps and Google Map's are different from Waymo's maps.

Google Maps are not HD as far as I know. They are basically just navigation maps.

MB's maps are generated entirely with camera vision only. Waymo's maps are generated with lidar.

MB's maps are still accurate and detailed but with less features than Waymo's HD maps which are incredibly accurate and with a lot of features.

Waymo is sticking with their super HD maps because they want to focus on accuracy and detail of the map in order to make the robotaxi safer. Since they are doing ride-hailing in 50-100 sq mi areas, they don't need super large scale maps. So they are prioritizing detail and accuracy of the maps over size of the maps. However, Waymo has greatly improved the speed and cost of creating and updating their HD maps by automating as much of the process as possible.

You can read more about Waymo's HD maps here:

The Waymo Driver Handbook: How our highly-detailed maps help unlock new locations for autonomous driving

For any emerging technology to be trusted, it helps to first be understood. In the past, people could see how their cars worked, looking under the hood and tinkering with them with the help of a user manual. In 2020, vehicles have so much technology that they’ve become difficult for the general...

blog.waymo.com

You can read more about the different types of maps here:

The Mapping Singularity Is Near

A high falutin’ title to convey how maps are converging and fidelities blending toward a new “medium definition” steady state.

medium.com

But when mapped road conditions change (e.g., construction and cones subsequent to the mapping process) we KNOW what happens to Waymo. It ain't pretty.

 

[Knightshade]

Fair enough. Certainly, MB has not proven their FSD approach yet. But I do like MB's AV mapping approach a lot. And they are using their maps in the real world so I think they are beyond the "marketing and guessing" phase as @Knightshade said.

I'm unclear how it's NOT guessing if they haven't proven the approach yet.

It's certainly EDUCATED guessing-- but so are the different approaches all the other companies- Tesla included- are trying.


Until one of them proves it works broadly in a consumer facing application, it's all just a guess that it's the 'right' way to do it, with varying degrees of evangelists for one approach versus another as most likely to be right.

 

[diplomat33]

But when mapped road conditions change (e.g., construction and cones subsequent to the mapping process) we KNOW what happens to Waymo. It ain't pretty.

Waymo says that their system is designed to work even if the map is wrong:

"We designed our self-driving system so it can navigate safely if something new or unexpected is encountered, such as new construction."

https://twitter.com/x/status/1286318085851553795

Now obviously, that does not always work out perfectly as we saw with the cones incident. There can be cases where the system still fails. That's why Waymo is still working on FSD. But those incidents are rare. It does not always fail every time the map changes.

 

[Knightshade]

Waymo says that their system is designed to work even if the map is wrong:


Now obviously, that does not always work out perfectly as we saw with the cones incident. There can be cases where the system still fails. That's why Waymo is still working on FSD. But those incidents are rare. It does not always fail every time the map changes.

MEANWHILE.....

https://www.reddit.com/r/teslamotors/comments/ojectw