Autonomous Car Progress

[Dan D.]

Really nice breakdown of this paper that was put out by Stanford last fall (article link: here). The Tesla part starts right at 9 minutes in if you want to skip ahead, but the whole thing is worth listening to.

Elon's recent tweet on v9:

https://twitter.com/x/status/1380599694154362882

Could be what Tesla is doing with v9 considering Elon specifically mentioned how much better it is in bad weather. Really has me interested to see what the next beta is like.

Sorry, I disagree entirely with his explanation of Tesla seeing through fog. For one thing he says at 11:00 that humans do not have a memory of what we just saw, as soon as a deer is gone we don't remember seeing it. Perhaps he's burned out his memory, but mine works just fine and I don't just rely on "instantaneously being able to see" something.

Then he compares image stacking and cumulative photon gathering over many hours (Hubble telescope) with how he imagines Tesla could accumulate pixels and form a better picture of what's in the fog. That is not even close to a good analogy. A few frames in real-time is nothing like that.

Also he mentions creating a pseudo-Lidar, pseudo-conformal-tomography landscape using AI vision and masses of processing. Or, you know, you could just use Lidar and get the actual tomography instantly with no time lag. Time lag & accuracy is everything in collision avoidance.

 

[Tam]

An updated timeline for MobilEye:

Delivery Service by 2023
Consumer vehicles by 2025

13 cameras
3 long-range LIDARs
6 short-range LIDARs
6 RADARs
EyeQ system-on-a-chip
Road Experience Management (REM) 3D map collected by EyeQ equipped cars.

 

[cwerdna]

WeRide FSD ride in a crowded street with lots of pedestrians:

I like that it seems to be in real-time. It is not artificially sped up. I also like that we get a clear view of the safety driver and we see that the safety driver is not touching the controls. I also like the perception visualization. The perception seems very accurate.

While doing a Google search, I stumbled across ‘Villages in the City’: WeRide Self-Driving on China’s Unique Urban Roads accompanying the above, that I hadn't read until now.

Chinese autonomous vehicle startup WeRide scores permit to test driverless cars in San Jose – TechCrunch

The California DMV, the agency that regulates autonomous vehicle testing in the state, said the permit allows WeRide to test two autonomous vehicles without a driver behind the wheel on specified streets within San Jose. WeRide has had a permit to test autonomous vehicles with safety drivers behind the wheel since 2017. WeRide is also restricted to how and when it tests these vehicles. The driverless vehicles are designed to operate on roads with posted speed limits not exceeding 45 miles per hour. Testing will be conducted during the day Monday through Friday, but will not occur in heavy fog or rain, according to the DMV.

 

[diplomat33]

While doing a Google search, I stumbled across ‘Villages in the City’: WeRide Self-Driving on China’s Unique Urban Roads accompanying the above, that I hadn't read until now.

Chinese autonomous vehicle startup WeRide scores permit to test driverless cars in San Jose – TechCrunch

Yes. That is correct.

 

[Bladerskb]

An updated timeline for MobilEye:

Delivery Service by 2023
Consumer vehicles by 2025

13 cameras
3 long-range LIDARs
6 short-range LIDARs
6 RADARs
EyeQ system-on-a-chip
Road Experience Management (REM) 3D map collected by EyeQ equipped cars.

Its not an updated timeline by the way. They are still aiming for 2022 in Israel.

However important quote in that article

“What this announcement really underscores is the commercial maturity and readiness of the Mobileye self driving system solution,” he said. “It is ready to scale at large, across tens of thousands of vehicles, multiple states, all the retail store partners in all cities, not geofenced or limited in any way.”

@powertoold

 

[dckiwi]

An updated timeline for MobilEye:

Delivery Service by 2023
Consumer vehicles by 2025

13 cameras
3 long-range LIDARs
6 short-range LIDARs
6 RADARs
EyeQ system-on-a-chip
Road Experience Management (REM) 3D map collected by EyeQ equipped cars.

It's fascinating to see other vendors going hog wild with sensors will Tesla goes vision only! They can't both be the right approach..........

 

[Tam]

WeRide

Nice! It's an L4 permit from California DMV and must "verify vehicles are capable of operating without a driver" and although there's no driver in the car, it still needs "remote operators".

 

[Tam]

It's fascinating to see other vendors going hog wild with sensors will Tesla goes vision only! They can't both be the right approach..........

What counts for me is which has fewer collisions (and fewer deaths of course).

 

[powertoold]

Still my opinion:

Mobileye is a joke. They're a joke because of the trash that comes out of Amnon's mouth:

Don't trust anything that comes out of Mobileye until it's deployed and laypeople are using it.

My prediction is that no consumer-accessible level 4/5 service of any kind will come out of Mobileye or be based on their technology.

 

[Bladerskb]

It looks like 2021 is finally the year that LIDAR has made it to consumer automotive use (not just exclusively for professional use like Waymo, Cruise...):

1) Toyota LS and Mirai Advanced Drive: It's an L2. It doesn't talk about city streets so I assume it's like from on-ramp to off-ramp Tesla Navigation on Autopilot for the highway only. It's been on sales starting 4/8/2021.

LS version L AWD= JPY15,340,000 = $140,222.17

LS version L AWD with Advance Drive= JPY16,320,000=$149,180.30

That's more than $9,000 for the option but by paying its competitor Tesla $10,000 FSD, Tesla will do city streets.

2) Honda Sensing Elite Traffic Jam Pilot: It's an L3 with a maximum speed of 31 MPH (50 kph). Auto Lane Change still has to be initiated by the driver's flipping the turn signal stalk (Tesla has an option to skip the driver's confirmation). Only 100 Honda Legend Hybrid EX with the system are for lease in Japan starting on 3/5/2021 for about $102,000. Its law holds the car manufacturer responsible while the system is activated and it's totally legal for the driver to be distracted with watching videos, reading e-mails... while the L3 system takes over.

And yes, the world will know that you have the system because you have to display the sign "Automated Drive" on the back of your car (currently in Japan only because the system is not sold outside of Japan):

Audi A8 also has that similar L3 since 2017 but it has not released to consumers because the laws were ambiguous with L3 driving: California consumers would still be ticketed for watching the video, holding the cell phone to talk while letting the L3 do the driving for them.

It'll be interesting to see how many accidents these LIDAR cars will get while being in the hand of consumers.

By end of 2022, there will be up to 10 cars using High resolution long range lidars and I'm not even counting the cars above.

 

[Bladerskb]

Still my opinion:

Mobileye is a joke. They're a joke because of the trash that comes out of Amnon's mouth:

Don't trust anything that comes out of Mobileye until it's deployed and laypeople are using it.

My prediction is that no consumer-accessible level 4/5 service of any kind will come out of Mobileye or be based on their technology.

What Trash? But ofcourse the nonesense that comes out of elon's mouth is biblical right?

Yet we should trust everything that comes out of Elon's mouth? Hows that Level 5 in 6 months working out? Its almost exactly 6 months.

 

[powertoold]

By end of 2022, there will be up to 10 cars using High resolution long range lidars and I'm not even counting the cars above.

Amnon says Mobileye currently has a LIDAR/RADAR-only car that has the same performance of their camera-only car. (source: 13:20 in the video)

Have you ever seen this car in action / video / picture?

 

[MP3Mike]

Amnon says Mobileye currently has a LIDAR/RADAR-only car that has the same performance of their camera-only car. (source: 13:20 in the video)

Have you ever seen this car in action / video / picture?

If that is true that would mean the performance of their camera-only car system has to be very crappy...

 

[Bladerskb]

Amnon says Mobileye currently has a LIDAR/RADAR-only car that has the same performance of their camera-only car. (source: 13:20 in the video)

Have you ever seen this car in action / video / picture?

17 mins 20s

 

[Tam]

...Don't trust anything that comes out of Mobileye until it's deployed and laypeople are using it...

That's reasonable.

Nevertheless, it might be still informative to get their rationales for the timeline.

For Tesla, the rationale has been that the improvement would be so drastic, so exponential. So after the slow start in 2016, and with so many cameras in the Tesla fleet, the AI could have so much data to quickly learn. Thus, the autonomous Coast-to-Coast demo could be done easily by the end of 2017 (4 years ago). Tesla would run 1 million robotaxis by 2020 (last year). And of course, the vision-only firmware coming in this month April 2021 (the download button is still missing from last month) would be so fantastic that Tesla RADAR will be removed!

For MobilEye, the LIDAR made by Luminar is still quite expensive in the $10,000 to $20,000 range so its system will be for commercial use first and with the help of Intel, it will be able to make cheaper LIDAR by 2025 for consumer use at around "few thousand dollar range".

I myself love science fiction, and I want more, just as long as the story is logical. And in science fiction, the story might makes sense but things might not happen in real life of course.

 

[Bladerskb]

That's reasonable.

Nevertheless, it might be still informative to get their rationales for the timeline.

For Tesla, the rationale has been that the improvement would be so drastic, so exponential. So after the slow start in 2016, and with so many cameras in the Tesla fleet, the AI could have so much data to quickly learn. Thus, the autonomous Coast-to-Coast demo could be done easily by the end of 2017 (4 years ago). Tesla would run 1 million robotaxis by 2020 (last year). And of course, the vision-only firmware coming in this month April 2021 (the download button is still missing from last month) would be so fantastic that Tesla RADAR will be removed!

For MobilEye, the LIDAR made by Luminar is still quite expensive in the $10,000 to $20,000 range so its system will be for commercial use first and with the help of Intel, it will be able to make cheaper LIDAR by 2025 for consumer use at around "few thousand dollar range".

I myself love science fiction, and I want more, just as long as the story is logical. And in science fiction, the story might makes sense but things might not happen in real life of course.

The Luminar lidar cost Mobileye $1,000.
The same lidare cost Volvo under $500 due to scale.
Livox, Innovision, etc all going into mass production this year and next year, all cost $250-$500.

The "Lidar is expensive and cost tens of thousands" myth needs to die!

 

[Microterf]

I've never cared about the cost of the lidar, that's always going to go down. I've always seen the problem as a scale problem for then and a sensor fusion problem for those companies. Maybe they have it figured out. Only time will tell who's right, or both approaches might be work, but cost structures will be different.

 

[mark95476]

Oh my... So many here try to convince others that Tesla was/is/will be totally being crushed by the competition.

Lets see what the volumes (or lack-of) amount to. These legacy OEMs manufacturer >10M vehicles a year while Tesla is at 0.5M/year.

It looks like 2021 is finally the year that LIDAR has made it to consumer automotive use (not just exclusively for professional use like Waymo, Cruise...):

1) Toyota LS and Mirai Advanced Drive: It's an L2. It doesn't talk about city streets so I assume it's like from on-ramp to off-ramp Tesla Navigation on Autopilot for the highway only. It's been on sales starting 4/8/2021.

LS version L AWD= JPY15,340,000 = $140,222.17

LS version L AWD with Advance Drive= JPY16,320,000=$149,180.30

That's more than $9,000 for the option but by paying its competitor Tesla $10,000 FSD, Tesla will do city streets.

2) Honda Sensing Elite Traffic Jam Pilot: It's an L3 with a maximum speed of 31 MPH (50 kph). Auto Lane Change still has to be initiated by the driver's flipping the turn signal stalk (Tesla has an option to skip the driver's confirmation). Only 100 Honda Legend Hybrid EX with the system are for lease in Japan starting on 3/5/2021 for about $102,000. Its law holds the car manufacturer responsible while the system is activated and it's totally legal for the driver to be distracted with watching videos, reading e-mails... while the L3 system takes over.

And yes, the world will know that you have the system because you have to display the sign "Automated Drive" on the back of your car (currently in Japan only because the system is not sold outside of Japan):

Audi A8 also has that similar L3 since 2017 but it has not released to consumers because the laws were ambiguous with L3 driving: California consumers would still be ticketed for watching the video, holding the cell phone to talk while letting the L3 do the driving for them.

It'll be interesting to see how many accidents these LIDAR cars will get while being in the hand of consumers.

 

[Tam]

Oh my... So many here try to convince others that Tesla was/is/will be totally being crushed by the competition.

Lets see what the volumes (or lack-of) amount to. These legacy OEMs manufacturer >10M vehicles a year while Tesla is at 0.5M/year.

My presentation does not mean that others are crushing Tesla.

In the context, why buy Lexus $9,000 option just for highway automation when you can pay an extra $1,000 and get Tesla City Street Robotaxi that can earn you a net profit of $30,000 annually?

My context is: The pros have accessed LIDAR before, but now, this year, people can pay to get their cars with LIDAR.

It's understandable that the pros don't collide their cars into a stationary police car, but now with regular consumers, let's see whether LIDAR can prevent collisions for everyday consumers.

It's not about competition but it's about which system is safer regardless of other factors (sales, price, popularity, volumes...)

 

[Tam]

...sensor fusion problem...

With the 2016 fatal autopilot crashing into and under the white tractor-trailer, Tesla explained the sensor fusion problem: "Neither Autopilot nor the driver noticed the white side of the tractor trailer against a brightly lit sky, so the brake was not applied."

But so far, none has voiced complaints against LIDAR and its fusion in such a scenario.

MobilEye describes its fusion in terms of separate systems of

1) Vision
2) LIDAR + RADAR

And when one fails, the other would take over