Unedited Mobileye’s Autonomous Vehicle & other CES 2020 Self Driving ride videos

[GrimRe]

The validation piece I think is much, much harder than what people give credit to. Tesla have a huge lead on the trials and tribulations of running a CNN in the real world. The MobilEye demo ware is very impressive because its a green field approach, there are no cars with the solution shipped right now to maintain and not have regress.

I would assume that Tesla have internal builds very similar if not better than what MobilEye demonstrated. Tesla are however (imo) purposely holding it back to prevent a) lots of accidents ending up on YouTube and b) surge in service centre costs to replace HW2+ to HW3.

 

[diplomat33]

Yup but the most important part is who is doing the driving policy.

I think Mobileye is providing the driving policy.

By the way the 4x parking cameras are fed into VIDAR so they definitely use it even outside of parking scenarios.

Good to know. Thanks.

 

[Bladerskb]

The validation piece I think is much, much harder than what people give credit to. Tesla have a huge lead on the trials and tribulations of running a CNN in the real world. The MobilEye demo ware is very impressive because its a green field approach, there are no cars with the solution shipped right now to maintain and not have regress.

I would assume that Tesla have internal builds very similar if not better than what MobilEye demonstrated. Tesla are however (imo) purposely holding it back to prevent a) lots of accidents ending up on YouTube and b) surge in service centre costs to replace HW2+ to HW3.

There are probably over a million cars with EyeQ4 and the chip includes a lot of what you saw in that presentation. Including the green semantic free space if that's what you were referencing.

https://i.imgur.com/b7ar1cR.png

 

[diplomat33]

There are probably over a million cars with EyeQ4 and the chip includes a lot of what you saw in that presentation. Including the green semantic free space if that's what you were referencing.

https://i.imgur.com/b7ar1cR.png

Yes. And Amnon even mentioned that Mobileye is collecting data from those 1M cars to help build their HD maps. So much for the argument that Tesla is the only one with a large fleet of cars on roads collecting data. LOL.

 

[Soda Popinski]

Yes. And Amnon even mentioned that Mobileye is collecting data from those 1M cars to help build their HD maps. So much for the argument that Tesla is the only one with a large fleet of cars on roads collecting data. LOL.

That's pretty cool. I can see them using VIDAR instead of LIDAR to generate those HD maps. Do we know how much data the cars are uploading?

 

[timewasted]

The validation piece I think is much, much harder than what people give credit to. Tesla have a huge lead on the trials and tribulations of running a CNN in the real world. The MobilEye demo ware is very impressive because its a green field approach, there are no cars with the solution shipped right now to maintain and not have regress.

I would assume that Tesla have internal builds very similar if not better than what MobilEye demonstrated. Tesla are however (imo) purposely holding it back to prevent a) lots of accidents ending up on YouTube and b) surge in service centre costs to replace HW2+ to HW3.

It’s been over three years since AP2.0 was unleashed upon the world, and every day since then people have speculated about the existence of a super secret code branch for autopilot that can actually achieve level 4 (or even 5!) autonomous driving.

Elon wants money, and what better way to get that than to release the first true autonomous vehicle? Smart Summon is proof that he doesn’t hold back things until they’re ready to be released. So let’s let the notion of this truly advanced capability being locked up somewhere in Fremont. It doesn’t exist.

 

[diplomat33]

Probably worth noting that the Mobileye demo highlights the following city driving scenarios:
- Non signaled 4 way intersection
- Yielding to a vehicle pulling out a parking space
- Road construction
- Coping with roadblocks and navigating oncoming traffic
- Unprotected left turns in heavy traffic
- Busy roundabout

These are often tricky driving scenarios. It is a very positive sign that Mobileye's car can handle these situations so well. It shows that Mobileye not just has excellent perception but excellent planning and excellent driving policy as well.

 

[arghx7]

Jerusalem has very little snow and rain and most of those roads were clearly marked. Mobileye may have very little capability in inclement weather. Autosteer does a decent job of at least holding lanes in bad storms, as long as all the sensors and cameras aren't blocked.

The other thing you have to keep in mind - if Tesla can do 80% of what the competitors can do for half or less of the cost, they are ahead. They are actively monetizing FSD, even if it's only level 2 and stays that way for a while (hopefully city NoA this year). People are paying real money for FSD, even if some are understandably getting impatient. It's likely a profit center for Tesla already, or will be soon as they recognize deferred revenue (forget about robotaxis). These CES demos are to try and get OEM's interested in dumping in millions to make a real production ready solution. Prototype vehicles can cost hundreds of thousands of dollars each; they are money pits.

I'm sure the Mobileye route was carefully chosen and most of the bugs worked out. It's a more complicated version of Tesla's FSD demo,that's all.

Tesla's betting that they can stick with the limited camera and sensor suite and progressively put more and more computing power into it, retrofitting the FSD computer multiple times. It's a reasonable gamble. It's way easier to swap one control module multiple times and reuse the same sensor suite. For other OEM's the customers will have to opt into the sensor suite. So they have millions of cars on the road that are definitely dead end for FSD capability. You can't take a million Rav4's in the market and install 8 cameras on them (I mean you could, but seems infeasible due to the places cameras are mounted). But you might be able to retrofit a million Model 3's to a new HW4 computer, when they were specifically designed to be upgradeable from the beginning. It's no more work than a major recall which often has millions of affected vehicles.

Teslas current sensor suite might work out with progressively more powerful FSD computers--they are taking a gamble, and it will be fascinating to see if they pull it off.

 

[VanillaAir_UK]

Teslas current sensor suite might work out with progressively more powerful FSD computers--they are taking a gamble, and it will be fascinating to see if they pull it off.

Well I hope the B pillar and non recording forward cameras have good coverage as 80% of traffic traversing this roundabout is invisible to the recording forward and right 'fender' cameras.

 

[diplomat33]

Well I hope the B pillar and non recording forward cameras have good coverage as 80% of traffic traversing this roundabout is invisible to the recording forward and right 'fender' cameras.

View attachment 498353

Not surprising. The B Pillar camera is the camera that is responsible to see that part of the roundabout.

 

[eli_]

I just see these Mobileye presentations as the "Missing Tesla FSD Roadmap", like if Tesla was actually transparent and did dev blogs or something, we would see an awful lot of similarities between the two. A few different priorities but in general, two vision-first self-driving systems are going to end up with a lot of the same implementation.

 

[Eno Deb]

The other thing you have to keep in mind - if Tesla can do 80% of what the competitors can do for half or less of the cost, they are ahead. They are actively monetizing FSD, even if it's only level 2 and stays that way for a while (hopefully city NoA this year). People are paying real money for FSD, even if some are understandably getting impatient. It's likely a profit center for Tesla already, or will be soon as they recognize deferred revenue (forget about robotaxis). These CES demos are to try and get OEM's interested in dumping in millions to make a real production ready solution. Prototype vehicles can cost hundreds of thousands of dollars each; they are money pits.

Ahem. Mobileye generated about $1 billion in revenue in 2019 with its ADAS systems, with about 17 million new vehicles using their solution. Many of them are collecting road data, partly in preparation of their autonomous driving solution.

I'm sure the Mobileye route was carefully chosen and most of the bugs worked out. It's a more complicated version of Tesla's FSD demo,that's all.

Tesla hasn't shown anything close to this test drive.

 

[VanillaAir_UK]

Not surprising. The B Pillar camera is the camera that is responsible to see that part of the roundabout.

Well they are going to have to sort the misting up and blinding of the B pillar cameras that are causing major loss of use for many of us this time of year. One thing temporary losing lane change, but another if it becomes a primary camera for junctions.

 

[Ludalicious]

Very impressive video. It really does make you think if Tesla's current camera's can really pull it off. The vehicle in the video has 12 cameras in total. All look to be well positioned too

 

[arghx7]

Ahem. Mobileye generated about $1 billion in revenue in 2019 with its ADAS systems, with about 17 million new vehicles using their solution. Many of them are collecting road data, partly in preparation of their autonomous driving solution.
Tesla hasn't shown anything close to this test drive.

In terms of volume the vast majority of those systems they are selling to other OEM's aren't even comparable to AP1 in capability. It's apples and oranges comparison. Mobileye does not have a massive fleet to query like Tesla does. They are just a supplier. Telematics are only now becoming standard to other OEM's new products, and it seems unlikely they are giving at-will access to Mobileye for data.

Mobileye may still be ahead of Tesla in some ways but it's not obvious that this demo means they are closer in a scaled high volume production solution. Also Waymo, Lyft, etc, are all bleeding tons of money on their autonomous test fleet with expensive 360 degree Lidar. How many driverless Uber rides does it take to make up for a small fleet (small relatively speaking) of retrofitted $100k+ cars and the standing army needed to keep them going?

it's just basic arithmetic. Amortizing these expensive low volume systems is going to take a long time. If Tesla can pull off comparable ability a year later for massively lower hardware costs they win, and even if they don't they are still selling $7000 level 2 systems.

If a 5 mile Uber ride with a driver running his 5 year old Corolla is I don't know, $10, how much does it have to be for a driverless Uber with a retrofitted $100k car to recoup that? Costs matter. Automation does not beat humans if it performs the same but it's far more expensive.

 

[Eno Deb]

In terms of volume the vast majority of those systems they are selling to other OEM's aren't even comparable to AP1 in capability.

Don't confuse gimmicky beta features (like Smart Summon and NoA) with capability. At it's core Autopilot still is just TACC + lane keeping, which many cars have.

It's apples and oranges comparison. Mobileye does not have a massive fleet to query like Tesla does. They are just a supplier.

Telematics are only now becoming standard to other OEM's new products, and it seems unlikely they are giving at-will access to Mobileye for data.

No idea what you mean with "at will access". Mobileye works with carmakers like BMW on road data collection. The carmakers know that it's in their mutual benefit.

Also Waymo, Lyft, etc, are all bleeding tons of money on their autonomous test fleet with expensive 360 degree Lidar. How many driverless Uber rides does it take to make up for a small fleet (small relatively speaking) of retrofitted $100k+ cars and the standing army needed to keep them going?

it's just basic arithmetic. Amortizing these expensive low volume systems is going to take a long time.

Not sure what that has to do with Mobileye.

If Tesla can pull off comparable ability a year later for massively lower hardware costs they win, and even if they don't they are still selling $7000 level 2 systems.

That's a big if. There is zero evidence that Tesla's hardware costs will be "massively lower" than a Mobileye system as shown in the video, if and when Tesla gains similar capabilities. Mobileye's EyeQ chips are manufactured in vastly bigger quantities than Tesla's Neural Net acclerator, so the economies of scale will more likely be in their favor.

 

[Wooloomooloo]

Tesla definitely should have stuck with these guys - that really does make Summon look like a bit of a joke. It also makes me wonder about the wisdom of Tesla making their own silicon, which is only a good idea if you're making better (more effective) chips than the competition, like Apple A-series chips. The jury is out on that one, but Intel are investing very heavily in AI-assisted tasks in their chip.

Very impressive demo, thanks for sharing.

 

[Bladerskb]

In terms of volume the vast majority of those systems they are selling to other OEM's aren't even comparable to AP1 in capability. It's apples and oranges comparison. Mobileye does not have a massive fleet to query like Tesla does. They are just a supplier. Telematics are only now becoming standard to other OEM's new products, and it seems unlikely they are giving at-will access to Mobileye for data.

Mobileye may still be ahead of Tesla in some ways but it's not obvious that this demo means they are closer in a scaled high volume production solution. Also Waymo, Lyft, etc, are all bleeding tons of money on their autonomous test fleet with expensive 360 degree Lidar. How many driverless Uber rides does it take to make up for a small fleet (small relatively speaking) of retrofitted $100k+ cars and the standing army needed to keep them going?

it's just basic arithmetic. Amortizing these expensive low volume systems is going to take a long time. If Tesla can pull off comparable ability a year later for massively lower hardware costs they win, and even if they don't they are still selling $7000 level 2 systems.

If a 5 mile Uber ride with a driver running his 5 year old Corolla is I don't know, $10, how much does it have to be for a driverless Uber with a retrofitted $100k car to recoup that? Costs matter. Automation does not beat humans if it performs the same but it's far more expensive.

good article explaining mobileye's advantages
Intel’s Mobileye has a plan to dominate self-driving—and it might work

 

[diplomat33]

good article explaining mobileye's advantages
Intel’s Mobileye has a plan to dominate self-driving—and it might work

Great article. I really like Mobileye's approach.

 

[vinny84]

To me this statement was the most interesting one.

Shashua expects the hardware for these initial self-driving taxis to cost $10,000 to $15,000 per vehicle. By 2025, Shashua is aiming to "reduce the cost of a self-driving system below $5,000."