Which ADAS is Really Driving? It’s Mobileye!

[pamalabama]

One question I have that we can speculate answers for:

how does the lane change assist work if the all the autonomy is based on the front camera MobileEye??

wonder if they can signal to MobileEye to do certain things? Speed up/down lane change. Or does it use the side rear view cameras too?

rivian gen 2 licensed a more expensive mobileye system that supports 360


rivian gen 1 can probably do these things too but rivian doesn't want to spend any resources into developing it.

Besides, tesla did it 10 years ago with autopilot 1 which was also a mobileye product. What caused mobileye to leave was the fact that tesla was doing auto lane change, navigate on autopilot, etc. with only a front facing camera. They were operating "blind" using ultrasonics and radars

Sponsored

 

[godfodder0901]

One question I have that we can speculate answers for:

how does the lane change assist work if the all the autonomy is based on the front camera MobileEye??

wonder if they can signal to MobileEye to do certain things? Speed up/down lane change. Or does it use the side rear view cameras too?

They don't signal MobileEye to do anything. They only get data out of it.

 

[portdirect]

They don't signal MobileEye to do anything. They only get data out of it.

If it’s MobileEye, then it’s quite likely that the mobileye unit is actually controlling the R1 whist active. Supervision (EyeQ 5 SoC) has the *exact* same capabilities that Rivian advertise currently for RAP: https://www.mobileye.com/solutions/super-vision/ and is also compatible with the sensor array in the gen2 r1 (360 camera, radar) - quite a co-incidence.

Edit: I just realized that the new kinda green light detector in the next update is not a Supervision feature, though not controlling movement - nice to see that they are for sure making progress on their own stack in places

 

[godfodder0901]

If it’s MobileEye, then it’s quite likely that the mobileye unit is actually controlling the R1 whist active. Supervision (EyeQ 5 SoC) has the *exact* same capabilities that Rivian advertise currently for RAP: https://www.mobileye.com/solutions/super-vision/ and is also compatible with the sensor array in the gen2 r1 (360 camera, radar) - quite a co-incidence.

Edit: I just realized that the new kinda green light detector in the next update is not a Supervision feature, though not controlling movement - nice to see that they are for sure making progress on their own stack.

Maybe read that post again...

Super Vision is significantly more capable than RAP...

 

[portdirect]

Maybe read that post again...

Super Vision is significantly more capable than RAP...

Chauffeur is, but Supervision Lite when running on a windshield mounted system like the EyeQ™ 6 Lite seems to match RAP in the gen2 pretty well.

Edit: sigh, sorry @godfodder0901 ; the marketing page for supervision has been updated to the EyeQ 6 version - I’ll see if I can find the old one on waybackmachine.

 

[pamalabama]

Also worth mentioning that training data alone does not make a good ADAS system. I would not be surprised if most of an ADAS system's performance is dependent on simulated data.

Waabi is doing self driving trucks and they are using almost 100% simulation for training

Also worth mentioning rivian cameras are 60hz which is not a good thing for self driving. You do not need that high of a framerate and it is better to utilize that sensor time to achieving higher dynamic range.

 

[pamalabama]

https://www.rivianforums.com/forum/...ous-driving-history-present-and-future.29179/

These are the various systems rivian is using

 

[George Kaplan]

It's going to be a LOOOOONG time before I think ANY system is ready for eyes off driving. Tesla isn't far ahead enough for it to matter for me, as I wouldn't trust it unsupervised any more than I'd trust Rivian's in their current state. I see videos of mistakes it makes constantly: hitting berms, going the wrong way on roundabouts, hitting stopped cars/ emergency vehicles, running kids over. That doesn't sound like something trustworthy to me.


All of these systems are ASSISTANTS, and just as I wouldn't trust my Assistant's work without checking it I don't trust these systems. It will be years of using them supervised without making dumb mistakes before I consider it. I'm not doing these stupid and simple things while driving, so me driving is the safter choice.

It’s like I have a twin.

I’ve stated this elsewhere: We gather here as people with a common interest and the echo chamber may inflate the concept of how much interest there is in L3 and more advanced ADAS.

I think the average person is somewhere between disinterested and scared shitless by the idea of their car “driving itself”. I read the breathless recaps from Tesla drivers about how their car drove for over an hour with “only two interventions“. But anything less than consistent, repeatable and dependable zero interventions is the only true FSD, and Tesla—which has a huge first-mover advantage over Rivian and everyone else—clearly isn’t there, while Musk continues to promise Cybercab—L5 driving from software which is currently L2+ at best—next year.

This is essential to him because he bet the future growth of the company on Cybercab, instead of designing and building a car smaller and lower cost than Model 3 to gain market and sales growth.

As for Rivian, Gen 1 isn’t even a good L2—lane centering remains an abstract concept—and I have no experience with Gen 2, but I can’t imagine it is leaps and bounds better. I’d rather the company deploy its engineering might towards making other refinements on the R1 vehicles. Gaining “new” features is always a Christmas in July surprise, but improving ride quality is even better.

If R2 is competently eyes-off out of the box with a clean sheet of engineering design paper, that’s great. But I think it’s a lost cause for Gen 1 at least, and would rather have 1,001 other features of the car improved and refined.

 

[scottf200]

It's only uploaded when you are at home, my car stays plugged in most of the time. So, I have no problem with any battery drain.

You still have vampire drain when plugged in, it's just shifted from your battery to your home's power circuit.
That daily 1-1.5kWh of energy has to come from somewhere. The vehicle will pull it from your 12V battery (which the vehicle will recharge from the big battery) or it will use the L1 or L2 connection to charge the big battery which in turn will keep the 12V battery charged.

Even at 20 Mbps we are talking just a couple hours. The Rivian can sleep the rest of the time. 20 Mbps is somewhat slow these days (eg. cable)

• Convert 20 GB to gigabits:
  • 20 GB × 8 = 160 gigabits.
• Divide by upload speed (20 Mbps):
  • 160 gigabits ÷ 20 Mbps = 8,000 seconds.
• Convert to hours:
  • 8,000 seconds ÷ 3,600 seconds/hour ≈ 2.22 hours.

 

[Donald Stanfield]

It’s like I have a twin.

I’ve stated this elsewhere: We gather here as people with a common interest and the echo chamber may inflate the concept of how much interest there is in L3 and more advanced ADAS.

I think the average person is somewhere between disinterested and scared shitless by the idea of their car “driving itself”. I read the breathless recaps from Tesla drivers about how their car drove for over an hour with “only two interventions“. But anything less than consistent, repeatable and dependable zero interventions is the only true FSD, and Tesla—which has a huge first-mover advantage over Rivian and everyone else—clearly isn’t there, while Musk continues to promise Cybercab—L5 driving from software which is currently L2+ at best—next year.

This is essential to him because he bet the future growth of the company on Cybercab, instead of designing and building a car smaller and lower cost than Model 3 to gain market and sales growth.

As for Rivian, Gen 1 isn’t even a good L2—lane centering remains an abstract concept—and I have no experience with Gen 2, but I can’t imagine it is leaps and bounds better. I’d rather the company deploy its engineering might towards making other refinements on the R1 vehicles. Gaining “new” features is always a Christmas in July surprise, but improving ride quality is even better.

If R2 is competently eyes-off out of the box with a clean sheet of engineering design paper, that’s great. But I think it’s a lost cause for Gen 1 at least, and would rather have 1,001 other features of the car improved and refined.

Gen 2’s lane keep is substantially improved over gen 1. Gen 2 is supposed to get hands off this coming update. I’ll use it, but with the idea I’m going to have to intervene occasionally. There is no system in existence I would let drive me without constant vigilance.

 

[George Kaplan]

Might very well be true. For level 3 you need to impose speed caps and also detect a lead car from a reasonable distance.

Tesla FSD E2E has about the least control you can imagine. It drives one speed and you have to accept the speed it drives. Turning on the turn signal does not even cause the car to change lanes anymore.

Level 3 will only be on premapped roads, no lane changes, with a lead car in perfect weather. That's not a hard problem to solve.

Correct. Mercedes Benz Drive Pilot system is approved only in Southern California and in Nevada, and only for use below 40 mph, no rain, a car in front to follow, and daylight. The restrictive nature is because at L3 the manufacturer (in this case, M-B) is legally liable in a crash, since the “driver” isn’t driving.

This makes it a remedy for rush hour traffic on the highway and little else. Germany has since raised the speed limit for Drive Pilot on the autobahn, but no increases so far in the US.

BTW, this should give pause to Tesla shareholders. It’s easy to get giddy about the parlor trick of the car becoming L4 (for current Teslas with the right hardware) or L5 (for Cybercab) , but Tesla would carry full legal responsibility for the safety and operation of the vehicles when the Tesla software is engaged (which for Cybercab would be 100% of the time). If the car is involved in any accident with a level of fault, Tesla is the liable party, not the car owner or passengers.

The software needs to be 101% locked-down solid if Tesla is to avoid crippling liability lawsuits from accidents from software fails.

 

[Riviot]

There is no system in existence I would let drive me without constant vigilance.

Does this extend to Mrs. Stanfield?

 

[Donald Stanfield]

Does this extend to Mrs. Stanfield?

No. I don't consider people in the same category as assistance systems. My wife drives an i4 M50. I'm not too worried when she's driving. There are some people I won't get in a car with when they are driving but not too many. Mainly because I don't find occasion to ride with too many people very often.

 

[pamalabama]

Correct. Mercedes Benz Drive Pilot system is approved only in Southern California and in Nevada, and only for use below 40 mph, no rain, a car in front to follow, and daylight. The restrictive nature is because at L3 the manufacturer (in this case, M-B) is legally liable in a crash, since the “driver” isn’t driving.

This makes it a remedy for rush hour traffic on the highway and little else. Germany has since raised the speed limit for Drive Pilot on the autobahn, but no increases so far in the US.

BTW, this should give pause to Tesla shareholders. It’s easy to get giddy about the parlor trick of the car becoming L4 (for current Teslas with the right hardware) or L5 (for Cybercab) , but Tesla would carry full legal responsibility for the safety and operation of the vehicles when the Tesla software is engaged (which for Cybercab would be 100% of the time). If the car is involved in any accident with a level of fault, Tesla is the liable party, not the car owner or passengers.

The software needs to be 101% locked-down solid if Tesla is to avoid crippling liability lawsuits from accidents from software fails.

As much as people love to use the phrase "my car tried to kill me," I have never in a modern version of FSD seen anything that would get anyone seriously hurt. Not at least one with an obvious solution. For example, the guy whose car drove onto the train tracks in LA could be solved with mapping. You can have navigation avoid crossing a train track entirely.

If you look at tesla's current disengagements, 75% of them are because of mapping issues. Maps are low hanging fruit because you just drive around, record all disengagements and fix the maps in those areas. Maps are done using auto labeling and machine learning. They are a relatively straightfoward problem.

Tesla then needs to fix sign recognition which can theoretically be done by mapping any offending signs that are critical to read. Notice how tesla reads not much other than stop sign, speed limit and traffic lights and can drive with mostly no problems. School zones/end school zone can be mapped as that requires just checking whether the light is flashing. And besides it would not be usually illegal to follow a school zone speed during off hours.

For example, tesla does not read yield signs as generally the roads themselves infer whether to yield or not without having to read the sign. The are however exceptions that might need to be mapped.

In the same vein , one way and no u-turn signs are integrated into the navigation and route planning.

you can also significantly reduce risk of injury or death by driving at a lower speed which is what waymo does when avoiding interstate travel.

The other advantage for tesla is the cybercab will cost $15K to manufacture and the panels are made of injection molded plastic. That means repair costs and totaling a vehicle will be inexpensive. The battery is only 40kWh which they achieve a good range by getting 5.5+ miles per kWh

Tesla accident rate just needs to be lower than human which is not a very high bar as humans get into accidents all the time. The cars are still insured and everything will play out as normal.

 

[Supratachophobia]

Id just like to add that the mobile eye system, for it's intended feature set, is pretty solid. We get into problems when we try to make it do something it wasn't intended to do.

Sponsored