MobilEye Powers Ahead In ADAS, Will Sell To Others Making Robotaxis

MobilEye, the driver-assist leader which recently returned to markets with a rare successful IPO, each year outlines their progress and plans in an address at CES which is rich with detail. This year they revealed numbers around their latest sales (design wins) with OEM partners, their strategy to move to robotaxi and more.

I’ve written about their strategy before and also interviewed their CEO, Amnon Shashua last year about some of the fine details. As such I will focus on things that are new or changed.

ADAS advances

MobilEye’s EyeQ chips are the undoubted volume leader in the ADAS space, with $17.3B in sales, 1/3rd of it from design wins in 2022. They are also now selling their “Supervision” system, an 11 camera plus radar package that is now in 70,000 Zeekr vehicles in China and has been sold to three other OEMS as a competitor to Tesla Autopilot and the driver-assist version of their FSD package. Different from almost all other players, MobileEye believes in an incremental approach, both within ADAS, and in moving from ADAS to robocar, and then finally within robocars. They feel the “moonshot” approach of working on a robotaxi first is incorrect and requires multiple moonshots. Among the others, Sterling Anderson, co-founder of Aurora Innovations, feels the incremental approach is like trying to get to the moon with taller ladders — so the battle is on.

Many companies focus on the urban robotaxi problem first, in contrast to the consumer robocar which people buy. This decision comes from the fact that making a system work in just a few cities is easier than making it work in a whole country (as consumers demand) and because the cars come home every night and can be refined and improved more easily than car in customer hands. MobilEye believes the reverse, that it’s easier to start from workable consumer ODDs (roads and conditions that the car can handle) like freeways, and expand one “blade” at a time.

MobilEye’s “blades” are:

1. Highway <60km/h, which Shashua believes is not very useful
2. Full speed highway operation
3. Ramp to ramp highway driving
4. Arterial roads
5. Rural roads
6. Urban driving, still with human on board
7. Unmanned urban operation, with remote assistance on rare occasions

While in the past, MobilEye has spoken of running robotaxi services, using its Moovit subsidiary, now they say they will build no more than 50-100 reference robotaxis just for their testing. They expect services to be run by partners, which are both OEMs building vehicles and fleet operators. One recent partnership is with Beep (an operator) and Benteler’s new Holon vehicle. MobilEye claims $3.5 billion in revenue has been booked from this, which dwarfs the booked revenue of any other player in the business.

MobilEye doesn’t want to be Uber or Lyft, they want to sell to companies like that. As such the MobilEye services rumored for Tel Aviv and Munich may not come to pass directly, though they may happen with other partners.

They claim they cost of a full robotaxi hardware package, with compute, 11 cameras, front LIDAR, 3 long range radars and 6 short range radars will eventually be under $5,000. That’s certainly a good price for such a system, since you can remove close to $5,000 in cost by pulling out driver features like the dashboard, wheel, pedals, adjustable seats and mirrors and many other functions, according to calculations by former GM VP Larry Burns.

While everybody has noticed that the attitude on robocars has swung like a pendulum from “any day now” to “2050,” Shashua feels they are indeed coming within just a few years, and they plan to have customers deploying by 2026, and billions sold by 2028.

Validation

To make it happen, you need extremely high MTBFs and proof you’ve done it, which are different. Once again, Shashua put forward the example of getting a 10,000 mile MTBF from their pure vision system, and a 1,000 mile MTBF from their LIDAR/radar only system. Last year I challenged him on how independent failures are, and suggested the idea of getting a 10M mile MTBF by combining the two is wrong. At the time, he agreed, but he said it again, claiming that the error types in the two systems are almost completely independent, though there is no way to prove that to regulators.

Their approach to validating their system is to do shadow driving with the customers using their ADAS systems. In this approach, drivers do ordinary driving using cars equipped with their more advanced chips. Those chips use spare capacity to calculate what they would do if they were driving, and you look for times when the system drove differently from the human, and see if any would cause accidents. They can put it in 100,000 cars, says Shashua.

MobilEye plans to use this approach only for validation, not for development or training as some others have proposed. One reason is that their OEM vendor partners insist on bounded projects with a time limit for such use of their cars. They don’t like to agree to projects that go on forever. This is one of MobilEye’s limitations — they don’t own or control the cars that use their technology, the way Tesla or Waymo do, and so they must get buy-in from car OEMs when they want to make use of them and their data.

The validation process must show there are no reproducible errors. Shashua fears that any reproducible error you knew about as a lawsuit waiting to happen with bad consequences. That means you must fix every bug you find and introduce no new ones in the process.

REM

MobilEye showed a video of the data collected by cars for their mapping system, known as REM. They have millions of cars scanning the roads and sending back distilled info from what they see. The data volume is low — about 50 cents per year in data costs — another constraint required from having car OEM partners. Nonetheless they are able to map all the main roads in a country within a day, most of the rest in a week, and just about everything in a modest number of weeks. That’s a valuable tool that few other companies have (though Tesla and other OEMs could duplicate it if they decided to) though none could have quite so many cars.

MobilEye is using the REM maps to build a simulation of the entire world. They take the road data learned from REM and “hallucinate the entire planet” based on the real roads. Then they have their systems do simulated drives through this photorealistic planet for validation and finding problems. It’s not clear to me that there is too much value in simulating buildings and terrain that aren’t there (though one can get data on such things from other sources.) But it’s a cool project.

MobilEye’s incremental approach is different from that of most teams, though Tesla and a few startups also follow it. Nonetheless, if anybody can pull it off, it will probably be MobilEye. They are certainly the one making money from it while others have yet to prove their business models. MobilEye is worried about the difficulty of proving robotaxi business models, so for now is content to let partners try that out. Of course, they have the option of changing their mind on this and making their own service. I asked Shashua whether it was OK to compete with customers in this manner, and at the time he felt it was no problem, though that view may have changed — and it can change again. Most companies have been obsessed with being the one who owns the end user as a customer — but not MobilEye.