MobilEye is an Israeli company, a unit of Intel, which has a dominant position in the driver assistance (ADAS) market among legacy automakers. They made the original Tesla Autopilot, but broke away from Tesla after a fatal crash. They have had their own ambitions in supervised “pilot” systems for many years, and also plan to make actual self-driving systems all the way up to robotaxis.
MobilEye’s business is selling chips and software to automakers, and while it briefly flirted with a plan to run its own robotaxi business, it currently hopes to do that, under the name “MobilEye Drive” in partnership with automakers, such as VW commercial vehicles, Rimac’s “Verne,” and some Chinese vendors. Under the name “MobilEye Chauffeur” they have several partners hoping to deploy part-time autonomy systems where a car can drive itself without human attention on freeeways, but a human driver is on standby when entering zones the system doesn’t handle.
When it comes to more basic ADAS, MobilEye’s systems are in a very large assortment of vehicles. It’s the most popular system in that area. Their “MobilEye SuperVision” system is in a variety of cars and competes with Tesla Autopilot, Ford BlueCruise, and BYD NOA, to name a few. In spite of this, MobilEye’s financial results this week disappointed, resulting in a market haircut of 20%. (A portion of MobilEye is public, the majority belongs to Intel.)
At CES 2025, MobilEye CEO Amnon Shashua presented a contrast between their plan to reach complete self-driving (sometimes incorrectly called “level 5”) and the plans, as he saw them, of Waymo and Tesla. Waymo is the undisputed leader, being the first company to make a working self-driving car, but Tesla perhaps gets the most press, though currently very far behind.
Shashua presented progress on two axes, borrowing names from data science and computer perception. He termed the safety performance as “precision” and a number of factors controlling the usefulness of the system, such as geographic breadth (service area) and affordability as “recall.”
While the safety performance goes from 0 to 100%, only values very near 100%–where you are ready to “bet your life” on the system as it drives better than most humans–are real self-driving, and can lead to an automotive revolution. Lower levels are still useful (and indeed, are almost all of MobilEye’s business) but they need a fully attentive supervising human driver, whose task they can make a little easier.
The other dimension is important though, as you want to be able to use it in lots of places and you want people to be able to get access to it and afford it. You really want both for the true revolution. If you want to sell a consumer car, it had better drive almost everywhere. On the other hand, a taxi service that just serves a large town like New York or San Francisco can be a lucrative business. Indeed, a system that just drive freeways can be a valuable luxury car amenity, and MobilEye is building that. (Mercedes already sells such a system but it is currently very limited.)
Shashua depicts Waymo’s path as working first to meet the full safety requirement, and then expanding the “recall.” Waymo has working vehicles in 3 cities and has announced expansion to several more for this year and the next. Some depict Waymo’s growth as too slow, limited by the path they have chosen. Waymo’s path is also the path of almost all other robotaxi projects, and indeed of MobilEye in their early robotaxi plans.
Tesla started with their Autopilot driver assist product, and has been improving it over time, hoping to make it reach the safety level needed. If they do, then they can quickly expand where it’s available to nearly everywhere. Their past approach failed, but they state they have completely re-architected their system and now it’s doing better, though it is still extremely poor in safety performance.
Some are critics of this strategy. Sterling Anderson, co-founder of Aurora and a past leader on the Tesla Autopilot team, calls it “trying to build a ladder to the moon.” Many self-driving teams rejected the idea of starting with ADAS and improving it; instead they worked to build self-driving from scratch. Tesla and MobilEye reject that it’s a ladder to the moon.
MobilEye advances a third, “zig-zag” sort of path. They began with ADAS, and have built their SuperVision product along a similar path to that of Tesla. Now, however, for their Chauffeur product, which will allow “eyes off” driving (but not “driver out” uncrewed operation) they plan to limit the circumstances they can drive it, but break above that critical safety (“precision”) level. They hope that from there, with the background they had in doing SuperVision on most roads, they can more easily grow the system into MobilEye drive, which wants to do the full driving task, with nobody in the vehicle, on most roads.
MobilEye has taken a different path from Tesla. They make AI processing chips and pair them with cameras (and sometimes radars) with their automaker customers. Unlike Tesla, they believe that radars and LIDARs can enhance the safety performance and allow independent systems to act as reinforcements and fallbacks to the vision systems which are not perfect. (Tesla believes radar and LIDAR are distractions, and that solving the problem requires so much AI that doing it with just vision will be enough. Of course, humans drive not just with vision, but with the human brain, something we have yet to replicate the full power of in silicon.) MobilEye was building its own LIDAR but has dropped that project and will work with Innoviz. They are full-speed ahead, however, on an imaging radar–a radar with increased resolution that can do many of the things a LIDAR can do, and a few it can’t, like see through all weather. Waymo’s design uses more cameras, more imaging radars and more and better LIDARs than anybody. They believe that it’s foolish to fight to save money on day one, since computer/electronics products always get cheap at scale. Their goal, which they attained, was to get the safety and driving working first.
MobilEye also has a different approach to mapping which I’ve discussed in earlier articles. Their chips are in many tens of millions of cars, so they constantly collect data on all roads every day, though they get much less data per car.
Which Path?
If moving from ADAS to self-driving is truly a ladder to the moon, then both Tesla and MobilEye are being distracted, though Tesla more so. At the same time, thinking on the futility of this has moderated with both the arrival of powerful new AI techniques, and the fairly long period it has taken teams aimed only at self-driving to get success. Several of them (Waymo, Cruise sort of, Gatik, Nuro, May, Baidu, Pony, WeRide and AutoX) have gotten some level of success–defined as providing service on public roads with passengers and no safety driver) though Waymo and Baidu seem to be at quite a higher stage of it.
At the same time, just about everybody has thrown out most of their early designs and systems and started again, nearly from scratch, and in some cases multiple times. Thinking on the role of the different sensors and what to do with their data has changed. The role of maps has greatly changed–some teams try to avoid maps in many areas other than lane-level path maps. Others started with dedicated mapping cars making highly detailed maps but everybody also drives “off-map” as is necessary in any construction area. More map data is coming from crowdsourcing at very low cost, or for cheap because any car that hopes to drive without a map is a car that can make a map as it drives, and the fleet can remember it for next time. Waymo, with its origins at Google, the world’s #1 mapping company, makes use of that expertise but is not finding mapping to be a barrier to expansion to other areas–there are many other barriers to expanding beyond just mapping for everybody.
It’s not clear if MobilEye’s Zig-Zag is the best plan. The hard truth it is results that matter today, not just plans. MobilEye has to deliver working freeway cars, and working self-driving shuttles and taxis to show they have the chops. Everybody does. It’s probably not possible to just make it work and suddenly work everywhere. There’s a lot that needs to be done to drive in a city–which means bet your customer’s lives and bet your business–than just making the software good enough. Two companies had serious (including fatal) incidents with pedestrians, and both projects were shut down and stripped for parts. The stakes are very high, but so is the reward. The “Revolution” on Shashua’s chart is real, and ideally multiple companies will reach it and compete. (It’s probably not “level 5,” though. Due to diminishing returns, the effort to complete the long tail of all roads with a certified guarantee of safety is pretty far in the future, if it ever comes.)
