Why The Tesla/Mobileye Fight Defines An Industry-Wide Schism
For starters, how we define a “fully autonomous car” is fundamentally wrong.
Mobileye and Tesla have begun trading barbs illuminating the real reason behind their split. These attacks mask an as-yet undiscussed schism in the sector that transcends their public statements.
“[Tesla’s Autopilot] is not designed to cover all possible crash situations in a safe manner,” said Amnon Shashua, Chairman and CTO of Mobileye, the Israel-based maker of collision detection and driver assistance systems. “[Tesla] was pushing the envelope in terms of safety.”
Tesla’s response? “When Tesla refused to cancel its own vision development activities and plans for deployment, Mobileye discontinued hardware support for future platforms and released public statements implying that this discontinuance was motivated by safety concerns.”
These statements highlight a distinct but unspoken truth in the burgeoning self-driving car sector. Mobileye—the company whose technology underlies the majority of ADAS (Advanced Driver Assistance Systems) and semi-autonomous driving suites on the market, may not be at the cutting edge of the technology on which they’ve built their reputation.
“They’re a failing company,” according to George Hotz, CEO of Comma.ai, who has promised to ship a $999 aftermarket unit that will make a late model Acura or Honda “on par with Tesla Autopilot.”
How is it possible Hotz could make such a claim with a straight face? Why does Mobileye’s Shashua really object to Tesla’s Autopilot? Why did Elon Musk claim a parting of the ways was inevitable?
Because Tesla/Musk and Comma/George Hotz are one side of a growing rift in the autonomous driving sector, and Mobileye—and traditional car companies—are on the other.
If you look NHTSA’s Autonomy chart you might think the path to fully driverless cars is a straight line. We’re at Level 1 now, with ADAS features like radar cruise control and Automatic Emergency Braking (AEB) approaching Level 2. Others may disagree, but I would say Tesla Autopilot is at or just beyond Level 2, meaning that it can drive itself for limited periods, but the driver must be ready to take over at any time. Level 3 will give the driver advance notice of transitions, and Level 4 is fully driverless.
Guess what? NHTSA's level system doesn’t work. Everyone in the sector knows it, but no one wants to say it. It doesn’t work because everything short of Level 4 is too vague. They can’t say it, because no company wants to be seen as falling off the self-driving car ladder.
Everyone wants to get to Level 4. Uber, so they can stop paying drivers. Tesla, so they can say they were first and somehow justify years of losses with a massive autonomous land grab. George Hotz, the founder comma.ai, so he can sell aftermarket units that migrate the world of pre-autonomy cars into the post. Most importantly, every legacy OEM on the planet must get there as soon as possible to prove their theory that Musk—and, soon, Hotz—are wrong.
That Musk and Hotz are wrong is the hinge of everything. The OEMs relying on Mobileye are orbiting ADAS, just short of Level 2, in a very deliberate holding pattern. Level 3—or Level 3 as Tesla (and Hotz) see it—is the devil. They fear that any autonomy, even as little as 60 seconds, opens the door to driver complacency, accidents and brand-wilting litigation.
Why else are the Germans hobbling their semi-autonomous driving suites in comparison to Tesla’s Autopilot?
The truth behind Mobileye, autonomy & tier 1 suppliers
There is a deeper reason. Traditional manufacturers are locked into traditional manufacturer/supplier relationships. They see all the components of ADAS—camera modules, radar, AEB, Radar Cruise Control, etc.—as building blocks on the path to Level 4. They buy these hardware modules with hard-coded software from suppliers like Mobileye and Bosch. Add enough of them, disabling any partial autonomy they see as dangerous, and eventually fully driverless cars will magically happen.
Fully or even semi-autonomous driving requires things that Mobileye and other Tier 1 suppliers can’t sell them. Things like what Tesla has, and George Hotz has baked into the Comma One. Things like crowdsourced driving data. Not merely map data, but driver behavior data. Things like machine learning and neural networks that continue to learn rather than being released, static, because they’re “good enough” until the next product cycle.
In a world of edge cases like Joshua Brown’s Autopilot accident, there is no such thing as good enough. Every time a manufacturer attacks Musk’s use of the word “Beta”, they’re indicating their own ignorance of how he’s using it. Musk doesn’t see “Beta” as unready, but as always getting better.
Everything is in beta, and car manufacturers will fail in the Autonomous Driving space if they think updates can wait months or years to update once in the real world.
The very system on which they rely, the very modules they think are stepping stones to Level 4, will never get better until manufacturers resolve problems inherent to issues of privacy, security and dealer relationships.
Commercially successful autonomy requires networking, and OTA updates/improvements to the car’s driving logic in near-real time. What Tesla calls Fleet Learning (and Hotz’s Comma One will include for $24/month) doesn’t currently exist on the menu of options from any traditional manufacturer. Even if their cars are networked, they’re firewalled and locked out of OTA upgrades by dealer franchise agreements.
If I have to go to a dealer to update my self-driving car software, I'll sell it.
Mobileye’s business model? Selling manufacturers incremental improvements to ADAS until the legal and cultural environment embraces anything further up the NHTSA chart. It’s a huge market that’s still in its infancy, and one that with little risk for Mobileye.
Working with Tesla? All downside. Too much risk being associated with a brand pushing the envelope of Level ⅔, right or wrong, and not enough units. Between the S, X, and 3, Tesla is unlikely to manufacture 400,000 in the next 24-36 months. Volvo? Daimler? BMW? VW? Tens of millions.
When hardware is commoditized, software is everything
Shashua would have us believe the OEMs are right, that Tesla has been disingenuous about Autopilot, and that Mobileye wants no part of Musk’s plan.
The truth is somewhere between they can’t...and that Musk wouldn’t share with them what they need and want, the most valuable thing Tesla has: the driving data being gathered under the Fleet Learning umbrella. Fleet Learning is the key to everything, and is why and how Tesla may commercialize Level 4 before anyone else.
Why should Tesla share data with Mobileye? That would be like anyone using Intel-based servers having to give Intel access to everything on them.
Hotz, who famously rejected a job offer from Musk to develop a Mobileye replacement, puts it all in perspective. “Mobileye makes a very low-power chip that has a lot of hand-coded rules about driving, and it really isn’t very good, because it’s not designed to be a self-driving car at all.”
What does Hotz think of Mobileye’s plan? “Mobileye wants to work with regulators to lower the safety rating of cars that don’t use their chip.”
That’s sounds like a great plan, but—contrary to most coverage in the media—it’s not going to deliver us Self-Driving Cars (even semi-autonomous ones) any sooner than Tesla will, and almost certainly won’t.
When hardware is commoditized, software is everything. Love or hate Musk, or Hotz, or any of the faces leading the new companies into the Autonomous future, hardware companies need to recogonize their place in the coming Autonomy wars. The two sides need each other, but he who has the best software will get to Level 4 first.
Much more on Hotz, Mobileye, fleet learning and Comma next week...