2017: The Year In Robocars
In November, Waymo’s pilot ride-hailing service in Arizona put the safety driver in the back seat, more or less fulfilling the promise of the company’s founding spirit, Alphabet’s Sergei Brin, who’d said true self-driving cars would hit the roads by 2017.
Earlier, Audi proclaimed that its soon-to-be-released A8 would offer Level 3 autonomy, meaning that it would let the person behind the wheel read the paper or sightsee, provided that he be prepared to take back the wheel after a warning. Cadillac—more careful with its terminology—characterized its new Cruise technology as Level 2; it allows the driver to keep hands off the wheel but uses internal cameras to ensure that he pays attention to the road.
True, there are still no truly commercial ride-hailing robocar services. The technology just hasn’t been proved safe, either to the satisfaction of road-safety regulators or even to that of the car makers themselves. Case in point: Volvo, the car company whose middle name is safety, has just deferred a driverless pilot program that had been scheduled for next year in Gothenburg; the program is now on for 2021.
Why all the backing and filling? Quite probably the more ambitious claims were made by suits in the C suite and then walked back by engineers in the lab. The original big, fat promise, by Brin, was certainly in line with Silicon Valley business culture, which is to talk big and apologize later, if at all.
Silicon Valley fellows have definitely tended to patronize the metal-banging car guys in Detroit and beyond. They reasoned that if the car of the future would essentially be a rolling supercomputer, who was better placed to design it than the computer scientists?
Alas, getting a car to kinda, sorta drive itself is surprisingly easy. Engineers were doing it decades ago. Getting it to drive itself with 99.9 percent reliability is harder, but with today’s digital devices, it’s still no biggie. But remember: human beings suffer only one death for every 100 million milesdriven in the United States, which is a whole lot further than just one “nine” past the decimal point. Getting a robotic car that good or better is still fiendishly hard. And even if you could do it, proving you had is harder still.
Some of the over-promising may have riled bench scientists enough to get them to leave. In late 2016, after Elon Musk implied that Autopilot had achieved self-driving capability, the head of that program distanced himself from the implication and, two months later, resigned to found a rival company, Aurora Innovation.
Other engineers have also taken back things they’d said. As recently as 2015, Chris Urmson, then the head of Waymo, said in a TED talk that he expected that his 11-year-old son wouldn’t need to take his driver’s license at age 16 because robocars would be a reality by then. But just a year later, Urmson changed his tune, saying that cars capable of driving themselves under all conditions might not come for 30 years. Soon afterwards, he left Waymo.
That’s not to say that the major car companies (and a lot of non-car tech giants) are simply blowing smoke. They really do believe robocar technology is coming, and fast—just not as fast as they promised. And the evidence is as solid as it gets: the companies are putting their money where their mouths are.
Here’s a very partial list of acquisitions made over the past 12 months:
- Intel bought Israel’s Mobileye for US $15 billion and made it the center of its autonomous driving program. It also invested $390 million in Here, a mapping company.
- Ford poured US $1 billion into Argo and made it the center of its robocar program.
- Broadcom bid $193 billion for auto supplier Qualcomm, which had itself bid $38 billion for NXP, a major robocar supplier—thanks to its own recent acquisition of Freescale. So far, only the Freescale acquisition has actually gone through.
- Delphi acquired robocar firm NuTonomy for $450 million.
- GM Cruise—created in 2016 when GM acquired Cruise—acquired Stobe, a flash-lidar company, for an undisclosed sum.
- Lidar companies are popping up like mushrooms after a storm, fueled by VC money, hope, and the fear of missing out. Ford’s Argo subsidiary doubled down on Ford’s earlier investment in Velodyne, and it hedged its bet by also investing in Princeton Lightwave. Delphi invested in LedderTech and Innoviz. TetraVue has received VC funding from Bosch and Samsung. Luminar is backed by Peter Thiel, among others.
- Meanwhile, a lot of companies are forming robocar partnerships—Bosch and Nvidia, Audi and Nvidia, Lyft and NuTonomy, and on and on.
As all that money blows around, it is accompanied by a giant sucking sound, the sound of companies sweeping in every last engineer who has the key robocar skills. The free-for-all has driven average salaries to around $300k.
How long all this goes on depends on when one or two robocar setups become the industry standard. Come the revolution, a handful of winners will profit, a world of losers will slink away or vanish, salaries will fall back to earth—and robocars will roam the roads at last.
Intel Buys Mobileye for $15 billion
Intel is buying Mobileye, the Israeli robocar firm, for $15.3 billion. It’s one of the largest robocar acquisitions in a two-year buying frenzy that has swept both the auto industry and the tech companies that want to eat its lunch.
Mobileye made its name selling machine vision systems for driver-assistance features, such as lane keeping and emergency stopping. Unlike many companies, notably Waymo, it has so far eschewed expensive lidar, choosing instead to depend on a single (“mono”) camera. Mobileye has done work for most of the major car makers in the world; the most prominent—but by no means the largest—such relationship was with Tesla Motors, which ended with some acrimony last year.
Intel has thus bought itself not only a full suite of robocar technology but also wide-ranging contacts in the auto industry. Its newly established self-driving unit also incorporates a 15 percent stake, which Intel acquired last month, in Here, a mapping company that BMW, Daimler and Volkswagen bought from Nokia in 2015 for $2.6 billion.
Intel’s self-driving unit will be run by Mobileye management, in Jerusalem.
The deal validates Mobileye’s strategy of resisting the found-and-flip routine that most Israeli tech startups have followed (though that trend may now be changing). It held out for top dollar and it got it: Not only is $15 billion the third-largest market valuation of any publicly traded Israeli company, it is high for an auto-parts supplier and not unrespectable for an OEM.
Look at Mazda’s market capitalization, which today stands at $8.7 billion. Or at the $8 billion that Samsung paid in November for Harmon, the car-audio tech company. The only deal that comes to mind that was comparable in size was NXP’s $12 billion purchase of Freescale, an auto chip maker, back in 2015.
By comparison, Uber spent just $680 million last August to acquire Otto, a self-driving truck startup. At the time, Uber said that it valued the expertise of Otto’s staff, above all the veteran robocar expert Anthony Levandowski, a key founder of what is now Waymo.
Uber thus had to pay just 4.4 percent as much as Intel just did for what might be called a roughly comparable suite of robocar knowledge. Uber got quite a bargain—you might even say it was a steal.
Nvidia vs NXP—Whose Robocar Brain Will Win?
NXP and Nvidia have announced computing platforms for smart cars, and each company claims that its platform is the best by far.
Which company is right? Or is each right in its own way? The answer depends on how you assign a grade.
NXP today announced the S32x automotive processing platform, apparently an elaboration of the BlueBox system unveiled last year. We said then that BlueBox “knits together devices the company already sells,” and NXP appears to be doubling down on such interoperability.
“All the parts that come out will have the advantage of being compatible with this platform—this vision processor, this radar processor, this torque managment device,” said Matt Johnson, who’s in charge of NXP’s product lines, software, and automotive processors.
NXP says that eight of the top 15 car makers have adopted the S32x platform for use in future models, and this isn’t too surprising. The company, based in Eindhoven, the Netherlands, instantly became the biggest supplier of automotive chips with its 2015 acquisition of the semiconductor manufacturing company Freescale.
And it says the new platform—together with the company-wide push toward interoperability—has created a uniform software environment as well. NXP says it lets coders reuse software, thus saving 90 percent in development time on a given application, and 40 percent across different systems.
Ultimately, the NXP platform means to pull the entire car’s electronics together. That means low-power chips handling mundane things, like automatic braking, and higher-end ones that fuse streams of data from many sensors at once.
Nvidia obviously has a different take. The Pegasus platform that it announced last week aims to do one thing but do it well: drive cars. That compute-heavy task is a perfect fit for the company’s powerful graphics processors, or GPUs. These specialized parallel processors won fame by powering gaming platforms and now, Bitcoin mining operations. (Don’t laugh: Bitcoin mining accounted for 6.7 percent of Nvidia’s revenues in the second quarter.)
“There is an insatiable appetite among car makers for compute horsepower,” says Danny Shapiro, who heads the Santa Clara, Calif. company’s automotive business. “And none of the other companies are stating what their compute horsepower is. No way anyone else out there comes close to what we have.” Pegasus can do the equivalent of 320 trillion operations per second, the work of roughly 100 servers, Shapiro says.
Carmakers need it all. Yesterday’s simple driver-assistance features, like automatic braking, use just a little bit of input data and require only a little bit of silicon muscle. But an entire integrated self-driving system is fed by data from lidar, radar, ultrasound, GPS, inertial guidance, and on and on, and it must process it all and produce a decision in a fraction of a second.
You also need parallel-processing capabilities to perform deep learning, the artificial-intelligence flavor du jour. That’s the trick behind programs that recognize stop signs, respect the edges of the lane or highway, and distinguish a flickering shadow from a child running into the street.
The two companies seem to be playing tag. Both last year and this year, they made their announcements back to back, with Nvidia going first each time. Neither company’s spokesmen will deign to mention the other by name, at least not for attribution.
But NXP and Nvidia aren’t really head-to-head rivals; they’re using different measuring sticks. NXP wants to be the friend of developers and system integrators, handling the entire car, with its hundreds of apps and chips. Nvidia wants to monitor the immediate environment and frame a path through it. Both systems could easily coexist in a car, and maybe they will.
Right now, though, the two companies’ robocar platforms are works in progress. Companies that want to test NXP’s product must content themselves now with software simulations. Nvidia’s Pegasus won’t incorporate the latest GPUs until next year, so for the time being, its customers must work with Drive PX, a forerunner to Pegasus that was unveiled last year.
There are other players, of course. Intel’s purchase of Jerusalem-based Mobileye in May makes it an instant auto electronics company, but Intel’s various chip architectures weren’t necessarily designed to play well together. Tesla, now using Nvidia chips, is rumored to be working on a home-built alternative, perhaps with the aid of AMD.
But right now, if you’re looking for an electronic brain for your car and you’re not in the mood to build one yourself, NXP and Nvidia would seem to be the go-to companies.
An abridged version of this post appears in the December 2017 print issue as “The Battle for the Smart Car Brain.”
GM Starts Catching Up in Self-Driving Car Tech with $1 Billion Acquisition of Cruise Automation
It should be obvious by now that autonomy is the future of cars. Obvious or not, large automotive manufacturers have been (predictably) slow to adapt, which is why the most sophisticated autonomous car you can buy right now, the Tesla Model S, comes from a company that didn’t even exist 15 years ago.
In an attempt to brute-force itself into relevancy, General Motors (the parent company of Cadillac, Chevrolet, and GMC, among other brands) invested US $500 million in Lyft “to create an integrated network of on-demand autonomous vehicles” back in January. And last Friday, GM followed this up with the acquisition (worth an estimated $1 billion) of a small autonomous car startup based in Silicon Valley called Cruise Automation.
Over the last few years, many of the largest automakers in the world have been trying to catch up in autonomous vehicle tech with increasingly frantic amounts of money. Last year, for example, Toyota announced a $50 million investment to fund autonomous vehicle research at Stanford and MIT, and followed that up with a $1 billion commitment to its overall robotics and artificial intelligence research, with a focus on automotive safety.
Meanwhile, General Motors’ investment in Carnegie Mellon University during the 2007 DARPA Urban Challenge (and beyond) has been significant, but many of the most experienced CMU researchers went to Uber last year. With other big companies (like Google and Tesla) hiring roboticists like crazy, and most of the academic research centers already cleaned out, GM needed to do something, and that something was pay a ridiculous amount of money for three year old autonomous vehicle technology startup Cruise Automation.
This is not to suggest that Cruise Automation doesn’t know what it’s doing. The company has been developing an aftermarket autopilot system that can be adapted to a variety of different consumer cars. The system would have cost about $10,000, and included sensors (such as millimeter-wave radar, stereo video cameras, GPS, and inertial sensors) and software to enable hands-free driving on highways (but not city streets).
GM isn’t likely to encourage the idea of a kit that can be applied to non-GM vehicles (like Audis, initially), and it sounds like Cruise will be integrating their technology and expertise directly into GM cars going forward. GM is at least willing to leave them mostly alone to let them do their thing (a popular strategy with tech acquisitions lately): Cruise gets to keep its San Francisco office and operate mostly independently from GM’s internal self-driving car team.
A billion dollar valuation for a 40 person company is certainly a lot, but it’s not necessarily an indication of a bubble in the self-driving car space. A bubble would imply a trend of systematic over-valuations, which ultimately reveal themselves as the bubble bursts.
However, there are two reasons why a bubble doesn’t seem likely in this case. First, there’s very little question as to the value of a good autonomous car, and that value is enormous to almost anyone who drives. Consumers have already demonstrated that they’re willing to pay thousands of dollars per vehicle for technology packages that include basic highway autonomy, and it’s reasonable to think that they’ll pay even more (perhaps a lot more) for enhanced safety features all the way up to full autonomy. With the technology in the process of real-world validation, acquiring talent in the space is a prudent (some might say necessary) medium- to long-term investment for a major auto manufacturer looking to stay relevant.
The second reason why this isn’t a bubble is that there simply isn’t enough talent out there to generate one. Because robotics involves hardware—and sophisticated robotics involves sophisticated hardware—getting education and experience in the field is a major commitment of time and money that is very unlike the environment that led to oversaturation in the dot-com and app markets. Finding people with the right education and experience is one of the biggest problems that robotics companies (of all kinds) have right now. Cruise, for example, is trying to hire ten new engineers in AI, machine learning, perception, computer vision, and other areas, and even with $1 billion in the bank, they’ll be competing for a very limited number of qualified and experienced people against Google, Toyota, and many other companies with massive resources. By acquiring Cruise, GM has managed to bring on exactly the talent that it needs all at once, which may have otherwise been difficult or impossible to do.
These factors suggest that small autonomous car startups have a lot of real value, both in their technology and in their people. We wouldn’t be surprised to see other major acquisitions of companies with proven experience, such as nuTonomy or Zoox. However many billions of dollars it takes to get this industry moving, we can at least be pretty sure that there’s going to be a tangible and very positive result at the end. And it’s about time.
Israeli Startup Innoviz Promises $100 Solid-State Automotive Lidar by 2018