At CES, BMW, Intel, and Mobileye announced plans to bring a fleet of 40 self-driving test vehicles to the road by the end of 2017. The name-recognition hierarchy among these three is obvious: Everyone knows BMW and what it does (it makes expensive cars). Most people know Intel and sort of get what it does (it makes computer chips and might still have those blue aliens as a mascot?). Then there is Mobileye, a developer of self-driving car sensor software and a comparative unknown — but when it comes to charting the future of autonomous driving, few companies are more vital.
Ignore for a moment the hype around driverless Ubers, futuristic Google pods, and Autopilot-enabled Teslas. Consider that old-fashioned automakers sold a record 17.5 million vehicles in 2016 in the United States (OK, that figure does include 40,000 Teslas). Many of these cars are equipped with the building blocks for autonomous technology through features such as adaptive cruise control and automatic lane centering. And many of these semi-autonomous cars are powered by Mobileye’s software, which uses single-lens cameras to see and interpret road obstacles. The company has partnerships with 27 automakers and claims that about 15 million vehicles on the road today make use of its technology, which is housed in a single computer chip about the size of a Starburst.
As these automakers move toward building fully autonomous cars in the coming years, many will turn to Mobileye for its expertise. Founded in Israel in 1999 by a professor of computer science and a businessman (Amnon Shashua and Ziv Aviram, respectively), the company long ago made a bet that using cameras and computer vision would be a more effective way to steer cars than competing solutions such as radar and lidar. Using Mobileye’s system, a car driving down a busy city street can identify other vehicles, pedestrians, lane markings, and traffic signs in real time from the high-resolution images it captures. “Radars and lidars are good at seeing objects, but they’re not good at seeing what we call texture,” says Mobileye chief communications officer Dan Galves. “They can’t read traffic signs, they can’t see color. So because the camera both has more resolution and also has this really versatile ability, it makes sense that it’s the primary sensor.”
But there are still several obstacles standing in the way of a safe and reliable camera-equipped self-driving car. First, cars will require multiple cameras to gain full awareness of their surroundings, and transforming all the imagery captured into useful data requires a huge amount of computing power. The next version of Mobileye’s computer vision chip, known as EyeQ, will be able to perform 2.5 trillion operations per second. That’s fast enough to achieve Level 3 automation, or the ability for a car to drive itself with only occasional human intervention, according to the company. The following Mobileye chip, due in 2020, is supposed to be able to handle full autonomy.
Second, a driverless car needs to have a comprehensive map of the world around it to interpret as it navigates the environment. Many experts believe this will limit the initial reach of driverless cars to specific geo-fenced areas that companies have already mapped out carefully via test rides. Mobileye believes it can build a larger map faster by leveraging the tech in today’s cars. All those semi-autonomous vehicles using Mobileye-powered cameras for lane centering and automatic emergency braking can send the images they capture to the cloud, which the company then uses to build a map for its autonomous driving system. It’s a huge data advantage that will be difficult for tech firms relying on test fleets to match. “The bigger you are as a car manufacturer, the bigger advantage you have because you have millions of cars on the road that are all sending the data to the cloud, and this map gets updated,” Shashua told the Financial Times last year. “This gives the car manufacturers a way to compete with Silicon Valley, because what Silicon Valley [companies] don’t have is the cars.”
Third, driverless vehicles must be able to maneuver on the road in a way that makes sense to human drivers, which isn’t the same thing as following the letter of the law. These cars have to be safe, sure, but to be truly viable, they also have to be fast. “Right now, the way autonomous cars are being developed is to be very conservative,” Galves says. “That’s fine if there’s like two or three autonomous cars running around, but if you have a city full of thousands of autonomous cars, and they’re all driving conservative, you’re just going to clog the roads.”
Teaching a car to follow a base set of rules is now easily achievable, which is why driverless cars perform great in controlled environments. But teaching a vehicle “negotiation skills” with hundreds of other fast-moving, unpredictable drivers is the Achilles’ heel of the industry, according to Shashua. A double-merging highway, where cars are simultaneously trying to shift left and right across lanes, is a big challenge for autonomous systems. Mobileye addresses the issue via reinforcement learning, a form of deep learning in which the driverless car simulates real-world scenarios and is rewarded when it performs correctly (Google also heavily relies on driving simulations).
If Mobileye can solve these three problems, it’s well-positioned to take on a dominant role in the driverless future. But with its business residing at the intersection of two colliding cutthroat industries (tech and automotive), things probably won’t be so simple.
As CES was winding down over the weekend, the North American International Auto Show was just beginning. The biggest announcement: Waymo, the new driverless car company established by Google parent company Alphabet, is bringing a test fleet of autonomous Chrysler Pacifica minivans to roads in Mountain View, California, and Phoenix later this month.
Though Waymo is barely a month old, Google’s interest in self-driving vehicles dates back to last decade. It was one of the first major corporations to invest heavily in driverless cars but has struggled to find a business model to commercialize tech. One avenue it’s not pursuing is building cars itself. “We’re not in the business of making better cars, we’re in the business of making better drivers,” Waymo CEO John Krafcik said earlier this year. “We’re a self-driving technology company.”
That leaves two primary avenues for making money: launching a ridesharing service, which is rumored, or licensing its self-driving tech to automakers. Either solution likely won’t involve Mobileye and could endanger the Israeli company’s business. Waymo’s self-driving system relies heavily on lidar, which performs better than cameras in the dark and, while expensive, is rapidly dropping in price. (Mobileye’s tech can incorporate radar and lidar as backup data but is centered on camera-based computer vision.)
Competition is healthy for an emerging market, of course, but there’s a chance that the industry’s refusal to settle on a standard tech could stall the rollout of driverless cars, or cause one approach to be demonized when an accident inevitably occurs. After a Tesla driver died while his vehicle was in Autopilot mode, Mobileye ended its partnership with the automaker. “We felt like the Autopilot system should have more safeguards to force hands on wheel,” Galves says. “For us to participate in something we felt like could lead to accidents that could create the perception that these cars were not safe, and then that would be associated with us, was very risky.”
Tesla has accused Mobileye of using safety concerns as a cover for what was really a business decision. According to Tesla, Mobileye cut ties after finding out the automaker was building its own computer vision technology and was unwilling to pay a higher price for Mobileye’s product. The fallout between the two companies has been messy, but the end result is that achieving the driverless revolution at scale becomes more challenging for both of them.
The field will only grow more crowded from here. Though Mobileye has signed on some big names, others like Ford are using acquisitions to try to develop self-driving technology themselves. The industry settling on any standard is likely a long way off, especially since tech companies are also actively discouraging governments from strictly regulating self-driving cars. Every car manufacturer and Silicon Valley giant is free to chart its own path. But know that behind big-name automakers, there might just be a scrappy AI company like Mobileye.