Mapping out paths for robot cars
Dozens of firms are vying to build HD landscapes that may serve as onboard navigation guides
On any given day, there could be half a dozen autonomous cars mapping the same street corner in Silicon Valley. These cars, each from a different company, are all doing the same thing: building high-definition street maps, which eventually may serve as onboard navigation guides for driverless vehicles.
These companies converge where the law and weather are welcoming — or where they can get the most attention. For example, a flock of mapping vehicles congregates every year in the vicinity of the CES technology trade show, a hot spot for self-driving advances. “There probably have been 50 companies that mapped Las Vegas simply to do a CES drive,” said Chris McNally, an analyst with Evercore ISI. “It’s such a waste of resources.”
Autonomous cars require powerful sensors to see and advanced software to think. They especially need up-to-the-minute maps of every conceivable roadway to move. Whoever owns the most detailed and expansive version of these maps that vehicles read will own an asset that could be worth billions of dollars.
Which is how you get an all-out mapping war, with dozens of contenders entering into a dizzying array of alliances and burning tens of millions of investment dollars in pursuit of a massive payoff that could be years away. Alphabet Inc.’s Google emerged years ago as the winner in consumer digital maps, which human drivers use to evade rushhour traffic or find a restaurant. Google won by blanketing the globe with its street-mapping cars and with software expertise that couldn’t be matched by navigation companies, automakers or even Apple Inc. Nobody wants to let Google win again.
The companies working on maps for autonomous vehicles are taking two different approaches. One aims to create complete high-definition maps that will let the driverless cars of the future navigate all on their own; another creates maps piece by piece, using sensors in today’s vehicles that will enable cars to gradually automate more and more parts of driving.
Alphabet is trying both approaches. A team inside Google is working on a 3-D mapping project that it may license to automakers, four people familiar with its plans confirmed to Bloomberg. This mapping service is different from the high-definition maps that Waymo, another Alphabet unit, is creating for its autonomous vehicles.
Google’s mapping project is focused on driver-assistance systems that enable cars to automate some driving features and help them see what’s ahead or around a corner. Google released an early version of this in December, called Vehicle Mapping Service, that incorporates sensor data from cars into their maps.
For now, Google is offering the service to carmakers that use Android Automotive, the company’s embedded operating system for cars. Google has named three partners for that system, but other automakers are reluctant to hand their dashboards over to the search giant. So Google is looking to expand the features on the mapping service and find other ways to distribute it, these people said.
“We’ve built a comprehensive map of the world for people and we are working to expand the utility to our maps to cars,” a Google spokeswoman said in a statement. She declined to comment on future plans.
At the same time, Waymo and the other giants with sizable driverless research arms — including General Motors Co. and Ford Motor Co. — are all sending out their own fleets to create rich, detailed, high-definition maps for use in driverless cars. There also are smaller start-ups hawking gadgets or specialized software to build these maps for automakers that find themselves further behind. Still other suppliers are working on mapping services for conventional cars with limited robotic features, such as adaptive cruise control or night vision.
These self-driving maps are far more demanding to create than older digital ones, prompting huge investments across Detroit, Silicon Valley and China. “An autonomous vehicle wants that to be as precise, accurate and up to date as possible,” said Bryan Salesky, who leads Argo AI LLC, a year-old start-up backed by a $1-billion investment by Ford. The “off-the-shelf solution doesn’t quite exist.”
The cartographic arms dealers
Making a driverless map, like making a driverless car, is a laborious task. Fleets of autonomous test cars, loaded with expensive lidar sensors and cameras, go out into the world with human backup drivers and capture their surroundings. Plotting the results helps train the next fleet, which will still have safety drivers at the wheel — and, in some cases, scores of additional humans sitting behind computer monitors to catalog all the footage.
It’s an expensive ordeal with a payoff that’s years, if not decades, away. “Even if you could drive your own vehicles around and hit every road in the world, how do you update?” asked Dan Galves, a spokesman for Mobileye. “You’d have to send these vehicles around again.”
Unlike conventional digital maps, self-driving maps require almost constant updates. The slightest variation on the road — a construction zone that pops up overnight, or a bit of debris — could stop a driverless car in its tracks. The risks of getting autonomous systems wrong were brought home last month when an Uber test car with a human monitor hit and killed a woman in Tempe, Ariz., as she was crossing the road at night. “It’s the freak thing that happens that’s going to make autonomous not work,” said McNally, the analyst.
Mobileye believes it’s more efficient and cost effective to let the cars we’re driving today see what’s ahead. In January, the Intel Corp. unit announced a “low band width” mapping effort, with its front-facing camera and chip sensor that it plans to place in 2 million cars this year. The idea is to get cars to view such things as lane markers, traffic signals and road boundaries, letting them automate some driving.
Mobileye says this will take less computing horsepower than building a comprehensive HD map of the roads; Mobileye’s Galves said the company will pair its sensor data with the maps from navigational companies and, over time, create a map that a fully driverless car could use.
That’s also the tactic of Google’s longtime mapping foes: Here and Tom Tom. These two European companies have positioned themselves as the primary alternatives to Google Maps, selling the dashboard mapping apps installed in today’s autos. Yet these “static” maps see only broad street shapes and capture snapshots in time. Now, both companies are working on replacement products: “dynamic” maps that represent lanes, curbs and everything else on the road. The hope is that car manufacturers will stick with old-guard mapmakers as vehicles move from somewhat intelligent to fully automated vehicles without steering wheels.
Here, owned by a consortium of German automakers, has a few examples on the road. Its mapping system enables limited handsfree driving for Audi AG, one of its co-owners, and plans to support safety features this year for BMW AG, another co-owner. (Intel also took a 15% stake in Here last year.)
Tesla Inc. is the car company most eagerly embracing the incremental march toward autonomous driving with its driver-assistance software, Autopilot. Tesla relies on cameras and sensors on its vehicles but has eschewed lidar. The company hasn’t disclosed what mapping service it is using for Autopilot, and a company representative declined to comment. Tesla had a nasty public split with Mobileye two years ago.
But Tesla has leaned on at least one other company, Mapbox Inc., to help assemble its maps. Tesla paid $5 million to Mapbox for a two-year licensing deal in December 2015, according to a regulatory filing. Mapbox has mostly sold its location data to apps such as Pinterest and Snapchat. Fresh off a $164-million financing round, the start-up has begun inching into automotive maps. Through its software installed on phones, Mapbox said it plots some 220 million miles of road data globally a day.
“We have more sensors on the road today than the entire connected car space will have by 2020,” said Chief Executive Eric Gundersen. Its pitch to carmakers is to use that location data as a base layer for future maps — pairing it with camera systems, such as Mobileye’s, or their own sensor data. And like other companies targeting automakers, Mapbox is happy to play neutral and work with anyone. “We don’t know who is going to win,” Gundersen said.
The new hotshot path finders
It’s not just that no one knows who will come out on top. The mapping industry doesn’t even know which strategy is best. Every selfdriving map looks different because each one depends on the sensor system of the vehicle that creates it. Therefore there isn’t a standard sensor package, said Spark Capital’s Nabeel Hyatt, an early investor in Cruise Automation, the autonomous-driving company bought by General Motors in 2016 for $581 million.
As a result, a slew of HD mapping companies are taking different stabs at the problem, each gobbling up venture capital and competing for lucrative contracts. Some of them disparage Mobileye’s approach, which relies on a seamless transition from semi-autonomous driving (what’s called Level 2 and 3) to driving without human assistance (Level 4 or 5).
“It’s very hard to climb the ladder from 2 to 3 and then to 4,” said Wei Luo, chief operating officer of DeepMap Inc. “There’s a very intense gap.” The best HD maps, Luo argues, are built with only driverless functions in mind. The startup said it’s working with Ford, Honda Motor Co. and China’s SAIC Motor Corp. (Mobileye also is working with SAIC, and Waymo is in talks with Honda.)
Waymo is in this camp too. The effort formerly known as the Google selfdriving car project started on maps in 2009, with Waymo’s Andrew Chatham and one other engineer doing the “super tedious” work of crafting them from scratch — shipping cars packed with sensors to capture a city’s surroundings then coding those 3-D images into a digital landscape. Chatham said cars may rely on perception systems alone to drive on the highway but would be helpless in other traffic conditions. Imagine pulling up to a busy, double-left-lane intersection you’ve never seen before. Now imagine a self-driving car trying to do that.
“That’s the advantage of having a detailed map,” Chatham said. “We can give the cars all the answers to the nasty questions.” He said Waymo is exploring solutions to mapping real-time factors such as construction updates, but declined to share details.
Thanks to its years of effort and artificial intelligence arsenal, Waymo is considered the leader in HD maps. But to date, the company has pitched its entire suite to prospective partners and landed few, although in March, Waymo announced a long-term development partnership with Jaguar-Land Rover that includes the purchase of 20,000 Jaguar electric SUVs to be fitted with Waymo driverless technology and used for ride-hailing services over the next several years. Chatham declined to say whether Waymo is considering selling its map as a separate product.
Plenty of newcomers are pitching carmakers on the need to catch up with frontrunners such as Waymo. Deep Map, started by veterans of Google and Apple, is banking on its intelligent software to cut down the time and cost of converting the images pulled from selfdriving car sensors into a single, high-resolution landscape. The start-up said it’s working with Ford, Honda and SAIC.
Civil Maps has technology that “fingerprints” sensor data, forming digital grids with each loop made by a mapping vehicle around the same area. It’s a bit like the way the mobile app Shazam recognizes a piece of music, said CEO Sravan Puttagunta. Ford is an investor, and Puttagunta said his company is in the process of raising additional money.
For now, most car companies are testing the waters rather than cutting massive, multimillion-dollar deals for maps. A Ford spokesman described its work with start-ups as “research.” Argo, the automaker’s selfdriving bet, has looked at a variety of suppliers but is currently relying on its own internal maps. At GM, spokesman Ray Wert said the company prefers to do its own mapping.
The new entrants know they can’t all survive. “It’s very similar to navigational maps or even the search engine,” said Deep Map’s Luo, a former Googler. “Whoever has bigger scale will have the advantage.”