Map makers hold the key to autonomous vehicles

One of the most intriguing automotive announcements at this January's Consumer Electronics Show (CES) lacked any of the usual fanfare, instead being buried in a Volvo/Polestar press release: Google is getting into the high-definition maps business with a new company called Geo Automotive. A single quote from the division head, Jorgen Behrens, summed up in the release what this potentially game changing new company will do: “Google’s new HD map is designed specifically for auto makers and provides comprehensive lane-level and localisation data that is crucial to powering the next generation of assisted and autonomous driving systems.” The maps will inform the new Volvo EX90 and related Polestar 3 electric SUVs, ultimately helping them drive autonomously for limited periods. Google will also expand its HD mapping knowledge using data extracted from the two new models, potentially adding more across the industry. “We see this as more of an ecosystem, and we hope they sign other car makers. The more join, the better quality the maps will be,” Martin Kristensson, head of product definition and partner management at Volvo, told Autocar. Cars already use on-board standard-definition maps to aid assisted driving. Models equipped with the Volkswagen Group’s Travel Assist system can automatically slow the car ahead of a tight bend, for instance. But HD maps are much more precise, locating the car to within centimetres rather than metres and presenting it with an entire 3D picture of its surroundings to help it make better decisions. It's a crucial enabler to hands-free driving, at least according to the map companies. “The map is probably the most important sensor in the car. Everything feeds into that,” Edzard Overbeek, CEO of Dutch mapping giant Here Technologies, told Autocar. “A map looks around the corners. In automating driving, the car becomes aware of the situation and looks through buildings and round the next bend.” Here was bought in 2015 by a consortium that included Audi, BMW and Mercedes-Benz for €2.8 billion (£2.5bn at today’s exchange rate) and is arguably the leader in automotive location technology. But as money continues to pour into technology to enable hands-free driving, the leader is under threat. Geo CEO Behrens was poached by Google from Here, where he was chief production officer. Overbeek was outwardly unconcerned with the arrival of Google in its territory, however. “Google has now access to Volvo, and I don't want to downplay that, but it’s a little bit different. We have 35 million cars providing us sensor data,” he said. Not all of that is high-definition. To hoover up that data, you need a 3D lidar sensor, Overbeek said, such as that fitted to the EX90. Autonomous driving is changing the mapping business, however. The biggest players now want to provide location and mapping knowledge as part of a wider ‘stack’ complete with perception sensing and driving policy (a set of rules). For example, autonomous-angled chip-maker Nvidia bought out mapping start-up Deepmap, while lidar maker Luminar announced in January that it had bought lidar-data company Civil Maps. The tech company with perhaps the biggest bragging rights to its mapping knowledge is Intel-owned Mobileye, the chip and sensor-based autonomous specialist looking to sell a complete autonomous package to car makers. The Israeli company now reckons it has harvested 12 billion miles of “crowd-sourced” mapping from cars equipped with its EyeQ chip and forward-facing camera. In fact, 8.6 billion of those miles came in 2022, its CEO, Amnon Shashua, told a conference at CES. The car isn’t sending whole images to the Mobileye cloud but instead “sparse” data, indicating landmarks and lane markings, he said. This is cheap (about 50 cents per 20,000 miles), the CEO reckoned, but also enough to create what Mobileye calls an HD map (despite not being created via lidar) that goes on to inform its Road Experience Management (REM) database. Onto this Mobileye will then insert all the surrounding buildings and furniture not directly related to the road. “What we do is we take our REM and we hallucinate a three-dimensional world,” Shashua said. However, Mobileye’s tactic to use clients’ data to build its own mapping and knowledge has caused disquiet at one of its biggest customers, the Volkswagen Group, which at CES said it had switched to Qualcomm’s chips and driver-assistance software for some of its cars going forward precisely because it was fed up with giving away its data. The desire to grab that back inspired the German giant to create its own software division, Cariad. “That's a contract the Volkswagen brand did prior to Cariad. That's why Cariad was founded, because we are not having access,” Cariad CEO Dirk Hilgenberg said at CES. Read more: Cariad outranks VW Group brands as software takes centre stage Since then, Israeli website Autonews has reported that the Volkswagen Group has swung back to using Mobileye (perhaps having managed to renegotiate the terms of its deal for MEB-platform EVs), but one thing seems clear: who owns the data on crucial elements like mapping holds the key to autonomous driving. Are maps really that important, though? Not everyone thinks so. Tesla’s approach is much more vision-based: it thinks sensors and particularly cameras should be fine when mated to sophisticated software to allow the car to understand its surroundings to the point that it can take over. Cruise, the self-driving division of General Motors, is another that thinks the same way. “The big thing is the shift from AV [autonomous vehicle] systems that are 100% reliant on the data in the maps to one where the AV relies more on its own sensors and perception system,” Cruise head Kyle Vogt said last September on a GM earnings call. “Instead of relying purely on memory and driving with its eyes closed, it's actually paying attention.” There are seductive reasons to not rely on HD mapping. For one, mapping in high definition (versus standard definition) is “significantly more expensive” according to Remco Timmer, head of production management at Here. In a 2018 teardown of a Tesla camera-based vision system by the bank UBS, it estimated that it cost $5000 (£4200) compared with a total cost of $90,000 (£78,850) for a system based on four lidars and HD maps. The lidar system’s cost would sink to $7000 (£5820) by 2025, UBS estimated, but then the camera system would be down to $2000 (£1665). China has also complicated things. “A key reason Tesla has been reluctant to use HD maps is the limited global availability,” UBS said in report published recently. Predictably, you need regulatory approval to map China to centimetre accuracy. Lacking the software ability, Chinese car makers have gone for the HD/lidar approach, in which they’re helped by local state support as well as increasing amounts of smart infrastructure, such as roadside embedded sensors. It's easier to map motorways in HD, which is one reason why approval of level-three (hands-free) autonomous technology in Europe has been restricted to those roads. The first car to gain level-three regulatory approval, the Mercedes S-Class (in Germany and the US state of Nevada), uses Here HD mapping, as does the BMW i7, which could become the second to be approved. These aren’t exactly everyman cars, but HD maps are also key to enabling level two plus, which is eyes-on but hands-off. This will become much more widely available in new cars across Europe if regulatory approval is given. Can Google become the one that democratises this technology and pave the way for faster autonomous? Here at least is sceptical. Overbeek pointed to the near monopoly that it has on providing data to enable intelligent speed assistance across the European Union, where cars need to know the local speed limit at all times. “The tech companies don’t have that,” he said. Google’s future growth into this space depends on how fast it grows its crowdsource fleet, possibly including other customers of its Android Automotive operating system and Google Automotive Services, such as Ford or Renault. Of course Google has data from its Streetview fleet, but this is unlikely to be updated fast enough to be properly used to build HD maps usable for cars. For that you need the crowd-sourced data.