The Navigator: Mapping the Way to Safe Automated Driving

Automated Driving System developers should take heed of what HD mapping can bring to the party.

March 1, 2021

Sam Abuelsamid

Professor Amnon Shashua, Mobileye CEO and Intel senior vice president, speaking at CES 2020. (Walden Kirsch/Intel Corporation)

When it comes to automated driving systems (ADS), Tesla CEO Elon Musk often runs counter to the consensus of almost everyone else in the sector. His disdain for lidar sensing is well-known, but during Tesla’s April 2019 Autonomy Day, Musk said “high-precision GPS maps are a really bad idea.” Almost every other AV developer is on the other side of that idea – including sensor-systems powerhouse Mobileye – and makes extensive use of high-definition maps.

During his now-annual keynote at the CES conference, Mobileye CEO Amnon Shashua (top) discussed several aspects of the SAE Level 4 automated driving system the company is developing, but one section was devoted exclusively to mapping. In December 2020, Mobileye expanded its automated-driving development program to Detroit, with Tokyo, Shanghai and possibly New York being added this year.

For most ADS developers, expanding operations into a new area typically starts with spending four to six weeks just building a high-definition map of the area. Shashua claimed that Mobileye can begin testing almost immediately because of its crowd-sourced AV maps. Note that Shashua refers to them as “AV maps” rather than simply HD maps.

The typical approach to building HD maps is for the automated-development team to drive the roads in the area multiple times, capturing all of the raw sensor data. From this, the map is built with information about the road configuration, drivable paths, speed limits, traffic signals and static landmarks that can be used for localization. This is the physical map.

Since 2018, vehicles from BMW, Volkswagen, Nissan and Nio equipped with Mobileye’s EyeQ4-based driver assistance have been collecting data and feeding it back to the Road Experience Management (REM) platform, where it is aggregated into these maps. More than one million vehicles have now built up more than 700 million km (435 million miles) of physical maps with 8 million more km (5 million miles) being updated daily.

But what sets Mobileye apart from most other AV companies – and allows for quick startup – is the semantic map. Vehicles contributing to REM are providing data on the actual path driven by vehicles on the road, typical speeds and actual stop and yield points. This is important to enable an AV to fit in seamlessly with driving patterns in a particular region.

For example, a vehicle may stop at the designated line before making a turn and then ease forward to get a clearer view of cross traffic before executing the turn. Capturing the position where other vehicles actually stop and replicating that with the AV allows the development teams to quickly start testing as they move from Jerusalem to Detroit to Shanghai.

It’s never that simple, of course. As recent winter weather in Michigan has demonstrated, systems also need understanding of road conditions and other variables to safely manage driving behavior. All of that is doable. Building this semantic understanding of the environment into the ADS can make these systems safer and more robust. Other ADS entities need to think about this as they build their own map systems. Those relying on artificial intelligence to simply guess only from what cameras can see in real time probably will not succeed.