The new Mapbox Autopilot Map allows carmakers to equip their vehicles with precise, up-to-date maps that significantly expand the roads that autopilot systems can navigate. It achieves this by collecting data in higher volumes and at a higher frequency than other providers, so that the Autopilot Map covers more roads and is updated daily by streaming changes to vehicles in real time.
‘Coverage maps’ are typically used to illustrate phone network coverage and where you might expect dropped calls. When it comes to vehicle autopilot systems, vehicles also need sufficient data coverage to prevent autopilot disengagement. Instead of people asking “can you hear me?” in a dead zone, car autopilots have to ask their drivers, “can you drive me?” in a location with insufficient or outdated map data. Autopilot disengagement forces the driver to take control again, sometimes at short notice, and is at best an inconvenience and at worst a safety concern. This battle for autopilot data coverage consists of three primary fronts: road network coverage, data freshness, and map distribution methods.
Why vehicles need advanced data
Let’s first provide some context for why these maps matter. Simply put, sensors alone do not provide enough data for autonomous driving. Modern autopilot systems rely extensively on maps to tell them what in-vehicle sensors alone cannot. Highly detailed and up-to-date map data is critical to supplement in-vehicle sensors, for example when weather impairs visibility or obstructions block sensors altogether. This complimentary system is important both for advanced awareness of the road network, to anticipate tight turns and stop signs, and for autopilot systems to know where they are in context of the broader road network and the maneuvers needed to reach their destination.
Given the importance of map data, it is surprising how often existing sources struggle to meet carmakers' and drivers' needs. Here’s why: traditional "high-definition" (HD) maps rely on expensive fleets of survey vehicles to manually map road networks, which focus on major highways. This time- and resource-intensive mapping process results in slow data updates (on the order of once every one to six months). This outdated road information further exacerbates the coverage problem because carmakers are forced to disable autopilot on roads with stale data.
The Autopilot Map
The Mapbox Autopilot Map takes a more scalable approach to road data collection by enabling ordinary sensor-equipped vehicles to share changes that they detect back to Mapbox. When enough vehicles detect a change, a consensus is reached, and the change undergoes review and integration into the Autopilot Map. The update is then instantly streamed to vehicles as a bandwidth-efficient update that only includes the data that changed.
The Autopilot Map has better data coverage because everyday vehicles travel over diverse road types, not only highways or high-traffic areas. It is also more up-to-date because the same vehicles drive over roads all the time, often multiple times per day. Updates are immediately streamed to vehicles in real time as delta updates —small data patches instead of a complete download of the entire map from scratch, which saves on bandwidth costs for automakers too.
AI-enhanced accuracy and insights
The Mapbox approach relies on crowd-sourcing data, leveraging volume and frequency of data over the accuracy of a slow survey vehicle fleet. The Autopilot Map provides vehicle systems with detailed map information about physical road features such as road edges, lanes, and traffic control devices at 20 cm relative accuracy—or about the length of a smartphone.
The Autopilot Map also contains information about optimal driving paths in locations that lack physical cues such as lane markings, for example roundabouts and complex intersections. This "behavioral" layer is built using information from the aggregated movement patterns of millions of drivers and adds additional guidance for autopilot systems in complicated driving scenarios, such as performing unprotected turns. As a result, the Mapbox Autopilot Map will reduce the number of locations where autopilots get confused or simply stop working when relying on in-vehicle sensors alone.
Forward-looking flexibility
Following Mapbox design principles of flexibility and customization, the Autopilot Map is compatible with multiple data formats, like NDS or Opendrive. Carmakers can download the data in bulk or use the Autopilot tiled map service to stream the map directly to their vehicle from the Mapbox cloud.
The Mapbox Autopilot Map can also integrate into the vehicle as part of Mapbox ADAS SDK, which can feed the map directly into the vehicle's CAN bus in ADASIS format, and offers features like map matching and electronic horizon.
Build with the Mapbox Autopilot Map
Combining the capabilities of AI with data from millions of distributed vehicle sensors, the Mapbox Autopilot Map enables both higher quality and higher coverage autopilot capabilities. By equipping vehicles with precise map data that is dynamically updated to reflect the latest road conditions, the Autopilot Map helps to reduce autopilot errors and enhance driver confidence in autopilot systems across all road classes.
The Mapbox Autopilot Map is currently available to select customers for access in Western Europe and North America.
For further details and access to Mapbox Autopilot Map, please contact the Mapbox automotive team.
