Self-driving cars need a new kind of map
Add Axios as your preferred source to
see more of our stories on Google.
Self-driving cars currently lack the common sense needed to navigate using a traditional human map. Since they can't interpret context, they need to rely on a map signal that doesn't cut out in tunnels, waver in precision or fall out of date.
The big picture: A new class of machine maps have thus become an essential element of safe and predictable vehicle autonomy. But what’s obvious to human drivers can be incredibly difficult to replicate in code, as can collecting the necessary data.
Maps for humans, like those displayed on a smartphone, rely on GPS and simple receivers that are accurate to a few meters. But they can afford to be a little imprecise — even though the map doesn't know the difference between a sidewalk and a street, it trusts that you do.
To compensate for their inability to operate safely and reliably from context alone, AVs need more of the scenes they encounter to be pre-mapped. No amount of sensor data can substitute: Maps tell an AV there’s a traffic light coming, for instance, even when a large truck blocks the view. Up-to-date maps also let AVs reroute to avoid tough situations like unprotected left turns or intersections under construction.
These machine maps must meet several key demands:
- Incredible precision, so the car can compensate for its lack of understanding context and know where it is within 10 cm.
- Granular instructions, like which lane the car is in, the traffic rules that apply to that lane, and even overhead clearances and road elevation.
- Constant connection, which continues to provide information even when GPS signals are weak or missing.
The catch: Collecting accurate 3D data of cities and keeping the information on them up-to-date have both historically been incredibly expensive and time-consuming. Even a thorough one-time map is almost useless for autonomy because cities are constantly evolving.
What to watch: Since autonomy deploys in small pockets around the globe, shrinking the cost and size of sensors (some combination of cameras, lidar and motion-capture data) to create and update these maps is crucial to meeting market needs. The challenge is making the hardware footprint lighter and scalable, while also delivering sufficient accuracy.
Neehar Garg is head of product at Mapper.ai.
Go deeper: Why machines need maps of their own
