Breathless media coverage of when and how AVs will be deployed has largely ignored the reality that AVs can only drive on roads that have been mapped, mostly in cities.
Why it matters: If AVs were deployed today, they would be unable to navigate millions of miles of U.S. roads that are unmarked, unlit or unpaved, and the technology needed to do so is still nascent.
What's happening: Companies like Waymo test their AV fleets in big cities where they’ve spent thousands of hours labeling the exact 3D positions of curbs and lane dividers.
- Rural roads, meanwhile, have a diversity of surfaces that make them more complicated to map, and they also get less traffic, so there’s less incentive for companies to devote resources to mapping them.
What's needed: AV's current navigation capabilities could be enhanced if two types of solutions were adopted: automated maps and perception-based technologies.
Generating maps automatically involves taking aerial images and geographic data, and then converting them to machine-readable 3D maps.
- That’s the idea behind companies like Mapper, DeepMap and lvl5.
- This strategy still has limits, like the fact that it usually requires hiring people to physically drive around gathering images of roads.
Perception-based technology and systems would enable an AV to navigate without highly detailed maps, relying instead on cues like clear road markings.
- Google developed a system that can navigate a city without maps simply by using Street View images and learning the locations of landmarks.
- MIT created a framework that combines basic GPS data with onboard sensors, enabling its cars to drive themselves on unpaved country roads without a 3D map.
- These methods can be less reliable, though, when there’s low lighting.
The bottom line: With or without maps, AVs still can’t function in many scenarios, including unmapped roads. Additional advances in sensor technology, mapping, algorithms for perception and more will move AVs closer to full autonomy anywhere.
Daniela Rus is director of MIT’s Computer Science and Artificial Intelligence Lab and a professor of electrical engineering and computer science.
