The startup is creating road schematics accurate to within 10 centimeters.
The suburban streets of Albany, Calif., were mostly empty on a recent sunny mid-afternoon. Of the few motorists who were on the road, many were scowling at a white Subaru Outback with a spinning radar antenna bolted to its roof and decals reading “Civil Maps” plastered on its side.
The Subaru was driven by Anuj Gupta, co-founder of Civil Maps and a visiting scholar in artificial intelligence at the nearby University of California-Berkeley. He was demonstrating a mapping technology that he hopes automakers will pay big money for: software that can transform raw 3D data from LiDAR (high-resolution laser imaging), cameras, and other sensors on autonomous vehicles into a machine-readable format.
The motorists were scowling because of how slowly Gupta was driving, and how frequently he was changing lanes. Each time he did so, the lane markings displayed on an HP laptop mounted to the dashboard changed color, as did the rest of the road schematic. Gupta explained that the software not only knows where the car is to an accuracy of 10 centimeters, but it also knows which traffic signals and road markings apply to the lane that the car is in.
“We are the cognition layer for autonomous cars,” Gupta said. Civil Maps has already recorded every permanent street light, stop sign, lane marking, and other road signals in Albany and a few other cities in the Bay Area, and is now setting its sights on the rest of America. Through partnerships with automakers, the company’s software will be installed on semi-autonomous vehicles that are already available for sale—cars with features like adaptive cruise control and lane departure warnings.
Assuming the driver opts in via a process to be established by the automaker, data from the sensors on his or her vehicle will be aggregated along with thousands of others and sent back to Civil Maps HQ, where engineers will determine their accuracy and then add the results to their database.
It’s a sort of backdoor approach to creating machine-readable schematics of every mile of major roadways in the US. There are around 6 million of those miles, Gupta said, and it would be impractical for his company to drive all of them in the way that Google Street View has set out to do.
If everything goes according to plan, Civil Maps will have a valuable product to offer anyone interested in making a self-driving car: instead of sensors that constantly scan for infrastructure in all directions, the car will use the database to know exactly where to focus its lasers and cameras, ensuring that it will never miss a red light or a child crossing the street.
And it can do all that with relatively cheap solid-state LiDARs and low-power processors from Qualcomm and ARM. That approach is markedly different from the expensive sensors and GPU-accelerated processors Google, Toyota, and other large autonomous driving companies are developing.
“If you’re driving your Honda Civic and you want to use that for mapping, then you can’t use it, because it doesn’t have those computers,” Gupta said.
Of course, Civil Maps is a startup with a handful of employees, many of whom are fresh out of college. Its headquarters amount to a few rooms of a single-family house on a residential street in Albany. So it is taking a cheap approach that doesn’t involve intensive computing partly out of necessity.
Nevertheless, its concept is catching on. The Ford Motor Company was one of several investors that participated in a $6 million funding round in July, and Gupta said he is in talks with several more automakers. Like its competitors, Ford is also testing its own sensor- and GPU-laden self-driving cars on public roads in California.
If nothing else, then, Civil Maps’s early success is emblematic of the Wild West that is autonomous driving technology right now: everyone from Nvidia to Tesla to startups out of UC Berkeley has their own approach, and they’re all planning for a future in which cars can not only drive themselves, but can do it safely and reliably enough that the public—and the government—will trust them.