Autonomous vehicles like those being tested by Google, Uber, and major automakers rely on 3-D maps that record the position of curbstones and traffic lights with high accuracy. The maps are usually created by driving around in vehicles outfitted with expensive sensors.
Civil Maps wants to use consumer cars as a low-cost mapping workforce instead, taking advantage of the sensors being added to premium models for advanced cruise control and crash avoidance.
Those cheaper sensors can’t match those in a dedicated mapping vehicle. But pooling enough observations of the same stretch of road makes it possible to maintain high-quality maps and keep track of features such as speed bumps, road signs, and road markings, says Sravan Puttagunta, CEO of Civil Maps. “I think starting in 2017-’18 you’ll have a lot of cars that will meet our criteria for map contributions,” he says.
Civil Maps' software uses data from vehicle-mounted sensors to build detailed 3-D maps with the information that self-driving cars need.
Puttagunta hopes to persuade automakers to add software to their vehicles so they can contribute data from cameras, radar, and lidar, which uses laser light to map objects in 3-D. Civil Maps has developed software that combines different types of sensor data, and multiple scans of the same objects, to make and update maps. Car companies that contribute data would also be able to use the maps created, says Puttagunta.
He argues that this approach is the only economical way to get rich, frequently updated maps at a continental scale so cars can drive themselves anywhere.
Carmakers have already shown they are willing to spend money—and work together—to create the maps needed for autonomous cars. Audi, BMW, and Daimler teamed up last year to acquire Nokia’s mapping business HERE for $3 billion. Civil Maps recently received $6.6 million from investors including Ford. This week Ford pledged to have fleets of autonomous vehicles on the road within five years (see “2021 May Be the Year of the Fully Autonomous Car”).
Civil Maps relies on machine-learning software to interpret data from different sensors and combine it into maps. Its technology also uses data from cars to teach its software how an autonomous car should handle particular stretches of road. For example, Civil Maps software can learn that a particular lane is left turn only. It has figured out that some lanes on the Golden Gate Bridge change direction depending on the time of day.
Reilly Brennan, executive director of the Revs Program on the future of the car at Stanford University, says Civil Maps’ approach makes sense. Rapid improvements in the cost and quality of 3-D sensors make it a practical idea, he says.
One challenge for the crowdsourcing approach will be figuring out how to ensure good coverage everywhere, says Brennan. Routes that don’t get much traffic, or aren’t visited much by people with expensive new sensor-packed cars, would be mapped less frequently and in less detail.
Puttagunta says that problem is manageable. The cost of sensors is coming down so quickly that they won’t be limited to certain high-end vehicles for long, he says.
As rival companies gear up to put driverless cars on the road, they are racing to map the world in more detail than ever, says Hal Hodson
“The maps need to be there for the autonomous cars to be able to do what they need to do”
IT’S a 4-hour drive from Pittsburgh to Detroit – but there’s an app for that. You punch the destination into your phone and a driverless car soon swings to a stop next to you. You jump in and it whisks you north-west towards the I-80 on-ramp.
But as you merge with the highway traffic, the car pipes up: “This car runs on the Uber network, which does not cover Detroit. I cannot take you to your final destination. You will be dropped at an appropriate interchange point.”
The way things are going, this could be the short-term prospect for driverless cars. The companies chasing a future full of autonomous vehicles are each creating a closed system in which their cars will work, but their competitors’ won’t and it’s all to do with maps.
Driverless cars carry many different kinds of sensors – including cameras, lidar and radar – but they are not capable of fully understanding what they see. They may be able to steer themselves around obstacles and brake to avoid collisions, but can have trouble reading unfamiliar objects in the way humans can. For example, before an autonomous car approaches a junction, it needs to know exactly where the traffic light will be. Because of this, driverless cars need highly detailed 3D maps of the roads they are to navigate. These are not top-down charts like you get from a satnav or Google Maps, but representations of street layouts and roadside infrastructure like barriers and traffic lights – plus information about where other cars are likely to be.
Maps for driverless cars are like railways for trains, says John Ristevski at Nokia Growth Partners in Palo Alto, California.
“The map needs to be there for the autonomous car to be able to do what it needs to do.”
Companies are fighting to build their own version of such maps, using a variety of tactics. Last week, Uber hired Tesla’s head of mapping. Traditional car makers like Ford and Toyota are scrambling to take advantage of the millions of vehicles they have on the road already to harvest large volumes of data. Newcomers like Uber and Google are relying on their prowess with data science to give them an edge.
All of them have customised mapping vehicles crawling the roads of their target areas, trying to get ahead.
“A car has far more computational power than a phone and much better sensors ”
Small beginnings
Creating those maps for a relatively small built-up area like a mid-sized city is not hard. “I think San Francisco has about 2000 kilometres of major roads,” says Ristevski. “You can map that with one high resolution mapping vehicle in about two weeks.” But extending those maps across larger urban environments – and eventually whole countries – will be painstaking, expensive work.
It’s therefore no coincidence that the first driverless taxi service – announced in August – is launching in the tiny city-state of Singapore. The company behind it, a spin-off from the Massachusetts Institute of Technology called nuTonomy, mapped all of Singapore’s streets by driving around with a lidar scanner, says CEO Karl Iagnemma.
But these companies may soon hit a stumbling block. Each one is building proprietary maps that only work with the sensors in their own cars. At the moment, the driverless cars that Uber is testing in Pittsburgh cannot run on Ford’s map in Michigan, for example. The maps are incompatible, like railway networks that operate on different gauges. “It’s a patchwork,” says Ristevski.
However, Sanjay Sood at Chicago-based mapping company HERE – which was bought by a consortium of German car makers in 2015 – is not too worried. There is bound to be fragmentation early on, he says. But that will change.
“There’s going to have to be some standardisation,” says Sood. Whether that is in the format of the maps themselves or the sensors and software that drive the cars remains to be seen. “It’s super early,” he says. “The reason you’re not seeing standards is that we’re still in the research and development stages.”
The dark horse in the map wars is Tesla. Elon Musk’s electric car company currently has 140,000 cars on the road around the world. Some models have an autopilot mode – in which the car can drive itself along relatively easy stretches of road as long as a human driver is ready to take over at any moment – but none are fully autonomous. However, the cars are still fitted with sensors that are needed for the autopilot feature, and all the data they gather is beamed back to Tesla.
As the first company to put a data-gathering sensor network on thousands of public roads, Tesla could have its hands on data for far more locations than any other car company.
But Tesla’s lead might not last long. Toyota plans to include the sensors required for autonomous driving in all of its new cars in 2017. These millions of vehicles won’t be autonomous themselves, but will gather the data needed for Toyota to build its own maps. To deal with this vast amount of information, Toyota is also building a data centre in Plano, Texas.
Whoever wins, the maps on which driverless cars run are going to end up processing vast amounts of data beyond that needed for the cars to drive. They might include the location of hordes of pedestrians, roadworks, black ice and other weather hazards, for example