Technology & Innovation

The information driving driverless cars

June 28, 2016

Global

June 28, 2016

Global
Our Editors

The Economist Intelligence Unit

_____________________

Autonomous vehicles rely on the ability to integrate and interpret data at great speed

Sponsored By

The arrival of driverless cars on our streets and motorways is now all but inevitable. Precise predictions vary, but industry bosses expect fully autonomous cars to hit the road sometime between 2020 and 2025.

The speed with which driverless cars have gone from fantasy to imminent reality reflects rapid advances in the ability of computers to collect, integrate and interpret information. The environment in which cars operate is extraordinarily busy—not only do they have to navigate static obstacles such as roads and buildings but they must also avoid the traffic around them, which may include erratic human drivers. Driverless cars need a detailed understanding of what’s going on around them, and the ability to interpret it and respond instantly. The volume and variety of these data are truly astounding, and the speed at which they must be processed only compounds the challenge for carmakers.

The key data source for driverless cars is Lidar (an acronym for light detection and ranging), which allows driverless cars to perceive their surroundings. Lidar creates a high-resolution, 3D picture of the vehicle’s immediate environment to a distance of 200 feet (62 meters). Carmaker Ford says its Lidar sensors capture 2.5m 3D points per second. But to create a full understanding of the environment, Lidar data are integrated with a bewildering variety of other inputs.

Cameras capturing 20–60 megabytes of data per second add visual information, such as colour and shape. GPS co-ordinates are combined with data from tachometers, altimeters and gyroscopes to build a precise understanding of the vehicle’s position. Radar and sonar complete the picture, detecting nearby objects, their location, speed and direction of travel. All of this input requires a driverless car’s on-board computers to deal with as much as 1 gigabyte of data per second.

Much of the data are just noise, however, and do not contain any useful information. The vital data are there, but insights must be drawn out. Just as humans avoid information overload by learning what sensory input matters and what can be ignored, car computers need to sort the signal from the noise to yield the insights they need to drive, and to drive safely. It would not be possible, however, for engineers to determine in advance whether or not a driverless car should take heed of any pattern of sensor input that it may encounter—there would simply be too many rules to write. Machine learning—techniques that allow computers to improve their ability to identify patterns over time, based on experience—is therefore central to driverless automobiles.

Web giant Google has been running prototype driverless cars around California for years now, clocking up nearly half a million autonomous miles just in the period from September 2014 to November 2015. It also trains its driverless car software in simulated environments, adding 3m virtual miles to its memory banks every day. With each mile the software learns how to better predict the behaviour of the environment and how to behave safely in response.

These data do not remain siloed within the car itself. Some of them flow to remote computers for crunching and are used to update the highly detailed virtual road maps that driverless cars use to navigate their environment to ensure they are up to date. If many cars see the same newly installed sign, for instance, the collective map is updated.

One challenge that driverless car companies face is processing the vast quantities of data they collect so that they can be “understood” by the automated driving system. At Google, the mapping team uses a combination of human and machine intelligence to categorise the data collected from on-board sensors into important features such as driveways, fire hydrants and intersections.

The technologists behind the driverless car revolution have made great strides in developing and implementing this complex, distributed system of automated intelligence. The next, equally daunting task is to make it reliable and secure enough to support nothing less than the daily transit of millions, if not billions, of human beings.

For governments, regulators and the general public to welcome driverless cars, the technology needs to prove that it meets the highest standards of dependability. On the road, a software crash is a potentially lethal hardware crash, too. The critical computing challenge lies thus in the balance between understanding which data to discard and which to keep while also making sure that systems are up and running at all times.

It’s not an insurmountable challenge. After all, planes mostly fly themselves, operating layer upon layer of highly resilient technology to ensure their computers can be relied on at all times. But the environment in which cars operate is an order of magnitude more complex and crowded than the air, and carmakers have an unenviable information management task ahead of them. 

Enjoy in-depth insights and expert analysis - subscribe to our Perspectives newsletter, delivered every week