Google self-driving car prototype

What it was

The Google self-driving car was a modified production car fitted with sensors and software that let it handle ordinary traffic on its own. The New York Times described one version as a Toyota Prius with a funnel-like cylinder mounted on the roof^[2] — the housing for the equipment that lets the car see the road. Google said its automated cars used video cameras, radar sensors and a laser range finder to see other traffic^[1].

The three kinds of sensor cover each other's weaknesses. Cameras record shape and appearance; radar measures distance and closing speed to objects around the car; a laser range finder measures distance by timing how long a pulse of light takes to travel out and bounce back. Combining readings that fail in different ways is more robust than trusting any single one.

Mapping before driving

The cars did not work out a route from scratch. Google said its cars used detailed maps collected in advance using manually driven vehicles^[1]. Every test began the same way: a conventionally driven car was sent out first to map the route and road conditions, lane markers and traffic signs included, so the software would become familiar with the environment in advance^[1]. The enormous amounts of information the cars gathered when mapping terrain were processed in Google's data centers^[1].

This makes the driving problem much smaller than it first appears. If the lanes, signs and turns are already recorded, the software's job on the day is mainly to compare what the sensors report against what the map predicted, and to steer accordingly. The hard work is moved earlier, into the mapping drive, where a person is doing the driving anyway.

Speed limits, personality and a voice

The New York Times reported that the car had been programmed with the speed limit for every road in its database^[2]. It could also be set to different driving personalities, from cautious — more likely to yield — to aggressive, more likely to go first^[2]. Those settings are worth noticing: how a car behaves in traffic is not a fixed property of the machine but a choice made by the people who wrote the software, and it can be tuned.

The car also spoke to the person at the wheel, announcing an approaching crosswalk or a turn ahead in a pleasant female voice^[2]. Sound is a cheap channel for keeping a human informed without asking them to look away from the road.

Two trained people in every car

Google said the cars were never unmanned. A trained safety driver sat behind the wheel, and a trained software operator sat in the passenger seat monitoring the software^[1]. The division of labour matters: one person watches the road and is ready to take over, the other watches the code. Neither is asked to do both at once, which is exactly the situation that makes human supervision of automation unreliable.

Where it sits

The prototype is best read as a public account of a method rather than a product on sale. Its ingredients — several kinds of sensor, maps built first by manually driven cars, processing of the collected data in data centers, and a human safety net in the vehicle — were described plainly enough that readers could see how the trick was done: map first, drive second, keep people in the loop, and let the car say out loud what it is about to do.