D

When driving,Clara-Marina Martinez takes down any unusual behaviour she sees on the road.She then feeds these into machine-learning algorithms(算法),a form of Al,which are intended to produce a system reliable enough for a car to drive itself without any disturbance from the driver,and cope with all situations on the road.But this proving hard to achieve, carmakers choose to scale back many attempts to do so.Last year,for instance,Uber,a ride- hailing service,sold off its unit developing self-driving cars at a low price.

Autonomous vehicles are boasted as being not just convenient but potentially safer. However,just as people take time to learn how to drive safely,machines are no exception.The RAND Corporation,an American think-tank,calculates that to develop a system 20%safer than a human driver,a fleet of100 self-driving cars would have to operate 24 hours a day,365 days a year, and cover 14 billion kilometers.At average road speeds,that would take about 400 years.

Carmakers such as Porsche therefore accelerate the development process using simulators (模拟器).These teach software about danger only rarely encountered in reality.Dr Martinez and her colleagues employ"game engines",the programs that generate photo realistic images in computer games,to do this.These are used to create virtual worlds through which the software can drive.

How quickly,if ever,all this will translate into reality remains to be seen.Both regulators and customers will need to overcome doubt that a software driver really can be safer than human. From Porsche's point of view,though,there is one other relevant question.Given that much of the reason for owning a sports car is for owners to show off what they regard as their driving skills,just how big a market will there be for a version where software takes those boasting rights away?