评论
打印
Science and technology
科技版块
Self-driving cars
自动驾驶汽车
Hands off the wheel
双手离开方向盘
Teaching autonomous cars to drive with computer games
用电脑游戏教自动驾驶汽车驾驶
WHEN DRIVING, Clara-Marina Martinez makes a note of any unusual behaviour she sees on the road.
驾驶过程中，克拉拉-玛丽娜·马丁内斯会记下她在路上看到的任何不寻常的行为。
She then feeds these into machine-learning algorithms, a form of AI, which she is helping develop for Porsche Engineering, a division of the eponymous German sports-car company.
然后，她将这些数据输入机器学习算法，这是人工智能的一种形式，她正在帮助德国保时捷的子公司开发这种算法。
Those algorithms are intended to produce a system reliable enough for a car to drive itself.
这些算法旨在产生一个足够可靠的系统，使汽车能够自动驾驶。
Such a fully autonomous car, known in the industry as Level 5, should be able to complete an entire journey without any intervention from the driver, and cope with all situations on the road.
这在业内被称为5代的全自动汽车，应该能够在没有司机任何干预的情况下完成整个旅程，并应对道路上的所有情况。
But this is proving hard to achieve, and many attempts to do so are being scaled back.
但事实证明，这很难实现，许多这样做的尝试都在缩减。
Last year, for instance, Uber, a ride-hailing service, sold off its unit developing self-driving cars.
例如，去年叫车服务公司优步卖掉了开发自动驾驶汽车的部门。
Autonomous vehicles are touted as being not just convenient but potentially safer.
自动驾驶汽车被吹嘘说不仅方便，而且可能更安全。
However, just as people take time to learn how to drive safely, so do machines.
然而，正如人们需要时间来学习如何安全驾驶一样，机器也是如此。
And machines are not as quick on the uptake.
而且机器的理解速度也没有那么快。
The RAND Corporation, an American think-tank, calculates that to develop a system 20% safer than a human driver, a fleet of 100 self-driving cars would have to operate 24 hours a day, 365 days a year, and cover 14bn kilometres.
美国智库兰德公司计算出，要开发一个比人类司机安全20%的系统，一支由100辆自动驾驶汽车组成的车队必须一年365天每天24小时运行，行驶140亿公里。
At average road speeds, that would take about 400 years.
在平均路速下，这将需要大约400年的时间。
Carmakers such as Porsche therefore accelerate the development process using simulators.
因此，保时捷等汽车制造商使用模拟器加快了开发过程。
These teach software about hazards only rarely encountered in reality.
这些模拟器教给软件的是现实中很少遇到的危险。
Dr Martinez and her colleagues employ “game engines”, the programs that generate photorealistic images in computer games, to do this.
为了做到这一点，马丁内斯博士和她的同事们使用了“游戏引擎”，即在电脑游戏中生成逼真图像程序。
These are used to create virtual worlds through which the software can drive.
这些是用来创建虚拟世界的，软件可以通过这些虚拟世界来驱动。
Objects in these virtual worlds are assigned their physical characteristics (ie, buildings are hard, people are soft) so that the sensors in vehicles, such as cameras, radar, lidar (a form of radar that uses light) and ultrasound transceivers respond in the appropriate way.
这些虚拟世界中的物体被赋予了它们的物理特征（即，建筑物是坚硬的，人是柔软的），因此车辆中的传感器，如相机、雷达、激光雷达（一种用光的雷达）和超声波收发器都会以适当的方式做出反应。
Once the software has been trained, it is tested in real autonomous vehicles by re-creating those situations on a test track.
一旦软件经过训练，它就会在真实的自动驾驶汽车上进行测试，在测试轨道上重现这些情况。
How quickly, if ever, all this will translate into reality remains to be seen.
这一切究竟会以多快的速度转化为现实，需要我们拭目以待。
Both regulators and customers will need to overcome scepticism that a software driver really can be safer than a wetware one.
监管机构和客户都需要克服疑虑，即软件驱动程序真的可以比人脑更安全。
From Porsche’s point of view, though, there is one other pertinent question.
不过，从保时捷的角度来看，还有另一个相关的问题。
Given that much of the reason for owning a sports car is for owners to show off what they perceive to be their driving skills, just how big a market will there be for a version where software takes those bragging rights away?
鉴于买跑车的主要原因是车主炫耀他们的驾驶技术，如果一个软件剥夺了这些吹嘘、炫耀的权利，它的市场还会有多大呢？
译文由可可原创，仅供学习交流使用，未经许可请勿转载。
