Robotaxis need to be tested in real traffic - FT中文网
登录×
电子邮件/用户名
密码
记住我
请输入邮箱和密码进行绑定操作:
请输入手机号码,通过短信验证(目前仅支持中国大陆地区的手机号):
请您阅读我们的用户注册协议隐私权保护政策,点击下方按钮即视为您接受。
自动驾驶

Robotaxis need to be tested in real traffic

To achieve safe, cost-effective autonomy we need to see how other road users react to the vehicles
00:00

{"text":[[{"start":5.45,"text":"The writer is co-founder and CTO of Chinese autonomous vehicle company Pony.ai "}],[{"start":11.7,"text":"At this year’s Beijing auto show, robotaxis moved from the margins to the centre for the first time. Automakers, mobility platforms and technology groups all presented their versions of autonomous ride-hailing. The show signalled that robotaxis are no longer an experiment conducted by a handful of specialists but the next phase of how we travel."}],[{"start":32.2,"text":"Still, a crowded race is not the same as a mature one. Launching a robotaxi is becoming easier but building one that can scale commercially and operate safely remains difficult. "}],[{"start":43.35,"text":"For much of the past decade, the autonomous-driving industry has behaved as if the decisive breakthrough would come from a lab: larger models, more training data, more simulation and more compute. That view is wrong. A robotaxi’s own reactions will change the behaviour of other road users in ways that historical data cannot always predict. "}],[{"start":67.1,"text":"This distinction matters because robotaxis are not judged by whether they perform well most of the time but whether they can operate safely in complex environments where the hardest problems are ambiguous and socially negotiated: a cyclist drifting between lanes; a scooter cutting across a pick-up point; a driver edging into a gap without quite committing. "}],[{"start":88.05,"text":"Merely collecting more human driving data is not enough. Yes, it can teach machines how people respond to people. But if an autonomous vehicle behaves differently, the surrounding traffic will respond to that. This feedback loop cannot be fully inferred from historical data. It has to be observed in real operations."}],[{"start":109.69999999999999,"text":"This is not only a technical challenge but a regulatory and operational one. In parts of China and the US, cities have allowed operators to move from testing to paid driverless services in defined areas, giving them more exposure to dense, mixed traffic. In the UK and Europe, however, regulatory approval has been slower. As a result, companies are accumulating real-world, driverless data at different rates."}],[{"start":134,"text":"This is where world models matter — not as a substitute for roads, but as a way to turn data collected by driverless vehicles into repeatable training and testing. It is a system for understanding cause and effect. If the robotaxi slows, will the scooter behind pass? If it behaves cautiously at a junction, does that create confusion? "}],[{"start":154.75,"text":"The next stage is to make the learning loop self-directed by creating a system that knows what training is needed. The most difficult cases are often mundane. They are ordinary moments of hesitation, negotiation and misread intention. "}],[{"start":169.85,"text":"Finally, the robotaxi industry’s cost curve is as important as its model architecture. If driverless vehicles remain too expensive to deploy widely, companies will not generate enough interactions to improve. This is an operations race disguised as a software race. Compute, talent and data all matter. But they do not replace live fleets. "}],[{"start":190.5,"text":"Regulators and passengers will not accept a business model that treats rare failures as statistical noise. To attain the gold-standard autonomy known in the industry as “Level 4” a robotaxi must maintain core driving functions even after an unexpected hardware or software failure, and execute a safe pullover if required. "}],[{"start":209.05,"text":"The commercial challenge is equally as severe. Robotaxis must eventually compete with human-driven ride-hailing services, public transport and private cars. If they only work in limited zones, under narrow conditions or with high remote support then they will not be economically scalable."}],[{"start":226.3,"text":"That is why the next three years will be decisive for the industry. As more companies set out their ambitions, the market will become noisier. But noise should not be confused with progress. It is on the road, not in the lab, that the autonomous driving race will be won."}],[{"start":247.55,"text":""}]],"url":"https://audio.ftcn.net.cn/album/a_1779703030_9962.mp3"}

版权声明:本文版权归FT中文网所有,未经允许任何单位或个人不得转载,复制或以任何其他方式使用本文全部或部分,侵权必究。

别再抱怨了,老实交税吧

福鲁哈尔:如果美国共和体制要延续下去,富有人群对避税行为的迷恋必须结束。

债券市场同样无法避免科技巨头的影响

即便在远离AI主战场的市场,债券投资者也越来越发现,他们无法躲开硅谷的宏伟计划。

特朗普退税红利消退,美国消费者面临开支收紧压力

随着霍尔木兹海峡遭封锁、燃油价格被推高50%,各方压力与日俱增。

伊朗战争恐使美国债务利息支出增加数十亿美元

在这场冲突持续三个月后,美国政府借贷成本已升至自2007年以来的最高水平。

美国石油生产商增产以从能源危机中获利

在全球供应紧缩导致油价飙升40%、拖累美国总统支持率之后,钻探业务随之扩张。

瑞典大臣批法国尼古丁袋禁令:“对瑞典生活方式的攻击”

法国对这种香烟替代品施加最高五年监禁的“荒谬”重罚,令斯德哥尔摩大为光火。
设置字号×
最小
较小
默认
较大
最大
分享×