{"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"}