notes on Chinese autonomous driving
i read a lot of Chinese AV news that doesn't make it to English Twitter. my Mandarin is good enough for Zhihu threads and WeChat articles (thanks mom), and i think there's a real information gap. so here are some notes.
the infrastructure approach
the biggest difference between Chinese and Western autonomous driving isn't the AI. it's the road.
Tesla's approach, and most of the Western AV industry, is pure perception. the car sees the world through cameras, builds an internal representation, acts on it. the road is dumb. the car is smart.
China is building smart roads.
the government's "new infrastructure" (新基建) initiative, launched around 2020, includes a major V2X component. roadside units at intersections that broadcast signal phase and timing. HD maps updated centrally. C-V2X (cellular vehicle-to-everything) communication over 5G. the car doesn't have to figure out everything from pixels. the road tells it things.
this isn't theoretical. in Wuhan, Baidu's Apollo Go robotaxi fleet covers a designated area of over 3,000 km², and the city has smart infrastructure installed at hundreds of intersections. the cars still have their own perception (lidar + cameras), but they also get signal timing, construction zone updates, and incident alerts from the road itself.
who's doing what
- Baidu Apollo, the biggest deployment. Apollo Go robotaxis are operating in Wuhan, Beijing, Shanghai, Shenzhen, Chongqing, and expanding. Wuhan is the flagship: fully driverless commercial service in designated zones. no safety driver. they also license the Apollo platform to OEMs.
- Pony.ai: robotaxi licenses in Beijing and Guangzhou, Toyota partnership, went public on NASDAQ late 2024. smaller fleet but strong technology.
- Huawei ADS, the one that gets underestimated. their Advanced Driving System ships in AITO vehicles (Huawei's partnership with Seres/Changan) and increasingly other brands. ADS 4.0 moved toward a more camera-centric approach, closer to Tesla's vision-first philosophy than most Chinese competitors.
- Momenta: partners with SAIC, BYD, and others for production ADAS that ships in actual cars at scale. less flashy than robotaxi pilots, arguably harder.
- WeRide: Guangzhou-based, running robotaxis and robobuses, also expanding into the Middle East and Southeast Asia.
the regulatory difference
in the US, AV regulation is a patchwork. federal guidelines exist but aren't binding. states make their own rules. California, Arizona, Texas, all different frameworks. companies navigate this by picking friendly jurisdictions.
China's approach is more centralized. the national government sets standards and designates testing zones. cities compete to be AV-friendly, offering streamlined approvals, dedicated test areas, infrastructure investment. when a city like Wuhan approves a company for driverless commercial service, it comes with specific conditions and a clear expansion path.
the part that surprises people: Wuhan allowed fully driverless commercial robotaxi service before San Francisco did. China's regulatory environment for AV deployment is, in many cases, more permissive than the US. that's not the story most English-language media tells.
what i think the West is missing
two things.
first: infrastructure investment changes the problem. V2X doesn't make autonomous driving easy, but it removes some of the hardest edge cases. a car that knows the signal timing can plan a left turn differently than one that infers it by watching the lights. we work hard on those exact problems at Tesla. in Wuhan, the road just tells you.
second: iteration speed. Chinese AV companies deploy and collect data faster than most people realize from the outside. Baidu alone has logged over 20 million cumulative robotaxi rides, doing 300k+ per week now. that's a lot of corner cases captured.
the counterargument is real: infrastructure-dependent systems don't generalize. you can't put RSUs on every road on earth. Tesla's pure-vision approach, if it works everywhere, scales to any road without infrastructure investment. that's a compelling bet.
but right now, in early 2026, the infrastructure approach is producing operational driverless services in major cities. the pure-vision approach is further along than it's ever been but still mostly supervised. both are working. neither is done.
i think the mistake is picking a side. the interesting question isn't which approach wins. it's what happens when they start borrowing from each other.