Podcast: China's autonomous vehicles and robotaxis
A deep dive into China's smart driving and AV industry, including government policies, key players, differing strategies, and shifting geopolitics
Watch or listen to the High Capacity podcast on:
In this episode, I speak with Lei Xing, former editor-in-chief of China Automotive News and co-host of the China EVs & More podcast with previous guest Tu Le. Lei Xing is a world expert on China’s EV and autonomous driving space.
Key takeaways:
China’s AV industry is supported by central and local government policy. Different Chinese cities have taken different approaches. Wuhan is more proactive in expanding robotaxis while Beijing is more conservative.
China’s smart driving players are taking different approaches. Some are moving up the ladder from “driver assistance” systems to L4 autonomy. Some are starting with robotaxis and then moving into private passenger cars. The LiDAR vs. camera-only debate is ongoing.
Many Chinese companies are taking a full-stack approach to autonomy. They’re developing their own semiconductor chips and their own AI models.
But the industry is also taking a mix-and-match approach. Many Chinese EV companies are using Nvidia chips and partnering with US mobility companies like Uber and Lyft. Western automakers are also relying on Chinese smart driving components.
China is moving fast on autonomous delivery and commercial trucking. Autonomous delivery vehicles have become commonplace in parts of China.
Follow Lei Xing:
Transcript
Kyle (00:00)
Welcome to the High Capacity Podcast. I’m your host, Kyle Chan, a fellow at Brookings. I’m thrilled to be joined today by my guest, Lei Xing, a world-class expert on Chinese electric vehicles and autonomous driving. He’s the former editor-in-chief of China Automotive Review and a co-host of the podcast China EVs and More with Tu Le, who was previously on the show. In that episode with Tu Le, we looked at China’s overall EV landscape. This time with Lei Xing, we’ll zero in on China’s push into autonomous vehicles and robotaxis. We’ll look at how China’s autonomous driving technology is evolving, how the Chinese government is supporting the industry, and how Chinese companies are scaling up robotaxi fleets around the world. Welcome, Lei Xing, and thank you for coming on the show.
Lei Xing (00:52)
Thank you for having me, Kyle. It’s nice to meet you through this platform and talk about AVs.
Kyle (01:02)
Absolutely. It’s a hot topic.
To begin, I’d like to focus on making sense of the technology. There are a lot of terms thrown around—smart driving, autonomous driving, ADAS, robotaxis, different levels of autonomy: L2, L2+, L4. For those who don’t follow the industry as closely as you do, what do these terms mean, and how do they differ in China?
Lei Xing (01:40)
This question alone could take hours, because the terminology is part of the narrative. “Smart driving” is definitely a Chinese term. In Chinese it’s zhijia (智驾). In China, “smart driving” usually points to ADAS or L2++—driver assistance—meaning you have to pay attention at all times, and the driver is still responsible in case of accidents.
“Smart driving” also makes it more appealing as a marketing term. Zhijia sounds better than “advanced driver assistance system,” which is jargon. But in practice, a lot of it is Level 2—sometimes called Level 2+, Level 2++—and “point to point,” which is where Tesla’s FSD is today. And Tesla still says in the fine print that you have to watch the road and pay attention.
The major difference comes when you go to L3 and L4, when responsibility shifts toward the OEM or the company—like a Waymo. In China, most recently, two vehicle models were approved to pilot Level 3 operation. Put simply, Level 3 is when you can take your eyes off. You can do other things. And that’s when, if the feature is commercialized, responsibility transitions to the OEM.
Level 4 used to mean “driverless,” like Waymo, Cruise, Baidu, Pony.ai, WeRide—we’ll talk about those. Historically, it wasn’t something a private consumer could buy, but you could use it as ride-hailing, either on their own network or through platforms like Uber or Lyft.
In recent months, there’s been talk about consumer Level 3 and Level 4, meaning you could actually buy a vehicle with Level 4 capability.
Kyle (04:52)
So these are like the robotaxis?
Lei Xing (05:10)
Eventually you can put that vehicle into a fleet, which is what Tesla is trying to do. FSD Supervised is essentially L2++. When you go from supervised to unsupervised, then it’s Level 4.
So now there are two tracks. Chinese companies like Xpeng are a good example. They’re pushing Level 4 consumer capability—what they’ll call “Robo.” They have Pro, Ultra, and then next is Robo, meaning Level 4 capability you can buy. At the same time, they’re developing robotaxis for fleets that you cannot buy.
That’s a recent and interesting change in direction for what’s possible in autonomy.
Kyle (06:12)
Yeah, the space is changing quickly. You have players that began with Level 2 “smart driving” features on consumer vehicles and are moving up. You also have companies that were mainly robotaxi-focused from the beginning. We’ll come back to that.
Before we dive into all of that, could you talk about Chinese government policies to support the industry—regulation, central versus local policy—and how they’re balancing safety and progress?
Lei Xing (06:41)
When you talk about China, you have to talk about roadmaps and overarching policy. I’d point to “Made in China 2025,” announced many years ago. There was a roadmap for EVs. If we bundle EVs, smart EVs, and autonomous vehicles together, in China it’s called the “intelligent connected vehicle” (ICV). ICVs encompass EVs, AVs, and everything else.
So you have that overarching roadmap—China wants to lead in these verticals: EVs, AVs. Then in recent years you’ve had a series of policies on safety. For example, we saw cybersecurity and data transfer rules announced this week: what could happen if a company like Tesla wants to use data collected in China outside China. That’s not currently allowed on both sides, but rules are coming into place that create pathways if you meet specific requirements.
I also mentioned the Level 3 pilot operation permits for the first two models. That’s on the national level. Locally, in tier-one cities like Beijing, Shanghai, Guangzhou, and Shenzhen, there are local pushes to pilot robotaxi operations. Pony.ai, for example—one of the big three robotaxi companies—is commercially operating in all four tier-one cities. They’re the only one currently, and that’s because of local policies supporting rollout.
Take Beijing as an example: permits are phased. First you test on roads with safety drivers. Once you reach milestones and accumulate experience and data, you move to the next permit—then you can have a safety driver as a passenger. Then you remove the safety driver completely. Baidu has achieved 100% driverless operations in the cities where they operate, but they had to go through testing and pilot operations before offering paid rides.
These local policies specify stages and thresholds for moving to the next step. So national and local policies work in tandem.
You also have requirements on the vehicles themselves. Years ago China was already piloting vehicle-to-road connectivity. Some cities—like Wuxi near Shanghai—have been test bases for vehicle-road communications. It’s a staged approach: communications, vehicles, testing, pilot operations—and now scaling. The policies make these developments possible. The overarching roadmap and local execution working together is important in China, and that’s true for many industries.
Kyle (11:52)
Yeah, definitely. This is a big theme for the whole podcast. And some of these things depend on the city—Wuhan versus Shenzhen versus Beijing. Beijing might be more restrictive for certain reasons.
Lei Xing (12:15)
Right. Wuhan is more lax—more willing to bring AVs into the city and run them. Beijing is the political capital, so it’s more restrictive. When companies like Baidu and Pony.ai get permits there, it’s tougher and more limited geographically—mostly certain areas like Yizhuang in southeast Beijing. They’ve also started operating at Daxing Airport and one of the train stations.
Wuhan is much broader—pretty much the whole city. So there are regional differences even within China—policies and support vary. It’s similar to the U.S.: Arizona is more permissive than other states, which is why Waymo runs in Phoenix and why Uber did a lot of piloting there in the past.
Kyle (13:41)
That’s really interesting. On robotaxis: what’s the experience like for an average urban resident in a major Chinese city? How do they summon one? What’s it like to enter? Is there a safety driver or not? Can you change your destination mid-ride? Does it feel like a Waymo, or is it different?
Lei Xing (14:15)
Yes to all of that. The latest robotaxi I took in China was in September, in Guangzhou, with WeRide. They’re running the GXR, a Geely van-style robotaxi.
You can hail these in several ways: through their own app, through WeChat mini programs, and through maps like Baidu Maps or Amap—because they’re integrated. In China, it’s very seamless.
The ride experience varies by vehicle type—comfort level, whether the driving is aggressive or jerky. I’ve ridden multiple platforms in China and in the U.S. For me, when I get into a robotaxi, I expect it to work. I don’t get nervous. It feels like another vehicle: you get in, you ride, you arrive.
If I had to rank comfort, Baidu’s latest RT6—minivan style—is very comfortable. Pony.ai runs a lot of Lexus RXs, and they’ve updated to newer generations. Baidu used to have cheaper vehicles that weren’t as comfortable.
Globally, I’d still call Waymo a gold standard. I’ve taken many rides in Jaguar I-Paces, which are quite comfortable. The best description is that they’re boring—which is good. They just work, and you don’t feel much.
In China, platforms like WeRide, Pony.ai, and Baidu have gotten to that point. But in terms of coverage and scaling rapidly, people still watch Waymo. We just saw their latest financing round and scaling plans. They want to be global.
But AVs are still regional. Chinese companies are arguably more global right now because they’re going into the Middle East and Europe. Waymo is moving toward places like London and Tokyo.
For consumers, the appeal is privacy—no other person in the vehicle—and then affordability, wait time, and ease of hailing. For robotaxi companies, the big questions are scaling and breaking even. Most now have fleets over 1,000 vehicles. By the end of this decade, many are targeting six digits—hundreds of thousands.
Kyle (18:42)
Things are moving fast. We’re just getting started on global expansion—robotaxi services coming out of the U.S. and China. A lot of them are using mixed strategies. Uber is partnering to offer WeRide in some cities. Waymo uses Zeekr EVs, or is starting to use them more. That could help drive down costs, especially outside the U.S. where Chinese EVs are heavily tariffed. So there’s an interesting mix of technologies—it’s not purely one tech stack versus another.
Lei Xing (19:29)
The simple debate is vision-only versus LiDAR. You have Tesla’s route with FSD, eventually robotaxi, versus most others using LiDAR. The jury is still out. There are proponents on both sides.
Kyle (19:32)
So Tesla uses cameras only, and the others use LiDAR?
Lei Xing (19:54)
That’s why we look not only at robotaxi companies, but the whole supply chain. And these companies have different strategies: focus purely on Level 4, or branch into Level 2++ “smart driving.”
Strategically, WeRide and Pony.ai differ. WeRide has moved into Level 2++ along with Bosch, while Pony.ai is still focused on pure Level 4 robotaxis. Then you have Momenta, which started with Level 4 but is now embedded in many Chinese and foreign brands in China for point-to-point Level 2++. At the same time, they’re working with Uber to launch robotaxis in Munich, and with Mercedes to launch both Level 2+ in China and robotaxis outside China.
Beyond that, there are mobility platforms: outside China, Uber and Lyft; inside China, DiDi, CaoCao, OnTime, T3, Meituan, and others. Hello Bike is moving into robotaxis. CaoCao Mobility—backed by Geely—has a goal of deploying 100,000 robotaxis.
DiDi has a joint venture with GAC Aion and recently produced a dedicated robotaxi. DiDi is a bit late. And then there are chip companies, compute, and other suppliers.
Even Lenovo is involved. They have a vehicle computing unit providing a domain controller with Nvidia Thor for the WeRide GXR robotaxis.
Kyle (22:39)
You mentioned LiDAR. Going back to the camera versus LiDAR debate: LiDAR used to be much more expensive, which is why Tesla leaned toward cameras only. Waymo has used LiDAR and radar, then refined its sensor suite over time. But LiDAR costs have dropped dramatically, and Chinese players seem to be leading. Can you say more?
Lei Xing (23:23)
There are “big threes” in each vertical. For robotaxis, there are big three companies; for LiDAR, there are big three companies too. The main leaders are Hesai, RoboSense, and another major player. Hesai and RoboSense are the biggest rivals. They’ve put LiDAR on vehicles at scale and reduced cost into the hundreds of dollars, compared to thousands just a few years ago.
At CES three years ago, the industry talked about the potential for a million LiDAR units annually. Now those two players alone are already delivering a million LiDAR units annually.
Part of it is that LiDAR is a selling point. Early adopters like NIO and Li Auto put roof-mounted “showerhead” LiDARs on vehicles. Now, with consumer Level 3 and Level 4 discussions—and robotaxi scaling—the addressable market is much larger.
Hesai, for example, doubled production capacity from 2 million to 4 million LiDAR units per year because Level 3 and Level 4 vehicles may require multiple LiDARs—top and sides. And foreign automakers and Chinese EV brands will standardize LiDAR on lower-priced vehicles because costs have dropped.
Kyle (26:09)
Right.
Lei Xing (26:09)
They can keep prices steady while adding LiDAR as a feature. Watch BYD this year—LiDAR will be a big part of their strategy to offset pressure. When you’re number one, it’s the most dangerous position. One of BYD’s tactics this year is pushing better features in their God’s Eye ADAS and standardizing LiDAR on cheaper vehicles.
Kyle (27:11)
That’s interesting, because some people think of LiDAR as a premium feature you only put on luxury models. But you’re describing economies of scale from standardizing across more models—even affordable ones.
BYD offering its God’s Eye smart driving system on vehicles below $20,000 is a capability people might expect only in higher-end cars.
Lei Xing (27:55)
Right. It’s not a premium thing anymore. It started as premium hardware, but it’s commoditized in China.
Kyle (28:18)
And there’s almost a floor in terms of safety, especially in difficult weather.
Lei Xing (28:32)
One policy we didn’t mention is compulsory AEB kicking in January 2028. Automatic emergency braking will be required on all vehicles in China—not only passenger vehicles, but also some light commercial vehicles. Companies like Huawei and Li Auto have said they can do AEB up to 120 kilometers per hour, and to do that, LiDAR is required, because at night you need long-range detection—300 to 400 meters. So policies like this expand the addressable market.
Kyle (28:42)
What is AEB?
Lei Xing (28:56)
Automatic emergency braking.
Kyle (29:29)
You mentioned Huawei. What is Huawei doing in this space? They seem to be everywhere, but they’re not a conventional EV or AV company.
Lei Xing (29:45)
Huawei officially entered the automotive sector at the April 2019 Shanghai Auto Show. We all know why: restrictions pushed them to expand into other verticals, and autos became one.
Huawei isn’t an OEM. They’re more like Xiaomi or Apple, but without making their own cars. Through the Harmony Intelligent Mobility Alliance, they work with state-owned automakers. They’re active across ideation and design of premium vehicles, and deep into hardware and software: components, their own LiDAR, HarmonyOS, e-motors, charging—many branches supporting the auto industry.
In China, “Huawei” pulls consumers. Their consumer products—phones, foldables—created a strong premium tech association. Vehicles like Aito and Luxeed are designed with features and quality tuned to Chinese consumer demand, and in many ways they’ve outpaced what foreign premium brands offer.
There’s also a national sentiment factor. Huawei positions itself as an integrator and facilitator helping state-owned automakers. Adding Huawei is like giving them wings—Huawei is the “Red Bull” for these automakers. Sales have taken off.
Kyle (32:42)
It’s the Huawei factor. The premium association, and integration with Huawei’s ecosystem—phones and other devices—adds something beyond the vehicle itself.
Lei Xing (32:56)
Take Aito M9: it’s priced above 500,000 RMB. For every ten vehicles sold in China above that price point, seven are Aito M9s. The value is hard to pass up. A Porsche at that price point isn’t really an option in China—Porsches are usually over a million RMB. Many buyers are upgrading, and Huawei delivers.
Kyle (34:15)
You had a great chart showing Porsche struggling in China. For a long time, it was the premium brand for successful businesspeople, but now market share is shifting toward Chinese brands—even in premium segments. It’s not just a race to the bottom.
Lei Xing (34:44)
And bringing it back to today’s topic: the ADAS L2++ race is intense, especially for point-to-point capability. We can put Huawei in the top tier right now for Level 2++ point-to-point.
Kyle (34:58)
I wanted to ask about AI and autonomous driving—models powering these systems and how approaches are shifting. In the U.S., we’ve seen a shift from hard-coded modular systems to end-to-end training. Do you see something similar in China? What’s happening on the model side?
Lei Xing (35:54)
Yes. Over the last one to two years, things have changed dramatically. Previously it was more rule-based. Now the buzzwords are VLA, LLM.
Xpeng is pushing out its second-generation VLA later this year. They want to reach Tesla FSD’s level—FSD Supervised, version 14-point-something. In retrospect, Tesla is still the gold standard for many Chinese companies: vision-only and a strong model.
Xpeng has even ditched LiDAR for its point-to-point Level 2++. They want to move to vision-only. Xpeng is often seen as China’s Tesla—they try to follow Tesla’s direction.
VLA—vision, language, action—and more generally “input in, output out,” end-to-end. For that, companies are moving into their own SoCs so they can optimize compute and capability rather than relying entirely on third parties.
NIO has its model, Li Auto has its approach, Huawei has its model. I experienced Li Auto’s system in September in Beijing. The moment you leave the parking lot, you press the button and it goes. During the trip, you can say, “At the next intersection, turn right,” even if the navigation was set differently, and it responds and adjusts.
This is what Jensen Huang talked about at CES—thinking and reasoning models. In China, Xpeng, Li Auto, NIO, Huawei all have their own models, but the direction is end-to-end. If vision-only works, the jury is still out. For true Level 4 robotaxi, LiDAR may still be needed.
We’re also seeing integration with broader models—DeepSeek, Doubao, and others—being incorporated into experiences. But Tesla is still a major benchmark.
Kyle (40:49)
It’s interesting to see convergence: adding reasoning and language capabilities, not just reacting to traffic patterns, but interpreting what people might do and responding appropriately.
You mentioned SoCs—self-developed chips. Can you say more about how much Chinese EVs and AVs rely on Nvidia chips versus domestic suppliers?
Lei Xing (41:50)
Still a lot of reliance on Nvidia. Thor and earlier chips are in many Chinese EVs, and it’s a selling point. Qualcomm is also moving into autonomous driving; previously their chips were mainly for smart cockpits. Domestic chip companies include Horizon Robotics and Black Sesame.
The future is single-chip integration—one chip that can handle both cockpit and driver assistance. Xpeng’s Turing chip is an example of that direction.
WeRide’s GXR robotaxi uses Lenovo vehicle computing with Nvidia Thor in the domain controller. Many vehicles on sale have dual Orin-X chips.
Some companies pick different strategies. Li Auto, for example, uses Qualcomm heavily for both smart driving and cockpit. Others think more about cost. There’s also geopolitical risk considerations—like access to high-end chips—and de-risking. Xiaomi is developing its own SoC too, extending from phones into EVs.
At the same time, it’s not the end of globalization. Capability and “make vs. buy” decisions differ across companies.
Kyle (45:37)
So many are pursuing full-stack approaches—self-developed chips and models—while still integrating with suppliers.
Lei Xing (45:58)
And this is where Mercedes and Volkswagen come in. They can’t build everything in-house, so they buy. Mercedes is working with Chinese partners and with Momenta. Volkswagen is working with Xpeng for EV architecture. Volkswagen is also the first foreign customer of Xpeng’s SoC—publicly announced—at least in the Chinese market.
This is the contrast: it used to be Chinese companies depending on foreign firms, and now it’s flipped—at least in China. That’s part of why China is a global frontier in this space.
Kyle (46:58)
Yeah, it’s flipped—especially in EVs and autonomous driving, where Chinese technology is sought after, from CATL’s batteries to Momenta to EV startups like Xpeng.
Looking ahead: with geopolitics, do you see a fragmented global market for robotaxi and smart driving tech? And what are the chances we’d see Chinese AV players in North America?
Lei Xing (48:10)
As we speak, Tesla is testifying with the Senate Commerce Committee on self-driving regulations. Yesterday, the CEO of NADA said their association would support blocking Chinese auto manufacturers from entering the U.S. They won’t control what individual dealers do, but that’s the stance.
For robotaxis, being truly global is difficult. Chinese “big three” are going into Europe and the Middle East. Mobility companies like Uber and Lyft aren’t in China, because China has DiDi, CaoCao, OnTime, T3, Meituan, and many others.
So it’s more regional, but there are neutral locations where multiple players can operate. The U.S. is not one of those right now. We can’t buy NIO, Xpeng, Li Auto, BYD—at least not yet. Chinese AV companies have tested in California and Silicon Valley, but there’s more scrutiny now.
Even LiDAR companies have U.S. presence: RoboSense has an office near Detroit, Hesai is based in Silicon Valley. Baidu had a U.S. office years ago; in 2018 they were showing ride-hailing robotaxi concepts, but those efforts didn’t go far.
The U.S. is a tough frontier. But there are workarounds. Waymo uses Zeekr vehicles—essentially the base vehicle—and then retrofits sensors in the U.S. So you can ask whether something could happen the other way around. I’m not sure, but it’s possible in some form.
In the near term, companies expand where operating conditions are favorable—dry weather, supportive policy, and funding. That’s part of why the Middle East is important. Europe too, in certain regions. WeRide and Pony.ai are already launching commercial services in Abu Dhabi and Dubai.
Can Chinese robotaxis operate in the U.S.? I can’t see that in the near future, just as Waymo is not in China. For the foreseeable future, the home market is the biggest market, but everyone wants overseas expansion to be called “global.”
We also have more players now. It used to be Cruise, but Cruise is no longer around. Now there are new partnerships and players—Plus there’s Zoox coming into play. In China, more local players keep entering too. The space is still expanding, and the end game is still murky.
Kyle (53:36)
One last question: commercial and delivery autonomous vehicles. We’ve talked about passenger vehicles, but there’s also trucking and last-mile delivery. What’s happening there?
Lei Xing (54:04)
This is another area where China is far along. One company is UIC—its CEO is a friend of mine. Their company, Cargobot, works on platooning for autonomous trucking. The founder was an engineer in the U.S. at Delphi. They’ve been operating platooning in Inner Mongolia—delivering goods and coal. In a platoon, you might have one human driver, but the other trucks don’t have drivers.
For last-mile delivery, there are companies like Neolix and Zelos—two major players for small autonomous delivery vehicles. They were at CES. People are often surprised by how many are already running in China—tens of thousands. They’re already doing real work.
When I go back to Beijing, I stay in Shunyi district. In that area, SF Express, Meituan, and others have autonomous delivery vehicles integrated into daily life. They’re just part of the flow with bikes and cars—no one is shocked by them anymore. In Yizhuang, where JD Logistics has its headquarters, these vehicles run around for last-mile delivery—including groceries.
In the true “last centimeter,” you still have human couriers delivering to apartments, especially for fast food, but they often take bags from these autonomous vehicles.
UIC’s niche is moving luggage at airports—pickup and drop-off. They’re already in Hong Kong, and I think in Qatar, and other airports. These are companies that are only about ten years old. Like EV startups, many are 10 or 11 years old. Their niches differ, so they don’t have to compete head-on. They find a niche, scale, and grow. And there are many more beyond the ones I’ve mentioned.
Kyle (57:59)
There are so many startups. I’d love to go to an airport and have a robot take my luggage—especially when you’re traveling with kids and don’t have an extra hand.
This has been awesome. If people want to learn more about you and your work, where should they go?
Lei Xing (58:45)
I’m on several platforms. I co-host China EVs and More. I have a Medium channel, and I do some YouTube—videos from events and my thoughts. You can find me on YouTube, Medium, LinkedIn. My handle is leixing77.
And one thing I want to correct: I don’t consider myself an expert. I consider myself an enthusiast.
Kyle (59:26)
You’re being too modest.
Lei Xing (59:28)
Because there’s so much to learn, and things are moving fast. The verticals—EV, AV, mobility, robots—at the end, it’s all one thing: automobile, auto, autonomous, mobile. I’m still learning myself. I’m glad I’ve covered the industry for so long, and there’s still so much to learn.
The people you meet along the way—Tu, you, founders of these companies—some I know personally. It’s fascinating to be part of this revolution. And it’s great we can communicate and educate through these platforms. We try to be fair.
Kyle (1:00:30)
Yeah. In our lifetimes, it’s very cool to watch.
Lei Xing (1:00:46)
We don’t want to be sensational. We can’t only say positive things—we have to be critical. Bad things happen too. And it’s interesting to watch, to be on this journey.
Kyle (1:00:50)
Yeah, exactly. We’re all learning together. I’ll include links to where people can follow you. Thank you so much, Lei Xing, for an awesome conversation.
Lei Xing (1:01:23)
Thank you for having me.
Kyle (1:01:25)
If you liked this episode, please rate and subscribe on YouTube, Spotify, or Apple Podcasts. You can find episode transcripts and more information on the High Capacity newsletter at high-capacity.com. I’m your host, Kyle Chan. Thanks for joining, and see you next time.