Toward the end of a nearly 15-minute video, William Sundin, creator of the ChinaDriven channel on YouTube, gets off the highway and starts driving in the southern Chinese city of Guangzhou. Or rather, he allows himself to be driven. For while he’s still in the driver's seat, the car is now steering, stopping, and changing speed—successfully navigating the busy city streets all by itself.
“It’s a NOA, [Navigation on Autopilot] function but for the urban environment,” he explains to the people watching him test-drive the XPeng G6, a Chinese electric vehicle model. “Obviously this is much more difficult than simple highway NOA, with lots of different junctions and traffic lights and mopeds and pedestrians and cars chopping and cutting lanes—there’s a lot more for the system to have to deal with.”
His final assessment? The Navigation on Autopilot isn’t perfect, but it’s pretty “impressive” and a preview of more advancements to come.
Beyond a simple product review, Sundin’s video is giving his followers a close-up view into the production race that has sped up among Chinese car companies over the past year. And whether they are electric vehicle makers or self-driving tech startups, they all seem fixated on one goal in particular: launching their own autonomous navigation services in more and more Chinese cities as quickly as possible.
In just the past six months, nearly a dozen Chinese car companies have announced ambitious plans to roll out their NOA products to multiple cities across the country. While some of the services remain inaccessible to the public now, Sundin tells MIT Technology Review “the watershed could be next year.”
Similar to the Full Self-Driving (FSD) features that Tesla is beta testing in North America, NOA systems are an increasingly capable version of driver-assistance systems that can autonomously stop, steer, and change lanes in complicated urban traffic. This is different from fully autonomous driving, since human drivers are still required to hold the steering wheel and be ready to take over. Car companies now offer NOA as a premium software upgrade to owners willing to pay for the experience, and who can afford the premium models that have the necessary sensors.
A year ago, the NOA systems in China were still limited to highways and couldn’t function in urban settings, even though most Chinese people live in densely populated urban areas. As Sundin notes, it’s incredibly challenging for NOA systems to work well in such environments, given the lack of separation between foot traffic and vehicles, as well as each city’s distinctive layout. A system that has learned the tricks of driving in Beijing, for instance, may not perform well in Shanghai.
As a result, Chinese companies are racing to produce more and more city-unique navigation systems before gradually expanding into the rest of the country. Leading companies including XPeng, Li Auto, and Huawei have announced aggressive plans to roll out these NOA services to dozens or even hundreds more cities in the near future—in turn pushing one another to move faster and faster. Some have even decided to release NOA without extra costs for the owner.
“They are launching it quickly in order to create awareness, to try to build credibility and trust among the Chinese consumers, but also, it’s FOMO [fear of missing out],” says Tu Le, managing director of Sino Auto Insights, a business consulting firm that specializes in transportation. Once a few companies have announced their city navigation features, Tu adds, “everyone else needs to follow suit, or their products are at a disadvantage in the Chinese market.”
At the same time, this fierce competition is also having unintended side effects—confusing some customers and arguably putting other drivers at risk. And underneath the automakers’ ubiquitous marketing campaigns, many of these features simply remain hard to access for those who don’t live in the pilot cities or own the high-end models.
Don’t think of it as full self-driving—at least not yet
The autonomous driving industry divides its technological advancements into six levels: from Level 0, where humans control the entire driving process, to Level 5, where no human intervention is needed at all.
There are really only two levels in use today. One is the tech in robotaxis, led by companies like Cruise, Waymo, and the Chinese giant Baidu, which offer Level 4 technology to passengers but are often limited in certain geographical boundaries.
The other level is the NOA system, exemplified by Tesla’s FSD or XPeng’s XNGP. They are only Level 2, meaning human drivers still need to monitor most tasks, but the technology is much more accessible and is now available in auto vehicles sold around the world.
It’s easy to believe that commercially available vehicles are closer to fully autonomous than they actually are, because Chinese car companies have given their NOA products all kinds of misleading or meaningless names:
- Li Auto follows Tesla’s tradition and calls it NOA
- NIO calls it NOP (Navigate on Pilot) and NAD (NIO Assisted and Intelligent Driving)
- XPeng calls it NGP (Navigation Guided Pilot) and more recently XNGP (the “last step before full autonomous driving is realized,” the company says)
- Huawei calls it NCA (Navigation Cruise Assist)
- Haomo.AI, an AI startup, calls it NOH (Navigation on HPilot)
- Baidu calls it Apollo City Driving Max
Apart from just being hard to remember, the different names also mean a lack of consistent standards. There’s no guarantee that these companies are promising the same things with their similar-sounding products. Some might only cover the major beltways in a city, while others go into smaller streets; some use LiDAR (a laser-based sensor) to help improve accuracy, while others only use cameras. And there’s no standard on how safe the tech needs to be before it is sold to consumers.
“Many such concepts are invented by Chinese companies themselves with no reference or background,” says Zhang Xiang, an auto industry analyst and visiting professor at Huanghe Science and Technology College. “What are the standards for achieving NOA? How many qualifications are there? No one can explain.”
Last September, two Chinese companies were racing to be the first to launch a city NOA system in China: On September 17, XPeng, the EV company that has long centered its brand image around the use of AI, managed to win the race by making its product available in Guangzhou. A week later, Huawei—a tech giant that has made smart driving a focus in recent years—launched it in Shenzhen.
But the progress really hit the accelerator in 2023. In January, Haomo.AI, a four-year-old Chinese autonomous driving startup, announced that it would make its city NOA service available in 100 Chinese cities by the end of 2024. Then on April 15, Huawei raised its goal to 45 cities by the end of 2023; three days after that, Li Auto, another Chinese EV company, pushed it further to 100 cities by the end of 2023. XPeng, NIO, and more companies followed soon after with similar announcements ranging in plans to expand to up to 200 cities.
For homegrown EV companies to remain competitive in the market, they are developing Level 2 navigation technology in-house and selling city NOA services as an upgrade to their vehicles; like other advanced features, they often require additional payments every month or year.
At the same time, AI companies are also competing against more conventional automakers, and as they work toward Level 4 or 5 self-driving technology, they still need interim revenue. NOA services can mean quick cash, easy sales, and, crucially, access to more data to train AI models.
“If you are just an autonomous vehicle company, significant revenue is 10 years out,” says Tu. “If you are under pressure from investors to generate revenues today, what do you do? You create incremental revenue through selling your hardware and software stack, or licensing it.”
Tu’s analysis is in line with what Cai Na, Haomo.AI’s vice president, told MIT Technology Review: “We think going the [Level 2] route is more realistic. The L2 technology has already made the breakthrough from being a technology to being a product, and many companies have turned it from just a product to a commercially viable product.”
The complexity on the ground
Behind all the big promises is the reality that today, these urban navigation services are not available to much of the public.
In 2023, about 360,000 cars produced in China will be equipped with city NOA capabilities, according to market research by Western Securities, a Chinese brokerage company. These models are usually more expensive than normal cars because they need hardware upgrades, like LiDAR or other sensors. Some companies are charging users extra for accessing the software functions, similar to what Tesla does.
But to have a car merely capable of providing city NOA is not enough. You also need to actually live in one of the few first-tier cities where the function has been made available, like Beijing, Shanghai, Shenzhen, or Guangzhou. Because of that, city NOA remains a niche technology in China right now (though few companies have shared data on their tech’s adoption).
For example, in addition to the XPeng test drive in Guangzhou, ChinaDriven’s Sundin has also test-driven a Li Auto vehicle with similar features in Beijing. He can’t use it in his daily commute, however, because he lives in Changsha, a second-tier city where no car company has enabled city NOA functions yet.
He notes how even in cities where it’s offered, navigation is often obstructed by poor road markings, new construction, pedestrians, or two-wheeled vehicles. “There’s a lot going on in China. The city is being turned over year on year; roads are being repainted,” he says. “And do mopeds drive on the road? Do they drive on the pavements?”
Haomo’s Cai echoes this: “[T]he real urban environment is far more complicated than what is imagined. The planning, policies, driving styles are different in each city,” she said. “For example, the traffic lights we usually see have three signs—red, yellow, and green. But some cities have five-sign lights, Chinese characters as signs, or triangle-shaped lights.”
But even within these pilot places, city NOA products still only offer limited functionalities.
Lei Xing, the former chief editor at China Auto Review, drove a recent XPeng model for a week in Beijing to test its autopilot features. XPeng was the first Chinese company to bring urban autonomous navigation to China’s capital city. So far, its autonomous features are limited to Beijing’s major ring roads and expressways.
One night when the traffic was light, he drove from a train station on the outskirts of the city all the way to Beijing’s innermost ring road highway, and XPeng’s tech did the driving for the whole process. Xing was impressed enough, but it still didn’t fully meet his expectations, particularly when traffic picked up.
He believes the automakers oversell their NOA’s capabilities: “I think the reality is much more difficult. These goals are quite aggressive, and I’m doubtful [that they will become true].”
XPeng, Li Auto, and Baidu didn’t respond to questions sent by MIT Technology Review. A Huawei spokesperson responded in an email that Huawei’s Advanced Driving System “has reached the start of production (SOP)” and focuses on three major scenarios: “highway driving, urban driving, and parking.”
Accidents waiting to happen
Even those who have been impressed by the urban NOA systems say it’s still a stressful experience. “When the traffic is busy, it will occasionally attempt to change lanes and cut someone off…sometimes it was too aggressive and it felt like I could bump into the car behind me,” Xing says.
Sundin also felt stressed when he tested XPeng’s features in Guangzhou. “To be honest, if you are a responsible driver and you are working with these systems, you are under a lot more pressure,” he says, mainly because he couldn’t predict how the car was going to react to traffic situations. “It can make you tired if you are properly monitoring what the system is doing,” he says.
Some of the cars offer checks on drivers to make sure they are paying attention. Xing says XPeng’s system would sometimes ask him to steer the wheel a little just to prove his hands were still holding the wheel. If the driver fails to do so, the car will warn the driver every few seconds. He says he also needed to complete a driver education procedure in which he was repeatedly reminded that the driver needs to stay focused and ready to take over the wheel. Sundin, however, found the same education mechanism lacking. The driver is asked to complete several multiple-choice questions, but he warns that it’s hardly an obstacle if you just click through all the answers to finish it quickly.
The fact that not every driver could be using the technology responsibly also means others on the road are more at risk. Unlike robotaxis, which are usually clearly labeled as such on the exterior or have noticeable sensors and cameras, a car with experimental NOA systems looks the same as any other car on the road.
“I don’t want to be part of someone else’s pilot if I’m driving a vehicle on the road,” says Tu, who has mixed feelings about how the products are currently being used. He thinks the industry is only one or two severe accidents away from the public and regulators turning against it.
“[How can you] strike that balance between being realistic and safe with your system but also using it as a selling point for your cars?” asks Sundin. “It’s a difficult situation, and I don’t know what the solution is. But definitely, if you are rolling it out, the education around these systems needs to speed up fast.”
Correction: We updated the name of William Sundin’s YouTube channel. It should be ChinaDriven, not ChinaDrive.
This new data poisoning tool lets artists fight back against generative AI
The tool, called Nightshade, messes up training data in ways that could cause serious damage to image-generating AI models.
The Biggest Questions: What is death?
New neuroscience is challenging our understanding of the dying process—bringing opportunities for the living.
Rogue superintelligence and merging with machines: Inside the mind of OpenAI’s chief scientist
An exclusive conversation with Ilya Sutskever on his fears for the future of AI and why they’ve made him change the focus of his life’s work.
How to fix the internet
If we want online discourse to improve, we need to move beyond the big platforms.
Get the latest updates from
MIT Technology Review
Discover special offers, top stories, upcoming events, and more.