Credit: Visual China

BEIJING, September 26 (TMTPost) – The No. 1 contest for new energy vehicle companies this year is undoubtedly the use of NOA (Navigation on AutoPilot) on city roads.

NOA is a typical function of L2+ assisted driving, which enables point-to-point autonomous driving on highways, urban expressways, and ordinary urban roads.

With Huawei, XPeng, and Li Auto shouting out the aggressive slogan of intelligent driving entering the city, consumers' interest in intelligent driving has gradually extended from highway scenes to urban scenes.

However, just as the market's enthusiasm for urban NOA that was ignited, the carmakers unexpectedly presented a seemingly "compromised" solution: commuting mode.

Why is it called "compromised"?

Because commuting mode restricts the use area to the familiar route that each user frequently uses, rather than being universally applicable.

However, for autonomous driving, the innovation of the business model is obviously not attractive enough.

Therefore, a voice of doubt has also emerged: Autonomous driving doesn’t directly move from highway scenes to urban scenes, but presents a point-to-point "commuting mode". Is this a sudden inspiration in product design or a roundabout strategy due to the difficulty of implementing urban NOA?

Concerted Efforts to Promote "Commuting Mode"

Just as its name suggests, "commuting mode" allows consumers to achieve intelligent driving on their daily commuting routes without having to wait for updates of high-precision maps or for manufacturers to open up a particular city. They only need to undergo short-term manual training.

In March 2023, XPeng first proposed the concept of "commuting mode". Subsequently, car companies and autonomous driving companies followed suit.

In June 2023, DJI proposed the "Memory Driving Mode" based on low-cost hardware, which will be implemented in Baojun Yunduo.

Following that, in August, Li Auto announced its plan for Commuting NOA, which officially started internal testing for early users in September, and will be offered to full-scale AD Max users in January 2024, followed by AD Pro. In fact, since the end of June 2023, Li Auto has been offering the first city NOA that does not rely on high-precision maps to early users.

Zhiji Auto started public testing of NOA without high-precision maps in September, and is expected to cover more than 100 cities nationwide by 2024 and realize the mode of Door to Door (all-scenario commuting) in 2025.

Autonomous driving company DeepRoute released two intelligent driving products, D-PRO and D-AIR, which integrate parking and driving, claiming that they do not require high-precision maps and can achieve point-to-point intelligent driving in cities at lower hardware costs.

Yu Chengdong, CEO of Huawei’s Intelligent Automotive Solution (IAS) BU, previously announced on Weibo that the NCA in urban areas that does not rely on high-precision maps will be implemented in 15 cities in the third quarter of this year, and will increase to 45 cities in the fourth quarter. At the recent AITO M9 launch event, Huawei chose to adjust its goals and announced that it will directly promote the mapless version of ADS 2.0 nationwide by the end of the year.

Regarding Huawei's sudden choice of a more aggressive strategy, there is speculation that it may be influenced by Tesla's FSD.

As a loyal advocate of mapless driving, Tesla first launched a FSD Beta V12 test drive live broadcast in August, personally conducted by Elon Musk. Then, it listed "City Street Autopilot" as a "coming soon" feature and moved it to the available/deployed feature list. This has led to speculation that Tesla's FSD will end its more than three-year testing marathon and be deployed in the vehicles in the U.S. and Canada, achieving real-world applications.

In addition to the motives, there is also discussion about what Huawei's mapless version of ADS 2.0 is. One common speculation is that by the end of the year, Huawei is likely to promote the "commuting mode" of ADS 2.0 nationwide, rather than its full-scale mapless capability.

It seems that both car companies and autonomous driving technology companies are collectively intensifying their efforts in the "commuting mode".

"This scenario may be a transitional phase," said Zhong Xuedan, Vice President of Tencent Intelligent Mobility, in an interview with TMTPost App.

In Zhong’s view, urban-level mapping is a relatively complex process, and the commuting scenario is easier compared to other scenarios, either in terms of understanding and familiarity with the data on the road, or in terms of the mapping capabilities. Therefore, starting from this scenario is easier and more focused than directly covering the entire city.

The term "no map" refers to the idea of being independent of high-precision maps.

Starting in 2022, the development of high-speed NOA has rapidly extended to urban scenes. Car manufacturers and suppliers have released related products and announced their landing plans. In the process of expanding from high-speed NOA to urban NOA, the question of whether to use high-precision maps has become a hot topic in the industry.

Why do some players prefer not to use them? The high cost, slow update speed, scarcity of top-level mapping qualifications, and the development of perception technology are all reasons, with cost being the primary factor.

According to the white paper on high-precision map of intelligent connected cars, using traditional mapping vehicles, the mapping efficiency of decimeter-level maps is about 500 kilometers of road per day per vehicle, with a cost of about 10 yuan per kilometer. The mapping efficiency of centimeter-level maps is about 100 kilometers of road per day per vehicle, with a cost of possibly thousands of yuan per kilometer, resulting in costs in the hundreds of thousands per day.

"High-precision maps go hand in hand with high-precision positioning. If you use high-precision maps, you must also use high-precision positioning. And if you have high-precision positioning, you must also use high-precision maps. The cost of combining these two modules is even higher," said Yang Wenli, Founder and CEO of Beijing Lingjun Technology.

In the process of "abandoning" high-precision maps, there are moderate approaches, such as using "light maps."

Using "light maps" primarily has two approaches. One is to "subtract" from traditional high-precision maps, reducing some map elements based on actual needs. The other is to "add" to navigation maps, adding some elements specifically for intelligent driving needs.

There are also radicals that firmly choose the "no map" mode, such as Tesla.

"No map mode" sounds lighter and more reformative, but it is not easy to implement in practice.

An industry insider even bluntly stated, "When we say 'no map’, it means that maps are actually very important, otherwise the term ‘map’ would not even be mentioned in autonomous driving."

In terms of technical difficulty, Yang explained that the difficulty level of urban roads is at least 1-2 orders of magnitude higher than that of highways. The intersection structure in urban roads is complex and the standardization still needs improvement, especially during the morning rush hour when there are more road participants and the drivers' habits are more complex.

Therefore, the advantages of high-precision maps become apparent. Because high-precision maps are prior knowledge and have a long-range view, they can cooperate with high-precision positioning to provide centimeter-level positioning, accurate driving assistance information, and semantic information, which is equivalent to opening a "God's perspective" for advanced assisted driving.

Yang explained, due to these characteristics, although the vehicle has not driven in some places, it can know the road structure from a long distance through high-precision maps, such as the number of roads and whether a road is a left turn or a straight road. This allows the car to turn onto the target lane early. If high-precision maps are removed and only navigation maps are used, some information will not be as accurate, which will have higher requirements for real-time perception and decision planning. "Light maps" greatly reduce costs while significantly increasing technical difficulty, becoming the direction pursued by many autonomous driving companies and car manufacturers.

In the view of Meng Qingxin, Senior Vice President of NavInfo, the simple expression "no map" itself is a rough description. "High-precision maps can be divided into several categories, with different levels of accuracy and coverage range, making it a very complex field."

How Far Is the "Commuting Mode" from Urban NOA?

By building an intermediate zone between having maps and not having maps, the "commuting mode" provides a shortcut for Chinese autonomous driving mass production players to "enter the city" right now.

In the vision of these new forces, after multiple tested commuter routes are combined, it may provide a "base map" for the implementation of city navigation functions.

For example, in the ideal commuting mode function, human drivers are still needed in the early stage to drive the vehicle and achieve point-to-point driving from home to the office. However, at the same time, the vehicle will use its sensing hardware and map data on the body to perceive and record information about the routes it passes through, and then provide it to the ideal NPN neural prior network algorithm for extraction, invocation, and learning. Ideally, NOA will then be able to be implemented in densely populated areas that have been extensively tested with the commuting mode.

XPeng's "AI driving" can analyze the user's commuting routes and other relatively fixed routes in a short period of time, learn the user's driving style, and thus achieve personalized route customization. Moreover, the company stated that the routes of AI driving can be shared with other users.

However, high-precision mapping relies on professional mapping vehicles, while commuting mode relies on consumers themselves.

From the perspective of data collection time, mapping vehicles are often deployed during periods of low traffic such as early morning and midnight, while commuting vehicles are mostly used during rush hour when the traffic is congested, and the presence of numerous dynamic objects can obstruct the scene and cause a sharp drop in data quality.

It is thus not easy to form a closed data loop by training large models with massive data to improve the maturity of autonomous driving technology.

In addition to data quality, the current production scale of vehicles by automakers is limited, resulting in limited total data obtained. Besides, how to quickly select high-value data from massive data will undoubtedly test the annotation capabilities.

More importantly, if more and more people use the commuting mode, although it will provide manufacturers with a large number of road testers and much data feedback, the issue of data compliance and regulation cannot be avoided.

From the perspective of consumers, in addition to the limited scope of use, drivers in commuting mode also need to actively train the "commuting mode", which may require a very long training time, and the final usage effect cannot be fully guaranteed. Whether this mode can really help consumers is still a question.

Under various challenges, how far is it from commuting mode to urban NOA?

Yu Kai, CEO of Horizon Robotics, believes that people should not be too anxious about autonomous driving. By 2025, what is really needed to do is to achieve a smooth and affordable experience of autonomous driving on closed roads such as high-speed NOA and ring NOA, and the price should not be too high. At the same time, there needs to be considerable investment to truly make more complex urban NOA usable.

Meng also agrees with this judgment.

She explained that the government's regulatory trend is gradually loosening, and NavInfo has obtained the approval for autonomous driving in 120 cities in June this year. But when it comes to these open scenarios, is the car's functionality really safe? Are the data reliable and compliant? There are still many issues that need to be resolved, and 2024 and 2025 will be more practical landing time points.

Except for Tesla, all other companies' city navigation functions are currently still based on map solutions. XPeng, Baidu, Huawei, and Li Auto are all expected to launch their mapless solutions by the end of 2023.

The reason why the "commuting mode" has attracted much attention is because many people believe that the commuting mode is a shortcut to solve the limited scope of NOA in cities. However, from the actual situation, the commuting mode not only has its own unique challenges, but the real difficulties of NOA in cities may not be avoided by the commuting mode. Perhaps there is no shortcut to the success of autonomous driving.

(This article was first published on the TMTPost App, Author: Han Jingxian, Editor: Zhang Min)