Musk has long since predicted the doom of high-precision maps, but domestic smart driving is now heading towards the same end as Tesla?

Recently, many test videos of Ideal 5.2 Beta AD have appeared on the Internet, and it has a very good ability to cope with complex road conditions in urban scenes. A few days ago, Zhiji L6 officially started delivery, and the NOA function of cities without high-precision maps was launched simultaneously, including more than 10 cities such as Beijing, Shanghai, Guangzhou, Shenzhen, Hangzhou, Suzhou, and Nanjing. Not long ago, Haobo HT has started the city map-free NDA internal test in Beijing, and in July it will add 41 cities to the current 6 cities, and cover the whole country in September.

I believe you have also discovered that the common point of these new features is map-free intelligent driving. It seems that since last year, high-precision maps, which were once regarded as one of the guarantees for autonomous driving, have suddenly begun to be abandoned by domestic car companies.

Musk started to criticize high-precision maps as early as 2019, including his later rejection of LiDAR and even millimeter-wave radar in favor of a pure visual approach, which was sneered at by many people. But suddenly, it seems that China's intelligent driving is heading for the same destination as Musk?

What is a high-precision map?

High-precision maps are high-precision maps designed specifically for autonomous driving. They contain rich information with centimeter-level accuracy, such as various lane lines, slopes, road curvatures, traffic signs, road surface markings, etc., to help vehicles perceive the complex environment around them, thereby achieving stable autonomous driving and precise navigation.

Simply put, a high-precision map is like an extremely accurate and meticulous navigator, which allows the system to understand the detailed information of the road in advance and establish a sense of road conditions outside the range of the sensor. This is completely different from the navigation maps we use on our daily mobile phones and car computers. In fact, we cannot "see" the so-called high-precision map. It only exists in the equipment of the technical research and development department behind the car company, providing technical support for the entire autonomous driving system. It can be said that the role and help of high-precision maps for autonomous driving are unquestionable.

So why did we suddenly start to “go to the map”?

After all, high-precision maps involve a large amount of data collection, analysis and processing, which requires a lot of investment from map vendors. At the same time, the production cycle and data update cycle are long. Even highways can only be updated quarterly, which is not as frequent and timely as mobile phones or car computers. It is even beginning to fail to keep up with the speed of autonomous driving development. Solving these problems will undoubtedly require higher and higher costs, which are not affordable for map vendors, and it is obviously impossible to cover all cities in China.

For car companies, issues such as user feedback and map updates have always been resolved through map vendors, so they also hope to take the initiative in map resources. However, because map information involves various geographic and privacy data, and the country is extremely strict in supervision and policy, car companies still have a high threshold to enter the high-precision map market.

At the same time, with the rapid development of various on-vehicle sensors and hardware such as lidar, millimeter-wave radar, high-definition cameras, and autonomous driving chips, the vehicle's own perception capabilities have been greatly improved. Therefore, car companies have gradually begun to look for a new "map-free" or light-map route, reducing dependence on high-precision maps and paying more attention to the vehicle's own perception capabilities.

What is the so-called "no picture"?

The so-called "no map" does not rely on high-precision maps. Based on the basic data of map vendors and various data collected by the vehicle's own sensors, a local high-precision map is generated in real time. Compared with the traditional direct use of high-precision maps in the past, no map relies more on the vehicle's real-time perception of the environment, emphasizes the vehicle's own capabilities, and reduces dependence on pre-made maps. It can also be considered a route of "heavy perception and light map".

But one thing we must be clear about is that the absence of maps does not mean the absence of high-precision maps. There is nothing wrong with high-precision maps themselves. It is just that the form in which high-precision maps are formed has changed.

It should be noted that many car companies, including these new forces we are familiar with, have more or less demonstrated a technological advancement when promoting the cancellation of reliance on high-precision maps, making consumers feel that "de-mapping" is a more advanced route.

In fact, many industry insiders believe that autonomous driving cannot do without high-precision maps, whether as a core foundation or as a redundant collaboration. Moreover, from surveying and mapping qualifications to recruiting professionals, various car companies are actually increasing their investment in the field of intelligent driving maps. Now a large number of car companies are starting to take the route of heavy perception and light mapping, which is more like a result of forced transformation. Car companies want to take high-precision maps into their own hands, but they cannot fully realize it for various reasons.

The current map-free solutions still face certain challenges. On the one hand, they have certain requirements for vehicle hardware and computing power, and need to solve problems such as low latency and software and hardware integration. On the other hand, the accuracy of the map will also decrease to a certain extent. In some places where high-precision maps are already covered, car companies are still willing to use high-precision maps, just as the CEO of NavInfo once publicly questioned a new brand.

Therefore, it would be more appropriate to say that the “图” in “去图” refers to map vendors rather than maps.

This trend started in China at the beginning of last year. Representatives of new forces such as Wei, Xiaopeng and Li Auto all said internally that they would gradually abandon high-precision maps in the future. Later, Yu Chengdong also made it clear that Huawei's autonomous driving should not be overly dependent on high-precision maps in the future. Not only car companies, but also autonomous driving technology companies such as Horizon Robotics, Yuanrong Qixing and Momenta have also expressed their intention to reduce or even eliminate their dependence on high-precision maps.

We had a more intuitive experience at the end of last year, when AD MAX 3.0 in Ideal OTA 5.0 opened up NOA for urban roads in 100 cities, and Xiaopeng Xmart OS 4.4.0 realized XNGP for urban roads without maps in 25 cities. In the past two months, Huawei Intelligent Driving ADS 2.0 began to exert its strength, and Huawei brands such as Wenjie, Zhijie and Avita began to push NCA for urban roads without maps that can be driven throughout the country, which pushed the "no map" issue to a climax and became a keyword mentioned by many users who enter the store. Now it is getting more and more intense.

Can you play without pictures?

But if there is no high-precision map as a foundation, how can we achieve increasingly advanced autonomous driving?

As we mentioned above, the current main map-free solutions or light map solutions mainly generate real-time online local maps through the comprehensive technical applications of powerful real-time perception capabilities of vision and sensors, large models and super-resolution technology, and increasingly powerful computing power support, to replace the functions of high-precision maps in the past.

For example, the BEV+Transformer algorithm solution, which is currently more commonly used, perceives the environment from a bird's-eye view and transforms it in space and time through Transformer combined with AI deep learning, and combines the emerging occupancy network to more effectively mark obstacles to form the entire model. This is also the technical route advocated by Musk a few years ago, but Tesla uses a pure visual BEV, rather than the integrated BEV route adopted by most domestic car companies.

This greatly reduces the high cost of high-precision maps, but it places high demands on the algorithms and computing power of car companies. Even though chips are developing rapidly, there are still certain technical difficulties in achieving online generation of high-precision maps plus crowdsourcing or V2X. At present, map-free or light-map routes can basically solve the scenarios of ordinary intersections and regular roads, but special-shaped intersections, large intersections, construction intersections, and large traffic volumes are still difficulties that need to be overcome by the current map-free technical route.

However, autonomous driving without maps has achieved certain results in actual performance, especially in the past few months, in urban areas with relatively less coverage of high-precision maps, from Huawei ADS 2.0, Xiaopeng XNGP to Ideal City NOA, the actual experience of autonomous driving in urban areas without maps is remarkable. The process of de-mapping is smoother than we thought, and the pace of urban autonomous driving is faster than we thought.

Bang Reviews

For ordinary consumers, it doesn’t matter whether there is a “map” behind the function or what kind of “map” it is, as long as the experience it brings us is good enough. At least the overall feedback from users on the launch of the NCA in the urban area without map at the beginning of the year is still very positive. It is still one of the hottest topics in the field of autonomous driving. It is also the direction that car companies will continue to focus on in the field of autonomous driving this year. This also indirectly shows the success of the current attempts at the map-free route.