An article reviews the "no-map" intelligent driving solutions of various car companies-EEWORLD

Collect

origin

Industry development needs

In the second half of 2022, it can be considered that the problem of intelligent driving on highways and urban expressways has been basically solved. Huawei and Xiaopeng, which have invested heavily in intelligent driving, have found that the proportion of cars used on highways and urban expressways is less than 15%. Intelligent driving cannot be called a decisive factor for consumers in buying cars, and it cannot justify the huge R&D investment.

In order to further expand the scope of application of intelligent driving, various car companies have focused their attention on urban NOA, but they are faced with a thorny problem in this process: the coverage ratio and freshness of high-precision maps in urban areas are far from enough!

Although the country has seen the difficulties of the industry and the Ministry of Natural Resources announced the pilot projects for the application of intelligent connected high-precision maps in six cities, including Beijing, Shanghai, Guangzhou, Shenzhen, Hangzhou and Chongqing, in the second half of 2022, it still faces problems such as insufficient number of open cities, high cost of high-precision maps, and slow map review.

The process needs to go through: the list of cities where policies allow data collection--> map vendors collect and draw maps based on their qualifications--> the Ministry of Natural Resources reviews--> car companies develop and adapt, and then popularize on a large scale. The above process takes 2-3 years, which is obviously too long for the leading car companies that are already at the starting line.

Therefore, how to bypass high-precision maps has become a key issue in the industry's exploration. Some terms have emerged in this process: "SD+", "SD pro", crowdsourcing map construction, light map, etc. It should be noted here that the actual definition of "no map" is different for each company, but the same point is: no map means not relying on high-cost high-precision maps.

The above solution is still from the perspective of map vendors to solve the map problem starting from the "map construction" itself, and does not solve the city NOA from the actual needs of car companies for intelligent driving "using maps".

"Sparks"

Before October 2022, the "no-map" solution was rarely mentioned in public reports about the industry, but from 2023, the "no-map" solution began to be widely reported and disseminated. I personally think that the turning point occurred at Tesla AI Day on October 1, 2022. Tesla only used the geometric & topological relationship of the roads in the SD map in the proposed FSD Lanes Neural Network solution. The lane level, number, width, attribute information and occupancy features can construct a real-time lane-level topological structure, thus ending the difficulty that the lane-level topological structure can only be obtained from high-precision maps, and providing an alternative to high-precision maps in terms of technical solutions and implementation.

"The momentum of a prairie fire"

As a world-class new energy vehicle company leading the trend of autonomous driving technology, Tesla's technical solutions were quickly followed by various domestic leading intelligent driving car companies. Xiaopeng and Huawei turned to developing BEV solutions that do not rely on high-precision maps. Since then, "no-map" solutions have begun to be reported and appear, and reached their peak in early 2023:

--In January 2023, He Xiaopeng of Xpeng Motors clearly stated at the Xpeng Motors all-staff meeting that the 2023 X-NGP assisted driving will abandon high-precision maps.

--In January 2023, Li Xiang, CEO of Ideal Auto, stated in a letter to all employees that Ideal Auto's end-to-end trained urban NOA navigation assisted driving (not relying on HDMAP, that is, not relying on high-precision maps) will be implemented at the end of 2023.

--In March 2023, at Huawei's spring conference, Yu Chengdong, CEO of the Intelligent Automotive Solutions BU, revealed that the M5 and M5 EV advanced intelligent driving versions, which will be launched in April this year, will be equipped with intelligent driving solutions that do not rely on high-precision maps.

With the official joining of the recognized first-tier domestic intelligent driving companies (Huawei and Xiaopeng), the "mapless" solution seems to have become a development trend in the intelligent driving industry.

progress

Tesla

Tesla does not use high-precision maps, as shown in its 2022 public plan. Instead, it uses SD maps and can build topological information of intersections. Tesla has the ability to build maps, but it is mainly used to provide annotations for training data.

In 2023, Tesla's FSDv12 version has moved towards an end-to-end solution. Judging from Musk's video, the current end-to-end capabilities are similar to or slightly weaker than Tesla's previous solutions, but its evolution speed may be very fast.

Around 2019, Baidu developed end-to-end community automatic parking, which is similar to super memory parking, but the scenarios are very limited. But compared with Baidu, Tesla is fully capable of achieving end-to-end. In summary: As a global leader in autonomous driving, Tesla has realized the "no map" solution.

Xiaopeng

Before 2023, Xiaopeng's intelligent driving solution is also following Tesla. Its perception XNet relies more on pure visual BEV and can achieve super strong environmental perception capabilities. Xiaopeng's "mapless" is achieved by relying on XNet to build a "high-precision map" in real time.

Xiaopeng's NGP without map and NGP with map are based on the same technology stack. The only difference is that the original high-precision map input is replaced by the input of navigation map and the understanding of navigation information through real-time perception. To this end, XNet needs to increase the perception distance and provide the beyond-visual-range environmental information required for decision-making and planning; at the same time, by learning a large number of road and intersection features, it enhances the ability to perceive complex road structures. It is said that the total perception area of XNet is about the size of 1.8 football fields, with a 150% increase in longitudinal perception and a 200% increase in lateral perception.

In the first half of 2023, Xiaopeng began to develop a map-free solution based on SD. The plan is to launch the map-free solution by the end of the year, and enable automatic lane change, overtaking, and left and right turns in cities without high-precision map coverage. The difference between map-based and "map-free" information is as follows: the information missing from "map-free" needs to be obtained through perception.

The "no-map solution" has basically no problems when driving on roads, but the main problem is at intersections. It can be seen that there are indeed perception blind spots at intersections, so the corresponding decisions will appear cautious and hesitant. Through video comparison and private contact, it is understood that Xiaopeng's "no-map" solution does exist, and it does use only SD maps. Regulation, control and decision-making rely more on real-time perceived information.

From a stage perspective, Xpeng's map-free solution is currently in the stage of generalization and expansion from a single city to multiple cities. We should be able to see OTA on some models by the end of the year, and more detailed information may be revealed on 1024 next week.

Compared with the solution with a map, Xpeng's "no-map" solution has the advantages of being 4 to 10 times faster in generalization speed, completely solving the problem of data freshness, reducing costs and popularizing intelligent driving.

In addition to the "no map" solution, Xiaopeng also has another mode, "AI" driving, which only needs to be learned once to start the city NOA. "AI" driving is more of a mode between crowdsourcing maps and no map. During the learning process, it only remembers navigation points and turning information. In the actual operation process, it is inseparable from the "no map" capability. Learning is only for improving experience and safety.

The main purpose is to learn personalized driving strategies, not just for map building. From the problems encountered in urban areas, it can be foreseen that in the face of high-frequency changes in urban freespace roads, the intelligent driving system needs to have strong real-time perception and decision-making capabilities.

Huawei

From the information released by Huawei before the Shanghai Auto Show in April, we can see that Huawei uses laser + vision god network and SD map to achieve the "no map" solution. In terms of implementation, Huawei uses laser to achieve functions similar to the "occupancy" network, and can generate the topology of the road in real time. From the perspective of sensor configuration (laser + vision) and basic network (transformer), the solutions of Tesla, Xiaopeng and Huawei are very similar.

Huawei's plan is to popularize non-map solutions in 15 cities by Q3, but it has not been realized yet, and Xiaopeng has not achieved the goal either.

Huawei's intelligent driving experience is better than Xpeng's, and the "no-map" solution is planned to be available nationwide by the end of the year. Huawei initially adopted crowdsourcing to build maps, but after finding the reconstruction of Shanghai too difficult, it gave up. Judging from Huawei's PR propaganda, it is all in on the "no-map" solution.

NIO

NIO’s intelligent driving capabilities have always been considered to be in the second tier, with a certain gap from Huawei and Xpeng, but NIO’s intelligent driving team is also large.