Why does intelligent driving abandon high-precision maps?

In the wave of autonomous driving, high-precision maps were once regarded as the eyes of intelligent driving systems, and their integration with perception systems was one of the key technologies for achieving precise vehicle positioning and navigation. However, with the advancement of technology, different technological paths have emerged in the industry, and high-priced intelligent driving systems have begun to gradually abandon high-precision maps in favor of more flexible and efficient perception strategies such as "heavy perception + light mapping" and "mapless end-to-end."

Currently, nearly ten domestic driving assistance car manufacturers or technology companies, including Xiaopeng, Huawei, NIO, Li Auto, Horizon Robotics, and Momenta, have expressed that they will abandon high-precision maps in the future or adopt strategies that primarily rely on the vehicle's own sensors with high-precision maps as a secondary support to develop autonomous driving (assistance) technology.

Why are more and more car manufacturers abandoning high-precision maps?

01

The Shift from High to Low Temperature for High-Precision Maps

Advertisement

High-precision maps, also known as high-definition maps or high-precision navigation maps, refer to map data with decimeter to centimeter accuracy (as opposed to the meter-level accuracy of general maps).

High-precision maps are divided into many different layers. Within the map layers, in addition to the normal road shapes, directions, curvatures, etc., there are also details such as road gradients, inclinations, heights; types of lane markings such as solid lines, dashed lines, white, yellow; and roadside information like barriers and green belts. In the positioning layer, it includes information on traffic lights, traffic signs, pedestrian crossings, and other traffic facilities, as well as more detailed roadside landmark data.

Thus, it can assist the vehicle's sensors in better positioning themselves. The final dynamic layer contains high-frequency changing road information. For example, traffic accidents, traffic control, road congestion conditions, construction situations, weather conditions, changes in road markings, and changes in signs, etc., are all dynamic traffic information.High-precision maps contain multiple layers.

In simple terms, compared to ordinary navigation maps that serve "people," high-precision maps primarily serve "vehicles." In the three core processes of autonomous driving perception, decision-making, and execution, high-precision maps can participate in the perception and decision-making processes. With high-precision maps, vehicles can anticipate road information ahead. In the early stages when the vehicle's own perception capabilities are not yet powerful, relying on the guidance of high-precision maps can reduce the workload of driving control algorithms and achieve a certain level of autonomous driving.

For a long time, it has been a consensus in the industry that "high-precision maps are a must for autonomous driving above level L2." At that time, car manufacturers unanimously chose the technical route of "multi-sensor fusion such as LiDAR + high-precision maps," intending to achieve a breakthrough in high-level assisted driving functions first.

However, recently, the industry's attitude towards high-precision maps seems to have made a 180-degree turn.

02

Limitations of high-precision maps

If 2022 was the first year of mass production for high-level assisted driving in high-speed scenarios, this year is the first year for the expansion of high-level assisted driving services in urban areas. The shift from enthusiasm to coolness towards high-precision maps is behind the difficulties encountered when high-level assisted driving transitions from high-speed scenarios to urban travel.

The total length and complexity of urban roads far exceed those of highways. Urban roads are updated and changed quickly, are frequently upgraded and renovated, and temporary road closures are endless, which means that the map layers and positioning layers in high-precision maps, which do not need to be updated frequently, also need to be updated frequently, and the dynamic map layers even need to be updated in minutes.

This poses strict requirements for the freshness and cost of high-precision maps. Since the update of urban high-precision maps mainly relies on the survey vehicles of map merchants, it is already very difficult for leading map merchants to achieve monthly updates, and it is unimaginable to achieve weekly updates. Guo Sheng, Chairman of Lidar Space, once revealed that if you want to complete the high-definition mapping of the whole country, at least 7 billion yuan of investment is needed, and 30% of it needs to be updated every year. Such high initial large-scale mapping costs and subsequent high-frequency update costs are difficult for any car company to bear alone.Therefore, Tesla's FSD and Mobileye's REM both adopt the concept of crowdsourced mapping, where every vehicle equipped with the system acts as a surveying vehicle. The data collected during the driving process is uploaded to the cloud, forming a new dynamic layer. The more vehicles equipped with the system, the more accurate the map becomes, and the more timely the updates are. However, map surveying is a core interest of national security, and not everyone can obtain the qualification to participate in high-precision map surveying.

The high cost and lack of timely updates have become important reasons for car manufacturers to "unbind" from high-precision maps. In recent years, the development of autonomous driving technology in areas such as sensors, recognition algorithms, and computing power has made rapid progress. The computing power of autonomous driving chips has quickly reached over 200 TOPS from 30 TOPS two years ago, and the new generation of products, such as NVIDIA's DRIVE Thor, will even reach 2000 TOPS. Relying solely on the vehicle's own perception can solve the vast majority of problems, which is also the confidence for car manufacturers to enter the "mapless navigation" era.