HD maps were once a critical component of advanced intelligent driving. However, today, intelligent driving in passenger vehicles is gradually reducing its reliance on HD maps, delegating more road information to be acquired and understood in real time by the vehicle itself. This trend actually involves a key shift in autonomous driving: should vehicles rely on pre-prepared road priors, or rely more on their own ability to understand the road during driving?
01. Why Were HD Maps Once So Important?
An HD Map is not a simple upgrade of ordinary navigation maps, but a set of highly detailed road information designed for autonomous driving systems. It can describe lanes, road geometry, intersection topology, traffic facilities, and other information, providing vehicles with more structured road priors. Once a vehicle enters a road that has been covered by the map, it can combine this information for localization, path planning, and driving decisions. This was crucial for early advanced intelligent driving systems.
Vehicles do not just need to know where the destination is; they also need to understand the structure of the road they are on, where the current lane leads, and how the upcoming intersections connect. HD maps can structure a portion of this road information in advance, providing the vehicle with a stable environmental prior during actual driving.
The advantage of this approach lies in its clear information structure and stable road priors. Especially in scenarios where the road structure is relatively fixed, HD maps can help the system reduce the pressure of real-time environmental understanding. However, this direction also has some issues. Maps record roads after they have been collected and produced, whereas vehicles actually face continuously changing real-world roads.
Road construction, lane adjustments, intersection modifications, and changes in traffic organization can all cause discrepancies between real-world roads and historical map information. The larger the road coverage of the map, the more information needs to be maintained and verified. For solutions relying on HD maps, adding new cities, parks, or operational areas requires corresponding map data preparation and verification.
When the scale is small, this process is relatively easy to control; however, when the intelligent driving system needs to cover more and more roads, map production, updating, and maintenance themselves become a factor to consider for scaled application. Therefore, the problem with HD maps is not that they are inaccurate or valueless. Rather, when vehicles need to face an increasing number of roads that are changing faster and faster, can pre-produced road information always keep up with the real world?
02. Why Are Vehicles Beginning to Have the Ability to Understand Roads Themselves?
The reason why automakers are becoming less reliant on HD maps is that the vehicles' own environmental understanding capabilities are steadily advancing. Passenger vehicles continuously acquire surrounding environmental information through sensors such as cameras, and some models are also equipped with LiDAR, utilizing perception models to identify and model roads, vehicles, pedestrians, traffic signals, and drivable areas. Meanwhile, environmental representation methods such as BEV (Bird's Eye View), along with 3D scene modeling technologies like Occupancy Networks, are also enhancing the vehicle's online understanding of road spatial structures and the surrounding environment.
These technologies do not mean that maps are completely unnecessary, nor did they emerge specifically to replace HD maps. Their more important role is to equip vehicles with stronger real-time environmental modeling capabilities and provide a technical foundation for reducing reliance on certain HD map dependencies. The development of end-to-end models is also changing the way autonomous driving systems process information. Traditional systems break down tasks such as perception, prediction, and planning into different modules, whereas the end-to-end approach allows more driving tasks to be completed by a unified model or more tightly coupled models, reducing information transmission loss between modules and enabling the system to more directly utilize current environmental inputs to generate driving behaviors.
End-to-end models do not mean maps are unnecessary either; rather, when vehicles can obtain more driving-related information from real-time sensor data, their reliance on pre-prepared road priors decreases. On this basis, the way autonomous driving understands roads is also changing.
In the past, it was necessary to know in advance what the road looked like. Now, there is an increasing emphasis on the vehicle understanding the current condition of the road itself. For example, as a vehicle approaches an intersection, the system can not only call upon existing road information but also use cameras and other sensors to identify current lane markings, road boundaries, traffic lights, surrounding vehicles, and pedestrians in real time, and then combine this with the current traffic status to determine how to proceed. If the road is under construction and the original lanes have changed, the system no longer just faces the question of whether the map has recorded this change, but must determine what road structures actually exist on-site, which areas are drivable, and how other traffic participants are moving. This is where real-time perception truly demonstrates its value.
However, it must be emphasized that "mapless" does not mean completely without maps. In current passenger vehicle intelligent driving systems, navigation maps can still handle tasks such as destination search and road-level route planning. The so-called "mapless" approach is more about reducing reliance on fine-grained road priors at the lane and intersection levels in traditional HD maps, rather than deleting all map information. Therefore, a more accurate description is not a shift from having maps to having no maps, but a gradual transition from highly relying on HD maps to real-time perception and map priors working together, with the former taking on an increasing number of road understanding tasks.
03. What Does Reducing Map Reliance Truly Change?
From the perspective of passenger vehicle intelligent driving, reducing reliance on HD maps truly changes the vehicle's ability to understand and cope with the current road even when complete road priors are lacking.
If a system highly relies on pre-produced HD maps, the vehicle can obtain relatively complete road priors when facing a covered road; however, when the road changes or the vehicle enters an area with insufficient map coverage, the system needs to handle the discrepancies between the map and reality. Conversely, if the vehicle possesses stronger real-time perception and environmental understanding capabilities, the system can rely on current road information to make judgments. This does not mean the vehicle can just drive wherever it sees fit.
Real-time perception only yields raw environmental information; the system still needs to translate this information into structured understanding usable for driving. This includes understanding where road boundaries are, how lanes extend, whether drivable space exists ahead, how traffic participants are moving, and whether road changes mean the original driving path needs adjustment.
Therefore, after reducing map reliance, what truly increases is the real-time understanding tasks on the vehicle side. This is also why one cannot simply assume that mapless equals reduced technical difficulty. In fact, the two approaches merely shift part of the workload to different locations.
The HD map approach completes the collection and structuring of more road information in advance, which the vehicle can utilize these priors during operation. The real-time perception approach, on the other hand, reduces some upfront road information preparation, allowing the vehicle to take on more environmental understanding tasks during driving. This also explains why reducing reliance on HD maps, for passenger vehicle intelligent driving, is not only to lower map costs but also to improve the system's generalization ability, road freshness, and scaled coverage capability.
04. Final Thoughts
For autonomous driving, HD maps can still provide stable road priors, and in some scenarios, they can also serve as supplementary information and a redundancy measure beyond real-time perception.
Looking at the current market situation, lightweight maps are still being installed in some passenger vehicle models and scenarios, indicating that the industry is not simply moving towards completely abandoning maps, but rather choosing lightweight map, mapless, or map redundancy solutions based on the scenario. It is just that under the trend of becoming mapless, the position of HD maps has changed.
In the past, vehicles needed to rely heavily on maps to tell them what the road should look like; now, with enhanced capabilities in real-time perception, environmental modeling, and driving models, vehicles increasingly need to judge for themselves what the road looks like right now.
#AutonomousDriving #HDMaps #RealTimePerception