While these "offline" maps offer high accuracy and rich features, they often struggle to respond promptly to dynamic scenarios such as rapidly changing urban road conditions, sudden traffic control measures, construction barriers, and temporary road closures due to accidents, leading to information delays or outdated information. Real-Time Generated Maps (RTG Maps) have emerged to address this issue. They aim to dynamically collect and construct map features using technologies such as vehicle-to-everything (V2X) and cloud collaboration, enabling autonomous driving systems to make decisions and plans based on richer and more accurate real-time environmental perception, thereby significantly improving driving safety and comfort.
The core concept of real-time generative mapping lies in "drawing as you go," meaning that as vehicles move, map elements are updated in real time using online algorithms based on continuous input from multiple sensor sources such as cameras, LiDAR, millimeter-wave radar, and ultrasonic sensors. These elements include static elements—lane lines, curbs, sidewalk edges, traffic signs, and road textures—as well as dynamic elements—vehicles ahead, pedestrians, bicycles, electric vehicles, temporary obstacles, traffic light phases, and road construction barriers. Compared to traditional static maps, real-time generative maps offer many significant advantages.
First, real-time performance is significantly enhanced. Dynamic elements no longer rely on offline mapping and periodic updates; instead, they are mapped in real-time by the vehicle's perception module combined with online SLAM (Simultaneous Localization and Mapping) technology, achieving near-second-level updates. This allows autonomous driving systems to quickly identify and replan routes when encountering temporary road closures or accidents, avoiding detour errors or safety hazards caused by lagging map information. Second, coverage is broader. Relying on a large fleet of autonomous vehicles or ordinary mass-produced vehicles equipped with advanced driver assistance features, any vehicle on the road can become a "mobile mapping node," uploading perceived environmental elements to the cloud and merging them with data from other vehicles to form a dynamic map service covering the road network. Compared to data collection by a single mapping vehicle, this is more efficient and lower in cost.
To achieve efficient and reliable real-time generative maps, in-depth optimization is needed in the following technical aspects: First, perception fusion and online mapping. Vehicles need to deeply fuse semantic segmentation results from cameras with 3D reconstruction from LiDAR point clouds and velocity information from millimeter-wave radar to extract useful environmental elements. The online SLAM framework must not only solve high-precision positioning but also manage the topology of large-scale scenes, effectively merging and deduplicating new and old elements to ensure that map data is both accurate and compact. Second, transmission and edge computing. Network bandwidth and latency are key constraints between the vehicle and the cloud. By deploying map building services on roadside units (RSUs) or vehicle edge computing nodes, most of the computational load can be completed at the edge, with only key topology changes or map increments uploaded to the cloud, thereby reducing communication requirements and ensuring basic real-time performance even in poor network conditions. Third, spatiotemporal consistency management. With the continuous influx of real-time map elements, optimizing the map storage structure while preventing data conflicts and temporal discrepancies requires the introduction of mechanisms such as spatiotemporal indexing, differential updates, and version control to ensure that the maps read by different vehicles are both the latest versions and constantly matched with their own positioning.
In autonomous driving systems, real-time generative maps primarily serve the following core functions. First, precise positioning. While onboard GPS and inertial navigation systems can provide approximate locations, meter-level or even centimeter-level positioning still relies on map matching technology. When vehicles use lane lines and curb features from real-time generative maps for matching, sub-meter or even centimeter-level positioning errors can be achieved within a local map area, thus supporting higher-precision control and obstacle avoidance. Second, path planning. Based on real-time generated road network topology and dynamic obstacle information, the planning module can output safer and more efficient driving paths. Especially in congested urban areas or construction zones, planning algorithms can utilize the latest map data to implement lane-changing, detour, or low-speed following strategies, improving driving smoothness and road traffic efficiency. Third, environmental modeling and decision-making. Real-time maps provide scene context for perception and prediction modules. By associating the positions of vehicles and pedestrians with map features, the behavioral intentions of traffic participants can be more accurately inferred, and more reliable action sequences can be generated at the decision-making level. For example, at intersections, based on traffic light status and pedestrian gathering area data, it can determine whether to slow down or stop.
Currently, various real-time generative map solutions have emerged in the industry. Waymo's fleet not only uploads static high-precision maps but also integrates detected traffic control information into its map service in real time, enabling dynamic event annotation. Tesla, in its FSD (Full Self-Driving) system, uses on-vehicle neural network models to output structured road scenes online and merges them with offline maps updated via OTA, improving recognition accuracy and stability. Mobileye has deployed the RoadBook system on the RSU side, which uses roadside sensors and cloud collaboration to provide passing vehicles with the latest road condition information in all directions. Different solutions have different focuses in their architectural design, but they all innovate around the core idea of "dynamic map construction based on large-scale vehicle-to-everything (V2X) connectivity."
While real-time generative maps have shown great potential, they still face challenges in large-scale commercialization. The first is data privacy and security. Real-time maps involve massive amounts of vehicle trajectory and surrounding environmental data. Ensuring map quality while protecting user privacy requires technologies such as differential privacy and federated learning, as well as robust regulations and standards. The second is network capacity and computing costs. In high-density urban environments, the simultaneous uploading and downloading of map increments by a massive number of vehicles can put enormous pressure on mobile networks. The deployment and operation costs of edge computing nodes are also significant, necessitating intelligent scheduling and hierarchical storage strategies to dynamically adjust the allocation of network and computing resources. Finally, there is the consistency and robustness of multi-source data. Data from different types of sensors and different vehicle brands differ in format, accuracy, and timing. Achieving cross-platform fusion at the cloud or edge places higher demands on the system's algorithm design and engineering implementation.
Real-time generative maps will evolve towards greater intelligence, automation, and ecological sustainability. With the continuous improvement of 5G/6G networks, large-scale model inference capabilities, and edge computing power, the latency of map building and updating will be further reduced, resulting in wider coverage and finer-grained updates. Furthermore, distributed map training mechanisms based on federated learning and privacy-preserving computation will allow fleet sizes and cloud platforms to share models and knowledge without leaking raw data, enhancing overall map service capabilities. In addition, high-precision time-series maps deeply integrated with in-vehicle sensors will provide richer scene adaptation capabilities for autonomous driving in mixed traffic environments, such as pedestrian behavior prediction and dynamic planning of non-motorized lanes.
Real-time generative maps are a crucial element in the transformation of autonomous driving from "offline mapping" to "network-wide collaboration and real-time perception." They not only enhance vehicles' environmental adaptability but also provide strong support for improving the safety, comfort, and efficiency of autonomous driving. With the continuous maturation of related technologies and the industrial ecosystem, real-time generative maps are expected to become one of the infrastructures of intelligent connected transportation systems, laying a solid foundation for achieving true full-scenario autonomous driving.
