Autonomous Driving Map (HD/LD/SD MAP, Online Reconstruction, Real-time Generative Map) Industry Report 2025

Research on Autonomous Driving Maps: Evolve from Recording the Past to Previewing the Future with 'Real-time Generative Maps'

'Mapless NOA' has become the mainstream solution for autonomous driving systems. This solution reduces the reliance on offline HD maps whose development has encountered challenges. The so-called 'mapless' essentially means the shift from 'map prior' to 'real-time map construction' and then further development into 'world models', while ADAS algorithms tend to be 'data-driven' instead of being 'rule-driven'.

A mapless solution, very similar to the early SLAM technology, actually builds a vector map online and then matches it with offline LD maps to obtain positioning and navigation information at the same time. The early SLAM technology relied heavily on LiDAR. As BEV emerges, SLAM technology has been gradually eliminated, but it is still used in scenarios such as underground parking lots.

The evolution of autonomous driving maps:

• Before 2022: The industry chain focused on HD maps that value geometric accuracy, while traditional ADAS algorithms relied on preset rules to process environmental perception
• 2023-2024: With the development of mapless NOA, lightweight maps (LD maps) with topology, semantics and freshness were promoted and applied

After 2025: With the introduction of new technologies such as 3D Gaussian sputtering and NeRF (Neural Radiance Fields), autonomous driving maps will 'preview the future” instead of only 'recording the past'. 'World models' extract spatiotemporal patterns from massive driving data through self-supervised learning, integrate multimodal sensor data (cameras, LiDAR, etc.) and real-time crowd-source data, build a dynamically updated environmental knowledge base, and accomplish online reasoning of road topology, semantic information and traffic rules.

'World models' leverage historical scenario information and preset conditions to predict the future changes in intelligent driving scenarios and the response of the ego vehicle.

Development trends of autonomous driving maps: Low-cost automated mapping, application of vectorized HD map construction technologies such as MapTR and VectorMapNet

Baidu MapAuto 6.5 is the first 3D lane-level map and all-scenario human-machine co-driving map in China, providing comprehensive data services. Baidu MapAuto 6.5, based on Baidu's integrated data collection vehicles, multi-source data input (closed loop of automotive and roadside data), and map generation foundation models with billions of parameters, has improved the efficiency of map production exponentially, effectively supported the rapid updates of Baidu map data, and offered powerful and comprehensive data services.

Baidu MapAuto 6.5 can provide three types of data: SD (navigation maps), LD (lightweight autonomous driving maps) and HD (HD maps). In March 2025, Leapmotor released LEAP 3.5, which is a technical architecture equipped with the LD data of Baidu Maps.

Low-cost automated mapping is an important development direction of Baidu Maps, with core technologies including BEV static road scenario reconstruction and automated feature extraction.

Baidu's BEV static road scenario reconstruction uses Instance Query and Point Query similar to Huazhong University of Science and Technology's MapTR to detect road elements and element outline fixed points. It adopts a method similar to the Auto-regressive decoder in Tsinghua's VectorMapNet to output the topological relationship between feature points.

MapTR is suitable for real-time mapping of urban roads, L2+ ADAS, and embedded platforms with limited hardware resources. Its fixed-length setpoint output is convenient for connection with the planning and control module

VectorMapNet is ideal for scenarios likemodeling of complex interchanges on highways, map generation research in the field of scientific research, and special scenarios that require variable-length fine modeling (such as construction areas).

Development trends of autonomous driving maps: Integration with driving world models (DWMs)

NavInfo has proposed to add the spatiotemporal cognition capability of maps to the intelligent driving technology driven by world models, that is, 'let world models inherit the spatiotemporal cognition of maps' - 'Maps have evolved from static layers to dynamic data engines that are indispensable in the world-model-driven stage. They are irreplaceable 'prior sensors' in application scenarios such as improving the intelligence level of a single vehicle, reducing computing power constraints and responding to emergency warnings.'

DWMs are the core components of the next-generation autonomous driving systems. By predicting the spatiotemporal evolution of dynamic driving scenarios, they help vehicles perceive the environment more accurately, understand interaction logic, and optimize decision-making.

DWMs build continuous learning and prediction capabilities for the physical world by integrating HD map data, real-time sensor information (such as cameras, LiDAR), vehicle status data (such as speed, steering), and external environment data (such as traffic flow, weather). The goal is to enable autonomous driving systems to secure the trinity of 'understanding, prediction, and planning' through a closed data loop.

Core functions of DWMs:

• Environmental understanding: Accurately locate the vehicle position through autonomous driving maps and real-time perception data, and recognize key information such as lane lines, traffic signs, and obstacles.
• Dynamic prediction: Predict the behavior trajectory of other traffic participants (vehicles, pedestrians), and predict potential risks (such as cutting in, sudden braking).
• Global planning: Generate the optimal driving path and driving strategy based on long-term environment simulation (such as generalization of scenarios under different weather and road conditions).

Technical features of DWMs:

• Continuously optimize the models based on data, continuous input of massive high-quality data and AI algorithms (such as deep learning and reinforcement learning).
• Achieve closed-loop iteration and self-evolution of the models through a complete closed loop of data collection → model training → simulation verification → deployment optimization.
• Integrate reality with virtuality and accelerate model generalization by combining simulation environments (such as digital twins) with real road test data.

Core value of DWMs:

Scenario deduction: Generate the physical rationality and spatiotemporal consistency of future scenarios based on historical observations, and support autonomous driving systems to predict potential risks (such as bizarre accidents (for example, when there is a vehicle or obstacle blocking the view ahead, a non-motorized vehicle or pedestrian suddenly jumps out from the roadside, and the driver fails to avoid it in time, often causing an accident), dynamic changes in construction areas)

• Multimodal fusion: Integrate multimodal data such as 2D images, 3D point clouds, and Occupancy grids to improve environmental modeling accuracy (such as 98.7% BEV geometric consistency in nuScenes data set tests)
• Decision-making optimization: Achieve human-like driving capabilities through reinforcement learning and prediction, real difference fine-tuning (The measured traffic efficiency on Beijing's Fifth Ring Road increased by 28%).

Development trends of autonomous driving maps: OEMs explore and deploy NeRF technology in autonomous driving map reconstruction

• At present, many OEMs have begun to explore or deploy NeRF technology in the field of autonomous driving maps, especially in dynamic scenario reconstruction and HD map generation.
• NeRF technology can reconstruct 2D images into 3D scenarios, and then produce HD maps to achieve high-precision vehicle positioning and map matching
• NeRF technology can synthesize complex autonomous driving scenarios, enrich autonomous driving training data, and help autonomous driving systems perform efficient data enhancement
• NeRF technology can simulate harsh scenarios such as extreme weather and serious traffic accidents, and use simulated data to restore real harsh scenarios to improve the safety of autonomous driving

The AD Max 3.0 of Li Auto has built a triple perception architecture consisting of 'static BEV + dynamic BEV + NeRF enhanced occupancy'. By deeply combining NeRF technology with occupancy networks, it handles insufficient long-distance perception resolution which exists in traditional pure vision solutions:

Static BEV network: Transformer fuses data from multiple cameras to generate a bird's-eye view of the road structure. When some cameras fail, NeRF helps reconstruct the road edges and lane lines in the missing areas.

Dynamic BEV network: Thanks to the spatiotemporal attention mechanism tracking traffic participants and NeRF's spatiotemporal continuity modeling, the speed and acceleration estimation error of moving objects is less than 0.3m/s.

Occupancy network upgrade: The original Occupancy output resolution improves from 0.2m to 0.1m, sub-pixel details are generated through NeRF's radiation field rendering, and 30cm high curbstones and 5cm diameter manhole covers can be recognized

OEMs such as Xpeng, Mercedes-Benz, and Li Auto have taken the lead in mass production and application of NeRF technology, while Tesla, BMW, etc. are exploring deeper application through technical cooperation. In the future, with the improvement of hardware computing power (such as the Blackwell architecture) and open source ecology, NeRF is expected to become the underlying standard technology of autonomous driving maps, promoting the industry to evolve towards 'real-time generative maps'.