Abstract
Modern autonomous driving systems rely heavily on understanding "topology"—the complex connectivity and relationships between lanes, traffic lights, and road signs. However, current models are often tethered to specific vehicle sensor configurations, creating a scalability bottleneck. This paper introduces SMART (Scalable Map Priors), a framework that decouples driving topology reasoning from specific sensor suites by leveraging globally available geospatial data. By pre-training on massive datasets of satellite imagery and Standard-Definition (SD) maps, SMART creates a robust map prior that significantly enhances the performance of online driving models, achieving up to a 28% improvement on the OpenLane-V2 benchmark.
Introduction: The Scalability Bottleneck
For an autonomous vehicle to navigate safely, it must perceive more than just the physical location of lane markings; it must understand the underlying logic of the road. This includes knowing which lanes connect at an intersection and how traffic signals govern specific paths. Currently, most topology reasoning models are trained on data captured by specific camera setups. This creates a significant hurdle: any change in camera placement or sensor hardware necessitates the re-collection and re-annotation of massive datasets.
Furthermore, ground-level sensors are inherently limited by occlusions. Large vehicles, buildings, or weather conditions can obscure critical road features. While Standard-Definition (SD) maps like OpenStreetMap provide a global "skeleton" of road layouts, they lack the precision required for high-level autonomy. SMART addresses these challenges by asking: Can we use the vast, sensor-independent world of satellite imagery and HD map archives to teach a model the "logic" of roads before it ever sees a dashboard camera?
The SMART Framework
SMART introduces a two-stage pipeline designed to bridge the gap between global geospatial data and real-time onboard perception.
Stage 1: Learning the Map Prior
In the first stage, the model learns a universal map prior using two primary inputs: SD map polylines (from OpenStreetMap) and satellite imagery (from Mapbox). An SD map encoder processes road attributes such as lane counts and types, while a ResNet-50 backbone extracts visual features from satellite tiles. These modalities are fused into a Bird’s-Eye-View (BEV) representation using cross-attention mechanisms.
The model is trained to predict a lane graph (centerlines and connectivity) using large-scale High-Definition (HD) maps as supervision. Crucially, this stage requires no camera images, allowing the model to be trained on datasets up to 285 times larger than standard sensor-paired datasets.
Stage 2: Online Integration (SMART-OL)
The pre-trained SMART model is then integrated into existing "online" models, such as TopoNet or TopoMLP. In this configuration, the SMART prior provides a stable geometric foundation, while the vehicle's real-time camera feeds handle dynamic elements like traffic lights and temporary road signs that are not visible from space.
Data Scaling and Modalities
The researchers utilized the Argoverse 2 dataset for Stage 1 training, leveraging its 4.3 million HD map segments. For evaluation, the OpenLane-V2 benchmark was used to measure performance in lane detection and traffic element relationships.
A key finding of the study is the existence of a "scaling law" for map priors. Increasing the training data from a base level to 40x resulted in a dramatic jump in topology reasoning accuracy, with the TOPll score rising from 16.2 to 31.9. The study also confirmed that SD maps and satellite imagery are highly complementary: SD maps provide structural topology, while satellite images provide rich textures like lane markings and intersection layouts. Removing either modality significantly degrades the model's performance.
Key Results and Performance
SMART demonstrates state-of-the-art performance across several metrics:
- Overall Performance: Integrating SMART into existing models (SMART-OL) resulted in a 28.4% improvement in the OpenLane-V2 Score (OLS).
- Lane Perception: Even without camera input, SMART outperformed several established camera-based models in lane detection, scoring 37.9 compared to TopoNet’s 28.6.
- Zero-Shot Robustness: In "geo-disjoint" tests—where the model was tested in cities it had never seen during training—SMART-OL improved lane topology reasoning by 148.9% over the baseline.
- Data Efficiency: SMART-augmented models can match the performance of full-data baselines using only 40% of the usual sensor training data.
Limitations and Future Outlook
While SMART represents a significant leap forward, it has inherent limitations. Satellite and SD maps cannot account for dynamic or small vertical elements like traffic lights. Consequently, the model still relies entirely on onboard cameras for real-time traffic element detection. Additionally, because SD and satellite maps are not updated in real-time, the model may occasionally "hallucinate" road layouts if there is temporary construction or recent changes not yet reflected in the geospatial data.
Future work aims to scale SMART into a comprehensive "Map Foundation Model" and explore its utility in downstream tasks such as trajectory prediction and end-to-end motion planning. By decoupling the understanding of road logic from specific hardware, SMART paves the way for more scalable and robust autonomous driving systems.
Summary updated