Abstract
The prediction of surrounding agents' motion is a key for safe autonomous driving. In this paper, we explore navigation maps as an alternative to the predominant High Definition (HD) maps for learning-based motion prediction. Navigation maps provide topological and geometrical information on road-level, whereas HD maps additionally have centimeter-accurate lane-level information. As a result, HD maps are costly and time-consuming to obtain, while navigation maps with near-global coverage are freely available. We describe an approach to integrate navigation maps into learning-based motion prediction models. To exploit locally available HD maps during training, we additionally propose a model-agnostic method for knowledge distillation. In experiments on the publicly available Argoverse dataset with navigation maps obtained from OpenStreetMap, our approach shows a significant improvement over not using a map at all. Combined with our method for knowledge distillation, we achieve results that are close to the original HD map-reliant models. Our publicly available navigation map API for Argoverse enables researchers to develop and evaluate their own approaches using navigation maps.
Introduction
Most state-of-the-art motion prediction models rely heavily on High-Definition (HD) maps. These maps provide centimeter-accurate details of lane geometries, intersections, and traffic signs. However, HD maps represent a significant bottleneck for the scalability of autonomous driving: they are expensive to produce, require manual labeling, and are difficult to maintain in dynamic urban environments.
This research investigates whether standard, road-level navigation maps (such as those from OpenStreetMap) can serve as a viable alternative. While navigation maps offer near-global coverage and are easier to maintain, they lack the precise lane-level geometry of HD maps. The goal is to bridge this performance gap through architectural integration and advanced training techniques.
Methodology
The proposed approach integrates navigation maps into existing prediction frameworks through two primary mechanisms:
1. Map Integration
The authors developed an API to extract road-level directed graphs from OpenStreetMap (OSM) for the Argoverse dataset. They adapted popular architectures, specifically LaneGCN and HiVT, by replacing their high-resolution lane-level queries with coarser road-level queries. This allows the models to utilize topological information even when precise lane data is unavailable.
2. Knowledge Distillation (KD)
To recover the performance lost by moving from HD to navigation maps, a teacher-student training scheme is employed. A "Teacher" model is pre-trained on full HD maps with frozen weights. A "Student" model, which only has access to navigation maps, is then trained to mimic the teacher.
During training, the student is penalized if its internal latent embeddings differ from the teacher's. This encourages the student to infer missing lane-level nuances from the coarser road-level data. This method is model-agnostic and can be applied to various prediction architectures that utilize a latent fusion step.
Experimental Results
The approach was evaluated on the Argoverse Motion Forecasting Dataset using standard metrics: Average Displacement Error (minADE) and Final Displacement Error (minFDE).
- Performance Gains: Integrating navigation maps significantly outperformed map-free baselines. For the HiVT model, the Final Displacement Error (k=6) improved from 1.28m to 1.21m simply by adding the navigation map.
- Closing the Gap: With the addition of Knowledge Distillation, performance improved further to 1.17m. Remarkably, the best navigation-map-based model was only 9cm behind the original HD-map-reliant model (1.08m) in final displacement error.
- Information Threshold: The results suggest that road-level topology provides the most critical information for prediction. Knowing the general path of the road is often sufficient for accurate forecasting, even without precise lane markings.
Conclusion
This work demonstrates that navigation maps are a powerful and scalable alternative to HD maps for motion prediction. While HD maps provide superior accuracy, the performance penalty of using navigation maps is surprisingly small—approximately 9cm in final displacement—when paired with knowledge distillation. Future work will explore integrating additional metadata from OpenStreetMap, such as speed limits and turn restrictions, to further narrow the gap with HD-reliant systems.
Summary updated