TL;DR
This paper challenges the assumption that autonomous vehicles (AVs) require expensive, centimeter-accurate High-Definition (HD) maps for motion prediction. By using standard, road-level navigation maps (like OpenStreetMap) and a teacher-student knowledge distillation approach, the authors achieve prediction performance nearly as good as HD-map-reliant models, offering a scalable path for global AV deployment.
What Problem Does This Paper Address?
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 are the "Achilles' heel" of scalability in autonomous driving: they are incredibly expensive to produce, require manual labeling in complex urban areas, and become obsolete the moment a construction crew moves a traffic cone.
The authors ask a pivotal question: Can we predict agent motion using standard navigation maps instead? Navigation maps (like those found in OpenStreetMap) are freely available globally and are much easier to maintain, but they only provide road-level topology rather than precise lane-level geometry. The challenge is bridging the performance gap caused by this loss of precision.
Methods
The authors propose a two-pronged approach to integrate navigation maps into existing prediction frameworks:
- Map Integration: They developed an API to extract road-level directed graphs from OpenStreetMap (OSM) for the Argoverse dataset. They adapted two popular models—LaneGCN and HiVT—by replacing their lane-level queries with road-level queries.
- Knowledge Distillation (KD): To boost the performance of the "navigation-map-aware" model, they use a teacher-student training scheme.
- The Teacher: A model pre-trained on full HD maps.
- The Student: An architecturally identical model that only sees navigation maps.
- The Trick: During training, the student is penalized if its internal "latent embeddings" (the mathematical representation of the scene) differ from the teacher's. This encourages the student to "imagine" or infer the missing lane-level nuances from the coarser road-level data.
This method is model-agnostic, meaning it can be applied to almost any modern prediction architecture that uses a latent fusion step.
Datasets
- Argoverse Motion Forecasting Dataset: A standard industry benchmark containing 201k training and 39k validation sequences from Miami and Pittsburgh.
- OpenStreetMap (OSM): Used as the source for the "near-global" navigation maps.
- Contribution: The authors released their navigation map API for Argoverse on GitHub to allow the research community to experiment with non-HD map-based prediction.
Key Insights and Ideas
- The "Good Enough" Threshold: The most significant finding is that road-level topology provides the "lion's share" of necessary information. Knowing where the road goes is far more important than knowing exactly where the lane lines are painted.
- Distillation as a Bridge: Knowledge distillation acts as a powerful regularizer. Even though the student never sees the HD map during inference, the training process teaches it to extract more relevant features from the coarse navigation map than it would have learned on its own.
- Robustness to Change: Navigation maps are abstractions. Because they don't rely on precise lane markings (which can be obscured by snow, rain, or wear), they are theoretically more robust to fast-changing urban environments.
Key Results
The authors evaluated their approach using standard metrics: minADE (Average Displacement Error) and minFDE (Final Displacement Error).
- Significant Gains over Map-Free: Using a navigation map significantly outperformed "map-free" baselines. For LaneGCN, the Final Displacement Error (k=6) dropped from 1.28m to 1.21m just by adding the navigation map.
- Closing the HD Gap: With Knowledge Distillation, the performance improved further to 1.17m.
- The 9cm Difference: The best navigation-map-based model was only 9cm behind the original HD-map-reliant model in final displacement error. In the context of a car moving at city speeds, this is a remarkably small performance penalty for a massive gain in map availability and cost-reduction.
Limitations and Future Work
- Coarse Geometry Failures: Qualitative results showed that the model occasionally struggles with very acute turns where road-level geometry is too simplified to represent the physical constraints of the turn accurately.
- Feature Set: The current study only used road geometry. Future work could integrate other OSM metadata, such as speed limits, number of lanes, or turn restrictions, to further close the gap with HD maps.
- Beyond Prediction: The authors suggest investigating whether navigation maps can be used for other tasks like "traffic scene reasoning" (understanding the why behind agent behavior).
Summary updated