Semantic Control Points for Structured Trajectory Representation: Lightweight Geometric Knowledge Injection for Cross-Domain Robustness

Trajectory prediction serves as a core module in intelligent driving systems and is critical for downstream planning and control tasks. However, existing methods predominantly trained on in-distribution samples from single datasets exhibit significant performance degradation when applied to Out-of-Distribution (OoD) scenarios. This paper presents a novel approach using semantic control points for structured trajectory representation with lightweight geometric knowledge injection to enhance cross-domain robustness. Specifically, we parameterize trajectories using a small number of spatially semantic control points via Bernstein polynomials, and incorporate auxiliary prediction heads with geometric consistency loss to explicitly guide the model in learning physically plausible trajectory structures. This structured modeling approach not only strengthens the model’s focus on critical geometric features but also provides interpretability: control points serve as structural anchors reflecting future turning intentions, facilitating the capture of potential behavioral changes. Experiments demonstrate that our method can be seamlessly integrated into mainstream trajectory prediction models without increasing training data, significantly improving their OoD generalization performance across multiple datasets including nuScenes, Argoverse2, and Waymo Open Motion Dataset.