CMTF: Control-Point Guided Motion Transformer Framework for Robust Trajectory Prediction

Trajectory prediction in autonomous driving must achieve both high accuracy and strong generalization to cope with complex and dynamic real-world traffic environments. However, existing methods are trained predominantly on in-distribution data from a single dataset, and their performance degrades severely in Out-of-Distribution (OoD) scenarios. To address this limitation, this paper proposes the Control-Point Guided Motion Transformer Framework, CMTF, a general framework for trajectory prediction that improves robustness and interpretability under OoD conditions. CMTF builds on the classical Bernstein polynomial parameterization and represents each trajectory as a small set of control points that carry explicit spatial semantics. The key novelty of this work lies in a shift of the supervision paradigm. Control points are elevated from passively used auxiliary encoding features in prior methods to explicit global supervision signals. These signals operate throughout the entire framework, from structured input augmentation to the auxiliary prediction head and the geometric consistency loss, and guide the model to learn trajectory representations with clear structure and geometric plausibility. Without increasing the amount of training data, CMTF can be integrated into mainstream trajectory prediction models. Cross-dataset evaluations on nuScenes, Argoverse2, and WOMD show a consistent trend of improved OoD performance.