A Domain-Adaptive Ensemble Framework for Trajectory Prediction in Autonomous Driving

Accurately predicting the future trajectories of traffic participants is crucial for ensuring the safety of autonomous vehicles in complex and dynamic driving scenarios. While existing methods achieve high performance on single benchmarks, their generalization ability often degrades in unseen environments. Training on mixed data from multiple sources is a common strategy to address this, but notable domain distribution discrepancies lead to optimization conflicts and degraded performance. To address this, we propose a domain-adaptive ensemble framework for trajectory prediction which first learns specialized knowledge within multiple, separately-trained expert models, and then adaptively fuses their predictions. To achieve this, we design Gaussian Mixture Model (GMM)-based Out-of-Distribution (OOD) detectors to quantify the distributional shift between the current scene and each expert’s training domain. This quantified shift then guides a Bayesian uncertainty estimation mechanism that dynamically aggregates the experts’ predictions, enabling a principled, weighted fusion. We conducted extensive experiments on three datasets with significant domain shifts: nuScenes, nuPlan, and highD. Our framework demonstrates superior performance over a single model trained on a mixed dataset and naive fusion methods by adaptively identifying the scene context and fusing multiple expert trajectories to generate an optimal, dynamically-weighted prediction.