Tailored meta-learning for dual trajectory transformer: advancing generalized trajectory prediction

Abstract

Trajectory prediction has become increasingly critical in various applications such as autonomous driving and robotic navigation. However, due to the significant variations in trajectory patterns across different scenarios, models trained in known environments often falter in unseen ones. To learn a generalized model that can directly handle unseen domains without requiring any model updating, we propose a novel tailored meta-learning-based trajectory prediction model called DTM. Our approach integrates a dual trajectory transformer (Dual_TT) equipped with an agent-consistency loss, facilitating a comprehensive exploration of both individual intentions and group dynamics across diverse scenarios. Building on this, we propose a tailored meta-learning framework (TMG) to simulate the generalization process between source and target domains during the training phase. In the task construction phase, we employ multi-dimensional labels to precisely define and distinguish between different domains. During the dual-phase parameter update, we partially fix crucial attention mechanism parameters and apply an attention alignment loss to harmonize domain-invariant and specific features. We also incorporate a Serial and Parallel Training (SPT) strategy to significantly enhance task processing and the model’s adaptability to domain shifts. Extensive testing across various domains demonstrates that our DTM model not only outperforms existing top-performing baselines on real-world datasets but also validates the effectiveness of our design through ablation studies.