Miformer: A Minus-Inverted Transformer Fed by Historical-Future Interactions for Trajectory Prediction

Encoding historical traffic scenarios is critical for autonomous driving trajectory prediction. However, existing methods face two fundamental challenges. First, neglecting future interactions leads to myopic predictions. Second, although query-centric paradigms capture multimodal interactions, their typically static query design may limit the adaptability of cross-temporal context fusion. To address these limitations, we propose Miformer, a trajectory prediction model that enhances historical-future spatial-temporal context interactions. First, we introduce a time query mechanism that transforms static agent interactions into dynamic temporal modeling by enabling cross-timestep information exchange between different agents and road geometry. Building upon this temporal foundation, we design a historical-future spatial-temporal interaction module (HF-STIM) that leverages iTransformer to model bidirectional dependencies between past and future contexts. To address the redundancy inherent in inverted embeddings, we further propose the minus-inverted Transformer (Mi-Transformer), which incorporates a dynamic residual learning mechanism to eliminate irrelevant dependencies while preserving critical bidirectional information flow. Finally, we employ a DETR-like decoder to generate diverse multimodal trajectory predictions. Experimental results demonstrate that Miformer achieves state-of-the-art performance on the INTERACTION dataset and exhibits competitive performance on the Argoverse dataset. These results highlight its effectiveness in real-world motion prediction. Our code is available at https://github.com/Morphlingxxx/Miformer-Trajectory-Prediction.