ForceTraj: Modeling Realistic Intention and Heterogeneous Interactions for Multi-modal Trajectory Prediction

Vehicle trajectory prediction is a fundamental yet challenging task for autonomous driving, where accurate future motion estimation is essential for safe planning and collision avoidance in dynamic traffic scenarios. However, many existing methods overlook critical lane-level environmental cues or treat the influences of all neighboring vehicles equally, limiting their ability to model realistic driving intentions and heterogeneous interactions. To overcome these limitations, we propose ForceTraj, a multi-modal trajectory prediction framework that explicitly models realistic driving intentions affected by environmental context and interaction heterogeneity. We first introduce an intention perception module that encodes spatial priors indicative of lane-changing intentions. Next, an interaction force module modulates interaction strengths conditioned on the inferred intentions. Then we fuse intention and interaction encodings, while a spatial relation module further enhances spatio-temporal dependency modeling. Finally, a mixture-of-experts (MoE)-based multi-modal decoder enables the model to generate the probability distribution of future trajectories. Experiments on NGSIM and highD show that ForceTraj achieves significant improvements over baselines in trajectory prediction accuracy. Our code is available at https://anonymous.4open.science/r/Forcetraj-8A1E .