TrajDiT: diffusion transformer for sequence-aware trajectory generation

Trajectory data is fundamental for intelligent transportation systems but faces challenges in data collection due to privacy constraints, sparsity, and cost. While deep generative models offer solutions, existing diffusion-based approaches often neglect the sequential nature of trajectories. This paper introduces TrajDiT, a novel framework for controllable trajectory generation. TrajDiT integrates a Diffusion Transformer (DiT) architecture with Rotary Position Embedding (RoPE) to explicitly model sequential dependency inherent in trajectory data. This design captures fine-grained sequential dynamics, ensuring trajectories are both geospatially realistic and sequentially consistent. Evaluated on real-world datasets from Porto and Chengdu, TrajDiT outperforms state-of-the-art generative models in trajectory quality. Ablation studies confirm the critical role of RoPE in maintaining sequential integrity. TrajDiT effectively addresses limitations in sequential modelling for trajectory synthesis, providing a robust solution for data-scarce scenarios in urban mobility analysis and intelligent transport applications.