DyTTP: Trajectory Prediction with Normalization-Free Transformers

DyTTP, a trajectory prediction method, employs normalization-free Transformers with DynamicTanh (DyT) in the backbone, replacing layer normalization for a simpler, more stable architecture. It uses a snapshot ensemble strategy with cyclical learning rate scheduling to aggregate diverse model hypotheses, enhancing accuracy, speed, and robustness in various driving scenarios. DyTTP demonstrates the potential of normalization-free transformer designs in autonomous vehicle trajectory forecasting, offering a scalable solution for real-time predictions.