KANTraj: Kolmogorov–Arnold Network-Based Recursive Vehicle Trajectory Prediction With Driving Style Awareness

Accurate trajectory prediction of surrounding vehicles is essential for safe and efficient decision-making in intelligent transportation systems and industrial autonomous environments. Most existing methods adopt a single-shot open-loop inference paradigm, where interaction representations remain fixed during decoding and driving style diversity is often neglected. To overcome these limitations, we propose KANTraj, a recursive trajectory prediction framework that jointly models trajectory dynamics, interaction evolution, and driving style. Three dedicated encoders extract trajectory, interaction, and style features from historical data, while a Kolmogorov–Arnold Network-based vehicle kinematic model enables compact, interpretable, and nonlinear motion modeling in the decoder. Unlike traditional open-loop approaches, our recursive decoder iteratively refines predictions by updating state and interaction representations based on intermediate future states, achieving dynamic adaptation in complex traffic. Extensive experiments on the NGSIM and HighD datasets demonstrate that our method achieves superior performance compared to state-of-the-art approaches, reducing root-mean-square error by up to 38% on the two datasets at a long-term prediction horizon of 5 s.