StyleFormer: Multi-Agent Joint Trajectory Prediction and Planning in Urban Environments With Driving Style Awareness

Accurately inferring the driving intentions of neighboring vehicles is a critical challenge for autonomous vehicles (AVs) when handling complex interactions in urban environments. These interactions are complicated by diverse driving styles, which AVs often struggle to interpret, leading to overly cautious driving strategies. In this work, a trajectory prediction and planning framework, StyleFormer, is proposed which considers vehicles’ driving styles. The model classifies short-term driving styles using an unsupervised method and employs a vectorized representation to integrate map features, agent states, and driving styles. A Transformer-based attention mechanism is used to model interactions and intentions, enabling joint prediction of future trajectories for surrounding vehicles and multimodal trajectory generation for AVs. StyleFormer further adapts planned trajectories to different driving style pReferences and leverages diffusion-based optimization to enhance safety and feasibility. It effectively models multi-modal driving behaviors and ensures trajectory quality without relying on rule-based post-processing. Validated on the Argoverse 1 and nuScenes datasets, StyleFormer achieves superior performance in open-loop trajectory prediction and planning, and reaches a 96.33% success rate in closed-loop planning. The results demonstrate enhanced prediction accuracy, efficient and safe planning in complex scenarios, and strong generalization across diverse urban environments.