Bi-LSTM-Transformer Based Trajectories Tracking Control Method for Autonomous Vehicles Considering the Coupling of Lateral and Longitudinal Dynamics

During trajectory tracking, the nonlinear coupling between a vehicle's lateral and longitudinal dynamics poses greater challenges to control accuracy and real-time performance. This paper proposes a Bi-LSTM-Transformer based trajectory tracking control method for autonomous vehicles, incorporating coupled lateral-longitudinal dynamics through data-driven methodology. First, a nonlinear 14-degree-of-freedom vehicle dynamics model is established as the foundation for training the trajectory tracking controller. Afterwards, a Bi-LSTM-Transformer network architecture is designed to capture the temporal and spatial characteristics of the target trajectory. This framework establishes an effective mapping between the target trajectory and the corresponding vehicle control commands, enabling the generation of control command sequences over a predefined future time horizon. Subsequently, a joint training framework for the Bi-LSTM-Transformer network and the 14- degree-of-freedom vehicle model is proposed, enabling effective training of trajectory tracking controller without requiring labeled vehicle control commands. Finally, the real-time performance and effectiveness of the proposed trajectory tracking control method are validated through offline simulations and hardware-in-the-loop experiments under various driving scenarios, such as variable-speed lane changing, double lane changing, and sinusoidal maneuvers. The Simulink-Carsim co-simulation and experimental results demonstrate that the BiLSTM-Transformer trajectory tracking controller enhances the smoothness of front wheel angle and longitudinal torque control commands by an average of 34.61% and 43.79%, respectively, compared to model predictive control and sliding-mode variablestructure controllers, while maintaining trajectory tracking accuracy. Regarding real-time performance, the proposed controller demonstrates significant advantages over the model predictive control-based trajectory tracking controller and achieves a 9.68% improvement compared to the sliding-mode variable-structure controller.