Automatic posture adjustment is a key application in the domain of intelligent vehicles, playing a fundamental role in enhancing safety and optimizing vehicle maneuvering operations. Articulated Vehicles (AVs), due to their high degrees of freedom and the complex, nonlinear dynamics resulting from the joint between the tractor and trailer, present a more challenging control problem compared to rigid vehicles. The objective of this research is to design and simulate an automatic control system for articulated vehicle posture adjustment utilizing a Deep Reinforcement Learning (DRL) framework. This system can also serve as a foundation for more advanced applications, such as autonomous parking.

In this study, the precise modeling of articulated vehicle dynamics and the jackknifing phenomenon was initially carried out. The developed model was validated using the specialized software, TruckSim. Subsequently, to reduce computational complexity, the learning process was segmented into two distinct phases: maneuver preparation and final posture adjustment. For training the intelligent agent, Deep Deterministic Policy Gradient (DDPG) and Twin Delayed DDPG (TD3) algorithms were employed, which were optimized with neural networks comprising three to five hidden layers. Evaluation results indicated that the TD3 algorithm, owing to its superior ability to maintain the stability of the learning process, outperformed DDPG. Ultimately, the proposed control system, with the optimal structure for each phase, achieved success rates of 96.6% in the preparation phase and 94.6% in the final adjustment phase, thereby confirming the high efficiency and reliability of the DRL-based system in addressing the control challenges of articulated vehicles.