Abstract In this work, we investigate imitation learning and offline reinforcement learning approaches for controlling a tractor-trailer system with a steerable front axle during backward maneuvers. We first develop an expert policy using a kinematic model and Proximal Policy Optimization to generate high-quality demonstration data. Leveraging this dataset, we compare Behavioral Cloning, Implicit Q-Learning, and TD3-BC as representative offline learning methods. Our experiments evaluate sample efficiency, generalization to unseen initial conditions, and model parameter count highlighting trade-offs between pure imitation, value-based optimization, and model complexity. We further analyze trajectory-level behavior and robustness under varying starting configurations. Results show that the applied methods can achieve performance comparable to or exceeding the expert policy while requiring fewer demonstrations, providing insights into the design of data-efficient learning-based controllers for complex trailer systems.
Menz et al. (2026) studied this question.