The paper introduces a novel approach called conditional imitation learning, which allows autonomous vehicles to be directed by high-level commands while maintaining sensorimotor coordination. The method addresses the limitation of traditional imitation learning, where trained policies cannot be controlled at test time. By conditioning the imitation learning process on high-level commands, the trained controller can handle sensorimotor tasks while responding to navigational instructions. The approach is evaluated in realistic simulations and on a 1/5 scale robotic truck, demonstrating its effectiveness in complex urban driving scenarios. The paper also discusses the importance of data augmentation and noise injection for improving generalization and stability. The results show that the proposed command-conditional formulation significantly enhances performance compared to standard and goal-conditioned imitation learning methods.The paper introduces a novel approach called conditional imitation learning, which allows autonomous vehicles to be directed by high-level commands while maintaining sensorimotor coordination. The method addresses the limitation of traditional imitation learning, where trained policies cannot be controlled at test time. By conditioning the imitation learning process on high-level commands, the trained controller can handle sensorimotor tasks while responding to navigational instructions. The approach is evaluated in realistic simulations and on a 1/5 scale robotic truck, demonstrating its effectiveness in complex urban driving scenarios. The paper also discusses the importance of data augmentation and noise injection for improving generalization and stability. The results show that the proposed command-conditional formulation significantly enhances performance compared to standard and goal-conditioned imitation learning methods.