This paper develops a new vehicle-robot system, where an autonomous ego-vehicle carries service robots and sends them out to the desired location to perform a specific task (e.g., inspection, cleaning, transportation), and it retrieves them after the task. Specifically, this paper proposes a new pose classification network at the retrieval stage. With such a network, the robot can be navigated to approach the ego-vehicle. When it is close enough, the robot can identify the structured features (e.g., AR codes) and board the vehicle. The constructed network has the advantage of being small-scale and is suitable for the implementation of the mobile robot without sacrificing the positioning accuracy. It allows the robot to use purely visual information to board the vehicle without any additional communication or sensory measures. Hence, it can be applicable to scenarios where the structured features are far away or the lidar is not within the sensory zone. Experimental results are presented to validate the performance of the proposed method.