Deep reinforcement learning has shown its advantages in real-time decision-making based on the state of the agent. In this stage, we solved the task of using a real robot to manipulate the cube to a given trajectory. The task is broken down into different procedures and we propose a hierarchical structure, the high-level deep reinforcement learning model selects appropriate contact positions and the low-level control module performs the position control under the corresponding trajectory. Our framework reduces the disadvantage of low sample efficiency of deep reinforcement learning and lacking adaptability of traditional robot control methods. Our algorithm is trained in simulation and migrated to reality without fine-tuning. The experimental results show the effectiveness of our method both simulation and reality. Our code and video can be found at https://github.com/42jaylonw/RRC2021ThreeWolves and https://youtu.be/Jr176xsn9wg.