Abstract: Aiming at the problems in remote control of tunneling equipment, such as low decision-making ability of equipment, less tunneling efficiency, and large security risks, a decision-making control method of tunneling robot driven by digital twin is proposed. By analyzing and comparing the current research situation of digital twin technology in the field of coal mine, the framework of digital twin-driven tunneling robot decision control system is designed, including six modules：physical space, virtual space, twin data, planning layer, control layer and execution layer. In this system, the virtual prototype can make planning decisions autonomously and control the synchronous motion of the physical prototype remotely. Firstly, the local obstacle avoidance strategy in unstructured environment is studied based on virtual reality technology. Motion control model and sensor observation model of tunneling robot are established, and obstacles in roadway are reconstructed in virtual environment by laser radar, Ray-Col method is used to detect the collision between robot and obstacle, which lays the foundation for robot path planning decision. Secondly, the global path planning method based on virtual agent is studied by combining deep reinforcement learning technology, the Muti-PPO algorithm based on the improved PPO algorithm is proposed, and the virtual agent of tunneling robot is established through the reward and punishment mechanism, and training in Unity3D platform, The training results show that compared with PPO algorithm and SAC algorithm, the average reward value of Muti-PPO algorithm is increased by 13.82% and 11.31% respectively. Standard deviation decreased by 17.85% and 16.81% respectively; the maximum reward value is increased by 0.14% and 0.43% respectively, and its performance is optimal among the three algorithms. Finally, a decision control platform is built to send the decision instructions generated in the virtual space to the end-effector of the physical prototype, and drive the synchronous change of the virtual prototype through the sensor data of the physical prototype. According to the planning decision, bidirectional mapping and remote control functions of the system, path planning experiment and virtual-real co-movement experiment are designed to verify it. The experimental results of path planning show that the error between the end point of virtual agent path planning and the target point is within 1.2 cm under three different complexity conditions, and the control information can be transmitted to the physical space to control the robot motion remotely. The experimental results show that the virtual prototype and the physical prototype are consistent in the roadway during the operation of the tunneling robot. This method realizes the new unmanned decision-making control mode of “data-driven, two-way mapping, collision detection, autonomous decision-making, and man-machine cooperation”, which provides a new idea for the intelligentization of tunneling equipment.