In response to the problems of long debugging cycle, high debugging cost, high safety risk, and difficult quality evaluation in the current automatic forming and cutting control of roadways, a digital twin driven virtual debugging method for roadway automatic forming and cutting is proposed. Firstly, by using real-time appearance mapping (RTAP-MAP) technology to reconstruct the three-dimensional environment of the roadway, a control system model of the roadheader is constructed to form a virtual debugging environment. The virtual sensor technology is used to achieve precise mapping from physical space to virtual environment state. A performance evaluation method for roadway automatic forming cutting has been established to address the problem of difficulty in quantitatively evaluating the quality of section forming. Based on the recording of the section forming cutting control process in the data transmission center, the evaluation indicators of section forming precision, cutting efficiency and the number of oil cylinder switches, hard rock cutting adjustment, and over excavation and under excavation are mainly calculated. This provides precise feedback signals for the iterative optimization of deep learning algorithms. An automatic cutting control strategy that integrates reinforcement learning is proposed to improve the adaptability and precision of automated operations. To verify the effectiveness and accuracy of the virtual debugging method, an automatic control experimental platform for roadheader is built. The virtual debugging system is applied to the automatic control program for forming cutting of excavation roadway. The virtual simulation results show the following points. ① The maximum positioning errors of the X, Y, and Z axes of the debugged software at the control key points are 74.8, 72.93, 123.67 mm, respectively. It indicates that the positioning precision of the virtual debugging method meets the performance requirements. ② The trajectory of the virtual prototype is basically consistent with that of the physical prototype, indicating that this debugging method has achieved mapping to the physical space. The application result shows the following points. ① The reinforcement learning controller adapts to complex environments in virtual excavation testing, effectively converts virtual sensor inputs into precise control instructions, and verifies the feasibility of simulation reality transfer training. By processing real-time feedback on excavation precision and avoiding over excavation and under excavation, the controller learns and optimizes the strategy. ② The improved section forming cutting control performance has been improved. According to the control quantity timestamp records in the database, it takes 126 seconds, which is 8 seconds less than before the improvement. ③ After improvement, the maximum error in tracking the end trajectory of the cutting section is 6.0 mm, which is 0.3 mm lower than before. This avoids the under excavation caused by the shaking of the cutting trajectory and makes the trajectory and section smoother.