<p>Path planning and obstacle avoidance for shotcrete robots in underground mines are challenging problems. The environment is complex and dynamic. Safety and stability are difficult to ensure. To address these challenges, this paper proposes a hybrid path planning method based on spatiotemporal risk prediction, corresponding to two modules. Unlike end-to-end decision methods, this approach limits deep learning to an environment prediction module. The module provides prior risk information for planning. The path planning module is still executed by traditional and interpretable algorithms. In this method, a 3D spatiotemporal convolutional network is designed. The network predicts future multi-frame dynamic risk fields through regression. It captures the spatiotemporal distribution of environmental risk. The predicted risk is then mapped into a time-varying cost map. This map is integrated into a hierarchical planning framework. In the planning stage, A* generates a global path that is aware of risk. DWA performs local trajectory optimization and real-time obstacle avoidance. This design allows prediction information and motion planning to work together in a unified framework. To validate the proposed method, simulation experiments were conducted, and the proposed method was compared with several ablation variants. The results show that the proposed hybrid method achieves a higher success rate and better stability in the constructed scenarios, which include multiple tunnel topologies. The success rate reaches 93.1%, which is higher than those of the baseline methods. In addition, the results of environment generalization and robustness experiments demonstrate that the integration of spatiotemporal risk prediction and hierarchical planning can improve the path planning performance of underground mobile robots in complex dynamic environments.</p>