Dynamic environment path planning based on hybrid pre-training algorithm
Xinyu Yang, Taotao Jin, Weihao Xu, Chuanyue Qi, He Ma · Engineering Research Express · 2025
Abstract To address the challenges of real-time path planning in a dynamic environment and balance motion feasibility and global optimality, this paper proposes a hybrid pre-training algorithm. This method combines the traditional A* heuristic search with deep learning pre-training to develop a three-dimensional path planner that can learn the environmental topology and respond to dynamic obstacles. The key innovations include discretization modeling of the state space and the composite heuristic function based on reed-shepp, which effectively reduce node redundancy and improve path smoothness. The spatio-temporal obstacle path coordination was achieved by using the incremental search mechanism of pre-trained knowledge, and a dynamic response time of 76 milliseconds was realized. The simulation results show that, compared with the traditional A* and LSTM-related algorithms, the paths generated by this algorithm are significantly shortened and can accurately and safely identify and avoid various obstacle environments. Meanwhile, physical tests verify the navigation capability of this system in narrow channels and dynamic scenarios. This algorithm provides the optimal and practical solution for autonomous systems in complex environments.