Abstract: To address the issues of low efficiency,poor real-time performance,and insufficient safety in path planning for
unmanned mining vehicles in complex and dynamic environments,a path planning method integrating Graph Neural Network
(GNN) and attention mechanism is proposed. Firstly,a dynamic graph structure based on road topology is constructed,and
GNN is utilized to deeply extract the features of the road network. Secondly,a multi-head attention mechanism is designed to
capture the long-range dependencies between road sections,and a spatio-temporal attention module is introduced to handle dynamic
environmental information. Finally,an end-to-end training of path planning is achieved based on the reinforcement learning
framework. Simulation experiments show that compared with the traditional A∗ algorithm,the proposed method reduces the
computational time by 45. 3% and shortens the path length by 12. 7%. Compared with the Transformer method,the planning
success rate is increased by 19. 1%,and the obstacle avoidance accuracy is improved by 14. 4%. In actual mine tests,this
method can effectively deal with complex terrains and dynamic obstacles,with an average planning time of only 0. 3 seconds,
providing a reference for the safe and efficient operation of unmanned mining vehicles.

Key words: mining autonomous transport vehicles,path planning,graph neural network,attention mechanism,reinforcement
learning