In order to solve the problem that the motion planning of manipulator in complex constrained environment is very important for the realization of universal service robot, a transformer-graph neural network (T-GNN) framework is proposed. And aim to solve that problem of low computation efficiency of the traditional sampling-based planning method and the limitation of local dependence model in the existing learning-based method. T-GNN integrates the Transformer module for capturing global geometric dependencies, and utilizes GNN for iterative latent graph optimization to achieve efficient path exploration. The experimental results show that T-GNN significantly reduces the number of collision detection, improves the planning efficiency, achieves a high success rate in high-dimensional scenes, and maintains a good balance between path optimality and real-time performance under different environmental complexities.