Multi-robot systems are widely used in spatially distributed tasks, and their collaborative path planning is of great significance for working efficiency. Currently, different multi-robot collaborative path planning methods have been proposed, but how to process the sensory information of neighboring robots at different locations from a local perception perspective in real environment to make better decisions is still a major difficulty. To address this problem, this paper proposes a multi-robot collaborative path planning method based on geometric graph neural network (GeoGNN). GeoGNN introduces the relative position information of neighboring robots into each interaction layer of the graph neural network to better integrate neighbor sensing information. An expert data generation method is designed for the robot to advance in a single step, by which expert data are generated in ROS to train the network. Experimental results show that the accuracy of the proposed method is improved by about 5% compared to the model based only on CNN on the expert data set. In ROS simulation environment path planning test, the success rate is improved by about 4% compared to CNN and flowtime increase is reduced about 8%, which outperforms other graph neural network models.