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.

翻译：多机器人系统在空间分布式任务中应用广泛，其协同路径规划对工作效率具有重要意义。目前已提出多种多机器人协同路径规划方法，但在真实环境中如何从局部感知视角处理不同位置邻居机器人的感知信息以做出更优决策仍是重大难题。针对该问题，本文提出基于几何图神经网络的多机器人协同路径规划方法。GeoGNN将邻居机器人的相对位置信息引入图神经网络的每次交互层，以更好整合邻居感知信息。设计了单步前进的专家数据生成方法，并在ROS中生成专家数据训练网络。实验结果表明，在专家数据集上，本文方法相比于仅基于CNN的模型准确率提升约5%。在ROS仿真环境路径规划测试中，相较于CNN方法成功率提升约4%，通行时间增加量降低约8%，性能优于其他图神经网络模型。