CNN-based path planning on a map

Daniele Sartori, Danping Zou, Ling Pei, Wenxian Yu

Abstract: An approach for fast 2D path planning within a map is presented in this paper. The proposed solution relies on a Convolutional Neural Network to estimate the path generated with a traditional Theta* algorithm. The network receives as input a top-down image of the environment with the obstacles distribution and the location of the start and goal points. It provides as output the coordinates of the path waypoints. This approach is validated on three environment datasets representing realistic indoor scenarios. It is demonstrated that, according to the environment complexity, with the help of a refinement algorithm safe and reliable paths can be created within very limited computation time and with high success rate.