Robot Path Planning with Low Learning Cost Using a Novel K-means-based Pointer Networks

Abstract

Robot-path-planning is an important research that seeks the shortest path to optimize the motion cost for robots. In robot-path-planning, the computational time will significantly increase if the moving targets for a robot rise largely, while the current algorithms for the shortest path planning may be invalidated due to large input data. This work thus proposes a hybrid algorithm, called the k-means-based pointer network, to tackle the problem mentioned above. By combining the k-means clustering and pointer network, unsupervised and supervised learning respectively, this work demonstrates how to lower the learning cost drastically with smaller training data. The simulation results show that the computational time cost of the Held-Karp algorithm grows significantly when the input size increases in some amount, while the proposed algorithm climbs slightly during the increments of input size because of using smaller input data for Ptr-Net. In applications, the proposed work can be applied practically to the case of large input size, for example, the employment for the ball-collecting robot in a golf court.