Coverage path planning for kiwifruit picking robots based on deep reinforcement learning

Abstract

In this paper, a deep reinforcement learning-based path planning method for kiwifruit picking robot coverage is proposed. Compared with existing approaches, the novelty of this paper is twofold. 1. Using a LiDAR to collect the environmental point cloud information of the kiwifruit orchard and construct a two-dimensional grid map. In the process of constructing the map, the fruit coordinate information is collected in real time, and the fruit coordinates are projected onto the grid map to obtain the distribution of kiwifruit in the orchard environment. Combined with the effective picking area of a kiwifruit picking robot, a kiwifruit area division algorithm is proposed, which converts the traditional grid-based coverage path planning into a travelling salesman (TSP) problem of solving the traversal order of each area. 2. An improved deep reinforcement learning algorithm, the re-DQN algorithm, is proposed to solve the traversal order of each region. The model training results show that the algorithm is more effective than the traditional DQN algorithm, completing model convergence to a better solution. The experimental results of kiwifruit orchard navigation show that the coverage path length of the method proposed in this paper is 220.67 m, which is 31.56 % shorter than that of the boustrophedon algorithm. The overall navigation time is 1200 s, which is 35.72 % shorter than that of the boustrophedon algorithm. This shows that the coverage path planning method proposed in this paper can effectively shorten the coverage path of kiwifruit orchards and improve the navigation efficiency of kiwifruit picking robots.