Automatic Path Planning for Spraying Drones Based on Deep Q-Learning

Abstract

The reduction of the agricultural workforce due to the rapid development of technology has resulted in labor shortages. Agricultural mechanization, such as drone use for pesticide spraying, can solve this problem. However, the terrain, culture, and operational limitations in mountainous orchards in Taiwan make pesticide spraying challenging. By combining reinforcement learning with deep neural networks, we propose to train drones to avoid obstacles and find optimal paths for pesticide spraying that reduce operational difficulties, pesticide costs, and battery consumption. We experimented with different reward mechanisms, neural network depths, flight direction granularities, and environments to devise a plan suitable for sloping orchards. Reinforcement learning is more effective than traditional algorithms for solving path planning in complex environments.