A Robotic Complex Control Method Based on Deep Reinforcement Learning of Recurrent Neural Networks for Automatic Harvesting of Greenhouse Crops

Abstract

The modern development of technology determines the feasibility of the transition in agriculture from manual labor to automatic production. One of the promising areas is the automation of growing vegetable crops in greenhouse complexes. Necessary factors for intensive plant growth and unfavorable for human health, such as high temperature and humidity, as well as an atmosphere saturated with chemicals, make the task of robotizing agricultural operations urgent in this area. The method for controlling a robotic complex for automatic fruit collection in greenhouse complexes is proposed. Work in greenhouse complexes is characterized as non-deterministic and with partial observability of the environment; therefore, the deep recurrent neural network DRQN was used as the basis for the method of controlling the robotic complex. Deep learning with reinforcement was used for optimizing its weights. The presented simulation results demonstrate the efficiency of the proposed method and the need for its further development.