Supervised Reinforcement Learning-Based Collaborative Master–Slave Harvest Control Study in Wheat

Abstract

Aiming at the difficulty of controlling the longitudinal relative position of agricultural machines during the agricultural master–slave navigation cooperative operation and the weak adaptability of the unitary traditional control method in the face of the working conditions of complex farmland environments, this paper proposes a supervised reinforcement learning (SRL)-based longitudinal stable and safe control method applicable to master–slave navigation harvesting and unloading operations. Firstly, to improve the algorithm training success rate, a supervisor trained on actual driving data is introduced into the actor–critic reinforcement learning method. Secondly, in order to improve the effect of agricultural machine operation, considering the actual grain unloading operation scene and combining the smoothness of operation and the safety of unloading, a new reward function in the supervised reinforcement learning algorithm is designed. Finally, the performance of the proposed SRL control strategy is verified by simulation and agricultural machines following grain unloading tests. The results of field operation show that, when the harvester speed is 1.2 m/s, the average absolute deviation of the actual distance between the two trucks is 0.048 m, and the maximum deviation of the steady state is 0.26 m. In the variable speed test, when the harvester speed is 0.4 m/s and 1.2 m/s, the average absolute deviation of the actual distance between the two trucks is 0.079 m and 0.091 m, and the maximum deviation of the steady state is 0.20 m and 0.21 m, and the cooperative accuracy can fulfill the operational demands of harvesting cooperative unloading. The study’s results can serve as a technological reference for autonomous harvesting operations in the field.

Keywords

• following control
• automatic guidance
• longitudinal stability control
• safe unloading of grain