The Development of an Autonomous Navigation System for Weeding Robots in Paddy Fields Using Deep Reinforcement Learning

Abstract

This paper describes a method for deriving optimal operations to enhance the efficiency of a weeding robot designed for rice paddies. The robot's operation must cover the entire paddy with minimal overlap while avoiding collisions with obstacles. To enable the robot to perform such tasks, we developed a learning algorithm based on Deep Reinforcement Learning, specifically a Deep Q-Network. The agent's actions involve directing the robot in one of eight possible directions. The inputs to the neural network include the robot's position, orientation, and images representing the work progress. After one million training steps, we observed a trend of improving performance per episode as training progressed. In future research, we will aim to achieve task completion by increasing the number of training steps and incorporating improved methodologies.