Development of a Tomato Harvesting Robot: Peduncle Recognition and Approaching

Abstract

Over the last decade there is a growing effort to incorporate robotic technologies in agriculture to reduce costs and increase yields and quality. In this paper we present the methods and tools developed for automated tomato harvesting by a greenhouse robot. The system is comprised of a 6-dof manipulator arm, a custom gripping/cutting end-effector, and a depth camera with dedicated vision processor. Deep learning algorithms are employed to locate ripe tomatoes and their peduncles, exploiting depth information from the acquired images to guide the manipulator arm towards the identified cutting points. Through extensive experiments in a realistic setting, the overall success rate of the detection and approach procedure was found to be 65%, with 92.6% accuracy of the vision processing in locating the correct cutting point. Suggestions for further improvement of the system's performance are also provided.