Integration and preliminary evaluation of a robotic cotton harvester prototype

Abstract

Cotton is conventionally harvested with large and costly harvesters once at the end of the growing season. Fiber in the early-opened bolls is left on the plants till the end of the season to be harvested, resulting in its exposure to weather and thus degraded quality. Furthermore, heavy cotton harvesters can compact the soil and adversely affect soil health and crop yield. These issues can potentially be addressed by using small robotic cotton harvesters that can harvest multiple times during the season, picking seed cotton soon after the bolls open. Robotic cotton harvesters are not currently available. This article presents the system integration and performance evaluation of a new robotic cotton harvester prototype consisting of a 3-DOF (degrees of freedom) linear robotic arm with a custom three-finger end-effector, integrated with a deep-learning based perception module to enable fast and automatic harvesting of cotton. The perception system utilized a stereovision camera and an onboard computer with a trained YOLOv4-tiny algorithm for detecting and locating cotton bolls within the plant. Three different control algorithms were examined to control manipulation of the arm and run the end-effector throughout the harvesting process. Performance of the harvester was evaluated in terms of picking ratio – picked seed cotton over available seed cotton in the image – and cycle time through laboratory tests using real cotton plants. Results showed that a closed-loop control algorithm using continuous feedback from cotton boll images during picking was the most efficient, successfully harvesting 72% of the seed cotton with an average cycle time of 8.8 s.