Deep learning based autonomous harvest robot for kumqat fruit

The history of harvest robots has been shaped in parallel with technological developments in the agricultural industry. The first harvesting robots started with simple mechanical devices, usually mounted on tractors, and have evolved over time with sensor technologies, artificial intelligence and automation systems. These developments have allowed harvesting robots to become more precise, faster and more efficient. The use of harvesting robots helps manage resources more effectively by increasing sustainability in the agricultural industry and is considered an important step to meet future agricultural needs. These robots can have the ability to automatically collect, classify and package fruits, vegetables and other agricultural products. These robots also provide solutions to the labor shortages encountered in the agricultural sector, providing agricultural enterprises with a more reliable and effective harvesting process. In this thesis, an autonomous harvesting robot was developed to automatically harvest kumquat fruits. The basic structure of the robot consists of a 5-axis robot arm and a 4-wheel mobile platform. RGB-D depth camera was used to detect kumquat fruits and their depth. The YOLOv7 algorithm, which gives good results in terms of speed and detection performance from deep learning models, was used. Here, forward kinematics and MLP-based inverse kinematics calculations were made for the robotic arm, and the average error value was measured to be less than 1% in the total of the entire joint. In tests conducted on the kumquat fruit tree in its natural state, object detection performance was measured as 93% and harvest performance was measured as 75%. Here, an autonomous harvesting robot has emerged that can be a solution to the problems arising in fruit harvesting and can be used in the agricultural industry.