YOLO on GPU Implementation for a Vision-Based Fruit Select and Place System with a 5-DOF Robotic Arm

This article introduces a cost-effective approach for a vision-guided robotic pick-and-place arm and demonstrates how this produces a faster inference rate and greater efficiency than existing products, which are typically constructed with rather high-end equipment. A YOLOv4-tiny algorithm-based object detection model is trained to classify and localise the desired object using a custom-created dataset. A monocular camera was also calibrated to determine its intrinsic and extrinsic parameters, which facilitated the determination of coordinates. GPU-based NVIDIA Jetson NANO is utilised to process the live stream captured by a camera, not only due to its faster inference capability and computations but also because of its quicker inference capability. The fabrication of 5-degree-of-freedom robotic limb results in improved positioning, more precise target object placement, and a larger work envelope. This arrangement identifies and localises the target object within the image frame by recognising it in its surroundings. Later, the image coordinates are mapped to real-world coordinates and fed to the robotic arm, which applies the appropriate inverse kinematics to plan the trajectory and implement the pick-and-place task. This low-cost system resulted in 40 frames per second of real-time inference speed, which is nearly double that of existing models.