An Embedded Deep Learning System for Grasping and Classifying PVC Pieces in Cluttered Environments

Robot grasping and manipulation in clutter environments remain challenging tasks due to the need for multiple machine intelligence capabilities. In this research, we present a Deep Learning-driven machine vision intelligence with a robotic control framework to grasp and classify PVC pieces in stand-alone mode using a Niryo One robotic arm, an RGB-D camera, and a Jetson Nano. The Deep Learning-based algorithms were integrated using ROS to automate the object grasping, classification, and relocation (i.e., organization) tasks. The validation of the proposed system produced a success rate of 94% in the grasp-hold objects task, an accuracy for object classification in real-time attempts of 90.5%, and an accuracy of the overall object organization task of 86%. Additionally, the complete system was deployed in a Jetson Nano without relying on external computing resources. The CPU, GPU, and RAM usage were recorded below 65%, proving the feasibility of performing object organization on a computation-constrained board. These results establish a solid foundation for complex robotic manipulation systems used in collaborative or industrial applications.