I’m building a ROS 2 pipeline that must reliably spot nuts and bolts in a live camera feed and publish their positions to the rest of my stack in real time. Your job is to create the complete vision-detection module—from model training or fine-tuning through to a clean ROS 2 node that subscribes to an image topic and spits out the detected objects with bounding boxes (or masks) and a confidence score. OpenCV, TensorFlow/PyTorch and any of the common ROS 2 image-transport plugins are all fine as long as the final node runs on Humble and stays GPU-agnostic (CUDA acceleration is a bonus, not a requirement). I already have a test rig with a standard USB camera; if you need specific calibration images I can capture them for you. Please deliver: • Source code for the detection model and ROS 2 node • A launch file that brings everything up with default parameters • A brief README explaining setup, parameters and expected topics • A short video or bag file proving the detector works on my sample parts Acceptance criteria: at least 90 % precision/recall on my validation set and a minimum 10 fps throughput at 640×480. Let me know how many training images you’ll need, your preferred framework, and an estimated timeline when you reply.