Giving our drone sight – how we taught it to find probes

Modern drones are often flown by following predefined waypoints, i.e. fixed positons in space that are visited sequentally. This provides a short context for what “waypoints” mean in the following.We wanted the drone to do more than follow waypoints. It needed to see – to identfy ground probes, estmate where they are in meters, and hand clear, usable positons back to our mission logic. The result is a compact vision pipeline that blends modern object detecton with classical geometry and a likle statstcal smoothing – simple in concept, powerful in practce.Figure 1 Learning image with probes in RoboFlowWe started with data. Images were annotated in Roboflow to get consistent, high-quality bounding boxes for the probe types we care about. That foundaton made training straigh forward: after an inital run we finetuned the best model at higher resoluton to capture subtle visual cues on the ground. Training on a local machine…