Fine-tuned YOLOv8m that detects the two connector ports on the task board of the AI for Industry Challenge cable-insertion task (UR5e inserts a cable plug into a randomised task-board port).

This is the only learned component in our solution. Everything downstream of it — board pose, port lock, approach, and the force-controlled seat — is geometric.

Note the base yolov8m.pt has an 80-class COCO head; this checkpoint has a custom 2-class head and is not interchangeable with a stock checkpoint.

Training data is synthetic — rendered from the challenge's Gazebo simulation (RGB from the left/center/right scene cameras), not real photographs.

In our pipeline the box centre is back-projected onto the board's CAD z-plane, then gated to the target module before averaging — see Known pitfall below.

The detector fires on every port on the board, not just the target. The eval scene spawns three SC modules and multiple NIC cards, so taking a median over all detections of a class averages across distractor modules and lands the estimate tens of millimetres from the true port (we measured 58 mm on SC this way).

Select the target module first (we gate by board-frame Y), then take the robust median of the survivors. That single change took our SC port lock from 58 mm → 2 mm, with no change to the detector.

AGPL-3.0, inherited from Ultralytics YOLOv8.