An AI peach sorter has to get a grip

Agrophoenix’s peach-sorting installation had a slippery problem. QING’s cameras could identify fruit to reject, but peeled halves also had to stay in place and respond reliably to a vacuum gripper. Stäubli’s 8 October account describes eight parallel cells processing 15 tonnes an hour at the Greek plant.[1]

Robovision, which supplied the vision platform, describes a development process split between QING’s laboratory and the customer’s site. Engineers integrated cameras, lighting, conveyors, robot arms and control software; the application learned to identify abnormalities including retained kernels and crushed fruit. Installation included rebuilding and testing the system at the processor.[2]

A defect label needs a location

Robovision’s grading-and-picking documentation explains the information such a system can produce. Labelled examples teach a model to identify an item’s outline and class. Dimensions and orientation help determine a picking pose. Results reach the machinery in metric units aligned with the conveyor’s coordinate system, so an instruction can identify both what to remove and where to reach.[3]

That combination matters for a naturally variable product. A quality judgement attaches to one particular half, with its own shape and orientation. The downstream controller needs to preserve that connection as the item travels. Recognising a defect and successfully removing the corresponding object are two separate events that the line must bring together.

The belt is part of the measurement

QING commercial director Will Uijting told Stäubli that the team had to find a conveyor material that held the wet fruit in position. Picking and releasing halves with their hollow side facing down also required the right vacuum gripper and contact material.[1]

The general tracking mechanics are documented separately in Stäubli’s VALtrack library. Its vision configuration triggers an image while recording an encoder reading—a measurement of conveyor movement—then sends the detected object information to the robot controller. Calibration relates the belt’s geometry and movement to the robot’s coordinates. This is background on conveyor tracking; the peach case does not identify its tracking library.[4]

An encoder follows the belt. If a peach slides relative to that belt, its inferred destination can drift even when the camera classified it correctly. Holding the fruit steady therefore preserves information as well as making handling easier. Improving the contact surface can fix a failure that retraining the classifier would leave untouched.

A rejected half can come back

In the installation Stäubli describes, halves containing larger pit fragments go to automated re-pitting; those with smaller fragments go to manual inspection. Both streams return to the sorting stations. The robot is routing fruit for another operation, with a possible return to production.[1]

That return route gives the quality decision a practical consequence beyond a reject count. A useful check follows the item through the whole sequence: was the right half selected, did the gripper transfer it, and did the intended rework happen? A successful image classification supplies the first answer. The belt, gripper and downstream process supply the rest.

Report an error or suggest a correction