ACI SmartCheck: Computer Vision for the Picking Station

The picking station is where an order is assembled, and it is the one step nobody records. ERP sees the order and WMS sees the stock. What happens on the table between order accepted and box shipped stays invisible, so mistakes surface only after a shipment leaves or a return comes back.

That gap is expensive. In some operations as much as a fifth of orders come back, and in a dispute there is nothing to check against. CCTV exists, but finding one pack in the footage takes hours. The operator works without feedback, so a wrong item goes in the box with nothing to catch it.

ACI wanted to close that blind spot. The fix is a station that watches its own work: an overhead camera over the surface, a small NVIDIA Jetson computer beside it, and a screen that shows the operator what the system sees. The goods vary, from marked pharmacy and e-commerce items to small unmarked parts, so the table has to read codes and recognize bare objects.

ACI SmartCheck turns an ordinary pick into structured data, evidence and live hints. It does three things on the same camera and screen.

It counts the items on the table, by type, with no marking needed. It reads their markings, the QR, Data Matrix and barcodes, and checks each one against the open order: green means it belongs, red means it is extra, yellow means not read yet. And it records every pack, the camera footage plus a text log tied to the order, so a pick can be pulled up later by a link.

The operator gets the answer while the pick is still open: what is complete, what is missing, what is extra. The screen runs the whole pick, from loading the order to a finish-picking confirmation when the box is complete. Everything runs on the device, with no cloud and no internet needed. The catalog of item types, the icons and the models are tuned to the customer's own goods.

At LogiMAT 2026 in Stuttgart SmartCheck read 13 of 13 mixed items in one frame across 4 marking formats at once, including QR, Data Matrix, barcode and PDF-417, so nothing had to be re-scanned. It also counted mixed piles of unmarked parts and held up after changes in light and color, with close types merged in the catalog.

On a Jetson Orin Nano it processes a frame in about 472 ms, fast enough to guide the operator while they work. For the business that means fewer picking errors, because the operator sees a problem at the table while the pick is still open, and a searchable record that closes a dispute in minutes where CCTV took hours.

This was a proof of concept built for the show, so read these as directional figures from the demo bench, with a controlled study still to come in a pilot. The point it proved: the picking station can record itself and check the order in real time, on cheap hardware, with no cloud.