Topping and Tailing Onions at One Per Second: Five Washdown Robots, Ten Cameras, and a Jetson

Justin Eckhardt
18 hours ago
5 min read

Onions are a miserable product to automate. No two are the same size, the same shape, or even close to the same orientation when they arrive. The skin flakes, the roots trail dirt, and everything in the room gets hosed down at the end of the shift. This is the story of a machine we built for an onion processing facility that tops and tails onions robotically - the kind of job that has historically been done by a row of people with knives.

The problem shape

The facility needed the top and the tail cut off each onion, cleanly, with minimal waste, at a target rate of one onion per second regardless of size. Downstream of the cut, the outer skin gets stripped with compressed air. The environment is full washdown - the machine gets sprayed with water and cleaning chemicals, so everything exposed had to be stainless or IP69K rated.

The hard part is not the cutting. A blade is a blade. The hard part is that every onion presents differently, and the cut planes are defined by features - the root tuft and the stem - that move around depending on how the onion sits. That makes it a vision problem first and a robotics problem second.

Line architecture

The system runs five robot cells in parallel, fed by a vibratory shaker table that singulates and presents onions at a pickup zone. Each cell is a self-contained loop:

A washdown-rated Fanuc LR Mate robot picks the onion off the shaker table with a vacuum suction tube
Two Luxonis OAK-1-PoE cameras image the onion - one over the pick zone to locate it, one at a presentation station to find the root and stem after pickup
The robot reorients the onion based on what the cameras saw, then drives it through a stationary V-knife cutter - top cut, spin, tail cut
The cut onion drops to an air peeler, where high-pressure air strips the outer layers without touching the flesh

All ten cameras feed a single NVIDIA Jetson AGX Orin, which runs the detection models and talks to all five robot controllers over Ethernet sockets. A small PLC exists in the panel, but honestly it functions as a glorified timer relay - it controls how long the air peeler stays on. The real logic lives in the robots and the Jetson.

Component anatomy

The robots are Fanuc LR Mate units in the IP69K washdown package, each with its own R-30iB Mate Plus controller. We optioned them with DCS position and speed checking (software-defined safety zones), the Ethernet package for socket comms, and Fanuc's ZDT monitoring subscription. Area safety is handled by Keyence SZ series safety laser scanners watching the operator side, wired through Phoenix Contact safety relays.

Vision is deliberately cheap and rugged. The Luxonis OAK-1-PoE is an IP67 PoE camera with onboard compute, and at roughly a tenth the price of a traditional smart camera it made putting ten of them on one machine economically sane. They run fixed focus and fixed exposure - autofocus is the enemy of repeatability - under LED2WORK TUBELED IP69 washdown light bars. A 16-port PoE injector and a stack of TP-Link switches carry everything back to the Jetson on Cat8.

Detection is a two-stage pipeline on the Jetson: an Ultralytics YOLO model finds and boxes the onion, then a classic OpenCV blob detector finds the darkest blob - the root tuft - inside a cropped region. Deep learning for the "where is the onion" question, deterministic image processing for the "where exactly do I cut" question. Pixel coordinates convert to millimeters through per-camera calibrated scale factors and offsets, and land in robot user frames.

Vacuum handling is all Schmalz: dry-running vacuum pumps with constant suction, food-grade NBR bellows suction cups, in-line vacuum filters with transparent housings so you can see them loading up, and pilot-operated vacuum solenoid valves. An SMC vacuum sensor per cell confirms the pick - if there is no vacuum, there is no onion, and the robot skips the cycle instead of air-cutting.

The cutter is our own design - a V-knife arrangement with replaceable stainless utility blades in machined holders. The blade holders went through several revisions as we tuned the entry geometry, and the consumable is a standard blade you can buy anywhere rather than a custom grind.

The electrical enclosure is a 72x72x18 stainless sloped-top Saginaw with a NEMA 4X disconnect, Eaton M22 operators, an IDEC e-stop, and Icotek and Roxtec glanding for the cable penetrations - the sloped top matters in washdown, water has to shed, not pool. Five robot controllers, five vacuum pumps, the Jetson, the PoE gear, and all the terminal rail live in that one box, with a vortex cooler and later a proper air conditioner handling the heat.

The robot code

Each Fanuc controller runs a main select-loop program plus a set of background tasks - background logic programs that run continuously from boot. One background task manages the vacuum pump and pilot air as a unit. Another implements the blowoff pulse: because Fanuc background tasks cannot use wait or pulse instructions, the blowoff timing is done with a register counter acting as a homemade pulse timer. Another watches the vacuum sensor and sets a skip flag the motion programs check. A clear routine resets registers, I/O, and frame and tool data when the operator hits reset on the HMI.

The Jetson side is Python - one process per robot, each holding a socket to its controller and exchanging data through numbered registers: a PC-to-robot register carries commands and coordinates, a robot-to-PC register carries state. Simple, debuggable, and language-agnostic on both ends.

Commissioning lessons

We made repeated trips to the facility over multiple seasons, and the machine changed every time. The lessons that stuck:

Organic product variability is the whole project. The difference between an 80 percent demo and a production machine is entirely in the tails of the size and shape distribution. Adjustable cut parameters per size class beat any single fixed geometry.
Fixed exposure and fixed focus, always. Every automatic camera feature trades repeatability for convenience, and repeatability is the product.
Hybrid vision beats pure deep learning here. YOLO is superb at finding onions, but a threshold-and-blob pass on a cropped region localized the root more precisely than the box ever could - and it runs in a couple of milliseconds.
Vacuum on a dirty, papery product means filtration you can see. Transparent filter housings turned a mystery fault - slow vacuum decay over a shift - into a visual check.
Design the consumable around a commodity part. Custom blades would have made every blade change a purchasing event.
Put the sanitation people in the design review. Sloped tops, standoff spacers, metal-detectable rubber bands, and drainage paths are cheap on paper and expensive to retrofit.

What carried forward

This machine set patterns we still use: the Jetson-plus-PoE-camera architecture for multi-camera cells, the two-stage detect-then-measure vision pipeline, register-based socket handshakes between Python and Fanuc controllers, and the discipline of treating washdown as a first-class design constraint instead of an afterthought. Later vision systems we built for other food and medical device customers are direct descendants of what this row of five robots taught us.

If you are looking at a product everyone says is too variable to automate - that was this onion, once.

Downloads

Free, no signup. The knowledge file is designed to be dropped into an AI assistant's context. No customer information in any of it.