On mixed-SKU casting grinding, changeover time is rarely the robot program

Many foundries size a grinding cell on one representative part. Three months after start-up, output is eaten by two-hour changeovers — and by the half-shift of hand touch-up that follows.

A single-part cycle chart often dies on a mixed line

Quotes for automated grinding are almost always built on one part: flash length, gate stub size, minutes per cycle, whether dual stations overlap load and grind. The spreadsheet looks clean.

The shop floor runs a different clock. Valve bodies in the morning, brackets after lunch, two pallets of housings squeezed in the next day. The representative cycle time is still true. What is not true is dividing 480 minutes by that number. Subtract fixture swap, tool swap, probing, first-article grind, and quality release. That remainder is the real shift.

People call this “slow program change.” Calling a program on the HMI takes seconds. Locating the part, matching the tool, keeping the compensation datum, and getting first-article sign-off are slower. Three of those four have little to do with the robot brand.

Changeover is four layers, not one motion

Shops treat changeover as stop, swap, run. It is at least four layers.

Program: path, spindle speed, feed, force or compensation settings. This layer is fast if programs live under a part number, not in someone’s teach-pendant backup.

Fixture: clamps, locators, jaws, or a complete tooling plate. Casting sand wears pins. A hole that matched the drawing in month one does not match in month six.

Tool: wheels, burrs, mills, flap wheels are not universal. A hard wheel sized for iron gate stubs will tear an aluminum skin. An aluminum face mill on ductile-iron flash goes dull in minutes. A tool change usually needs balance or a touch-off. Skip that and the first pass vibrates the new path off.

Datum: probe points, vision templates, safe retracts. Points that sat on a valve body last week may land on flash or empty air on this week’s bracket. A wrong probe point is worse than no compensation.

If any layer is unprepared, a “15-minute changeover” on the form becomes an hour. Many acceptances run eight hours on one SKU. Changeover is never in the scorecard.

Fixtures are where most plants actually lose the hour

Stable families can live with a self-clamping fixture or a press-down unit and only swap locators. Plants with scattered geometries are different. A valve uses a register, a bracket uses two pads, a housing sits on the cavity. A “universal fixture” usually locates universally badly.

Simple pins are cheap and quick to change. They are also sloppy once parting-line shift shows up. Hole-break edges then vary lot to lot, machining sends parts back, and the simple fixture grows extra stops until it is no longer simple.

Self-clamping tooling locates well and behaves like a small machine of its own when you swap it. Lead time is long, storage is real, and door-swing interference on a three-station table often appears only at install.

AI grinding’s honest benefit on changeover is fewer hard locators. A tray or a coarse stop can be enough for the vision system to find pose. That does not turn a thin aluminum housing into a rigid body. Vision finds the part. Cutting force and chatter still need support.

When tools lock you to a material family, changeover becomes a cell change

Iron cells are built around wheel load and stub cutting. Aluminum cells more often run face mills, cutters, or belts. Steel adds hardness. Hanging an iron wheel and an aluminum mill on one spindle is possible on paper. Dust mix, filter fire risk, hood opening, and airflow are designed around one chip type. Few plants keep that mix for long.

Mixed SKUs therefore does not mean one cabin for every alloy. The cheaper split is usually iron on one cell, aluminum on another, then size bands inside each family. Changeover stays inside a family where fixtures and tools still share something.

Forcing a ductile valve and an aluminum housing through the same cabin puts spindle maps, dust mode, guarding, and base plates on the changeover list. The machine you saved is paid back in lost hours and process risk.

A stored program and a re-taught program are two different factories

Someone who can teach a robot is not the same as someone who owns a program library. Names like “valve2”, “valve2-fix”, “final” mean nobody will call the file three months later. Changeover becomes: open the nearest path and re-touch the critical points.

Compensation belongs with the part number: probe coordinates, allowable offset, force cap. Store only the path and the probes stay where the last product left them. First article either cuts into parent metal or leaves a ring of flash.

A Chinese teach pendant shortens training. It does not replace version control, first-article records, or a name on the last path edit. When the senior operator is off, that gap hurts more than a spindle alarm.

Vision shortens finding the part. It does not skip process prove-out

High mix, modest volume is the usual reason to look at AI grinding. You weld fewer hard fixtures. For housings and brackets that do not justify a full clamp set, that reason holds.

What does not hold is the hope that vision makes changeover a start button. The template says which product this is. It does not say how tall this gate stub grew after a tooling wear week. Tool strategy and compensation still sit per family. A new part still needs a first article and a feed tweak.

Treat AI as “no process work” and you will rework. It cuts fixture count and search time. The process window stays. On tight windows, first-article time after a change can exceed the fixture swap itself.

Not every high-mix shop should chase one cell for the whole list

Mixing on one cell usually needs several conditions at once: same material family, size inside one table, similar burr shape so tools can be shared, enough daily volume to keep changeover count tolerable, and a shop willing to keep fixture and program books.

Bad mixes are equally concrete: iron with aluminum, stub cutting with cosmetic faces, prototypes stealing a production cell, downstream machining that already uses the ground edge as a datum. A second cell looks expensive until you add changeover, touch-up, and scrap.

A middle path is common: one cell locked on two similar runners that make most of the tons, a third geometry left to a bench or a vendor. Automation does not have to cover every drawing number. Covering the two numbers that are 80% of output is often the better buy.

If the changeover sheet is vague, cell speed does not matter

On the machine side: dual stations so the door is not the cycle, HMI recall by part number, fixture quick-change interfaces, at least a quick-change tool flange, compensation stored with the program. Dedicated foundry cabins tend to hold these together because table, door, and dust are designed as one cycle. A general-purpose robot in a fence usually does not.

The plant owns the other half. Fixtures wiped before the swap, pins replaced on a calendar, wheel remaining diameter written down, first-article dimensions listed, a rule for whether a bad part means a path edit or a locator check. Without that sheet, a twenty-minute vendor changeover still dies in the afternoon.

Scheduling has to change with it. Three unrelated geometries in one shift will beat any spindle. Run a family together. Park changeovers at shift start. “Cut over any time” is a sentence from a presentation, not from a grinding bay.

Ask how many times a day you change before you ask if the robot can grind

Whether a grinding robot can physically cut a casting is the first sourcing question. On a mixed line it is not the deciding one. The deciding questions are whether those drawing numbers belong in one cabin, and which of the four changeover layers you have actually prepared.

A program that appears on the screen only proves the integrator shipped a file. Repeatable location in twenty minutes, tools that stay inside one material family, probe points that travel with the SKU, and a signed first article — that is changeover capability. Miss one, and you are back to the single-part chart that never described the shift.