Mixed-case depalletizing wasn't automatable until 3D vision could find a box it had never seen

October 1, 2026 · 3 min read

Unloading a pallet of identical cases was solved long ago. Unloading a pallet of mixed cases, different sizes, random order, some shrink-wrapped and some not, was a job only a person could do. Three dimensional vision with learned box detection changed what a robot can pick without being told what is coming.

The short answer

A robot can only pick what it can locate. For decades that meant depalletizing worked when every case was the same size in a known pattern, and failed the moment a distributor's inbound pallet arrived with twelve SKUs stacked by a person at the shipper. Three dimensional cameras combined with detection software trained on boxes in general, rather than programmed for a specific box, let the robot find the top of the next case, estimate its size and pick it, without a pattern file. That is what moved mixed-case depalletizing from research demos to warehouses.

Why this job resisted automation for so long

Classic depalletizing is a geometry problem solved in advance. Someone enters the case dimensions and the layer pattern, and the robot goes to the computed positions. A mixed pallet has no pattern. Cases shift in transit. Some are taped, some wrapped, some crushed at a corner. Two dimensional cameras struggled to tell where one brown box ended and the next began, and false picks, grabbing two cases or none, made the cell slower than the person it replaced.

Grippers were the second problem. A vacuum tool sized for one case either lost suction on a smaller one or could not lift a heavier one, so the cell needed either a tool changer or a compromise gripper that handled nothing well.

What changed

Structured-light and time-of-flight 3D cameras became affordable and fast enough to image a pallet layer in a fraction of a second. Detection models trained on large numbers of boxes learned to segment cases of unknown size, including the hard ones: same color, tight seams, torn flaps. Zoned vacuum grippers that activate only the cups over the case learned to pick a small box and a large one with the same tool. The result is a cell that approaches each pallet as it arrives and decides what to pick next.

What it looks like now

An industrial or heavy collaborative arm with reach for a full pallet, a 3D camera over the pick zone, a zoned vacuum gripper, an outfeed conveyor, and guarding or scanners per the risk assessment. The cycle is detect, pick, place, repeat, with the software choosing a safe pick order so stacks do not topple. Rates are lower than single-SKU depalletizing and vary with how messy the pallets are, which is why the honest sizing exercise uses your worst inbound pallets, not your best.

What still disqualifies an operation

Pallets with bags, pails or odd shapes mixed in with cases, which defeat box detection. Cases too heavy or too large for the arm and gripper class. Receiving volumes too low for the cell to earn its keep. Glossy or black wrapping that confuses some cameras, which is a test, not a guess. And a dock where the pallet quality problem is really a supplier compliance problem that a conversation would fix cheaper than a robot.

How to find out if yours is automatable

Photograph twenty inbound pallets at random and sort them into clean, mixed and ugly. The share in the third pile sets the project's difficulty, and any vendor worth talking to will want to run their vision system on exactly those pallets before quoting. Describe the dock to Joe, the concierge on this site, with the case weights and the number of SKUs. It will give you a preliminary read on the machine class and tell you which questions to put to the integrator.

Related on this site

If every pallet you receive is different, the robot's value lives in the vision system. Qualify that, with your worst pallets, before you qualify the arm.

Put it to the test on your own job

Tell Joe what you are running and he will apply this to your numbers, including telling you when a robot is the wrong answer.

More notes