The Pallet Built Backwards from the Shelf - by Jaimin

Atoms to Algorithms

SubscribeSign in

The Pallet Built Backwards from the Shelf<br>Wednesday, June 10, 2026 · Logistics & Warehousing

Jaimin<br>Jun 10, 2026

Share

In the last post robot had to dig. Today’s never does, because it does not store totes at all. Symbotic stores cases, and its hard problem starts only once a case is in the gripper: how do you stack a few dozen mismatched cartons, crackers, canned soup, bottled water, into one pallet that is dense enough to fill a truck, sturdy enough to survive the highway, and arranged so that when it reaches a store, a night crew can take it apart aisle by aisle in a single walk? Symbotic says it builds those pallets at more than 1,300 cases an hour per cell, and runs them 30 to 50 percent denser than a person stacking by hand. The robot arm is the easy part. The plan is the product.

Three days, three companies, three completely different bets. Amazon changed the box so the arm barely had to think. Ocado kept everything in identical totes and paid for density by making its robots move, digging through tightly packed stacks. Symbotic takes the third road: it never digs and never reshapes the product, but it pays in foresight, building a pallet whose final shape is dictated by a store it will not see for two days.<br>How it actually works

Picture a stream of cases arriving at a robotic cell in whatever order the warehouse happened to retrieve them, which is almost never the order the pallet wants. Each case is different: its own size, weight, balance point, and a “crushability” rating, how much weight it can take before the cardboard caves. For every case, as it arrives, the robot has to decide where it goes on the growing stack and which way to turn it, and it has to decide right then, because the next case is already on its way. There is no pausing to see the whole batch and plan the perfect pile. You commit, then live with it.

That would be hard enough. Symbotic piles on three more demands. The stack has to stay standing, so the planner builds in flat layers and uses heavy, solid cases to brace lighter ones with a high center of gravity. It has to obey business rules, caustic cleaning products go below food, allergens stay apart, fragile items ride on top. And, the clever part, it has to come apart in the right order. The pallet is built to a specific store’s floor plan, so that unloading it walks the aisles in sequence. The stack is, in a real sense, built backwards from the shelf it will end up on.

Symbotic’s own description is “a high-speed game of 3-D Tetris,” run by software it calls the Pallet Build Planner. The comparison is fair but generous to Tetris. In the game you see the next piece and you have one rule: fill rows. Here the pieces arrive unknown, a dozen rules interact at once, and “winning” is not rows cleared but a blend of density, stability, and how neatly the finished pallet unpacks in a stockroom.<br>The cleverest piece is what happens when something goes wrong. Lines hiccup: an expected case fails to show, or arrives crushed. A planner that committed to a flawless pile is suddenly holding a blueprint with a hole in it. Symbotic’s software replans on the fly, slipping in a different case or spreading the rest to close the gap, so a robot can still finish the pallet instead of handing it to a human. That ability to bend without breaking, not the tidy best case, is what separates a production system from a demo.<br>New this week

Researchers published a method in January 2026 that packs against the near future instead of just the present, peeking a few cases ahead and searching possible stacks the way a chess engine searches moves, which lifted performance by more than 10 percent exactly when the incoming mix of goods shifts, the same hiccup Symbotic’s replanner is built to absorb (arXiv 2601.02649). A March 2026 system called STEP optimized for speed rather than tightness alone and reported cutting operational time by 44 percent without losing density, a direct nod to the metric that actually decides whether one of these cells pays for itself, cases per hour (arXiv 2603.07800). And a benchmark robot from last year grounds the hype: it packs real boxes on a bare pallet at about 10 seconds each, roughly 19 boxes a pallet, a useful reality check against the production numbers companies quote (arXiv 2504.04421).<br>What to notice

The visualization puts all three of this week’s companies on one axis: where each spends its hard work. Amazon spends it on the scene, Ocado on motion, Symbotic on the plan. Next to it sits a single mixed pallet in cross-section, heavy braced cases at the base, fragile ones on top, an aisle-order label running down the side. The point worth sitting with is that the same business, getting groceries to people, splits into three engineering bets that barely overlap, and only one of them, Symbotic’s, builds an object whose shape is decided by a building two days down the...