Pick arms and the unfamiliar item

A pick arm does not fail on the items it was trained on. It fails on the box that arrived yesterday in different packaging, the item wrapped in a reflective film the training set never included, the bag that slumps instead of standing. Those are the picks that end in an exception, a flashing light, and a person walking over to clear the cell.

The catalog on a fulfillment floor changes faster than any vendor's model does. New items, new suppliers, new packaging for old items. The vendor's answer is usually a model update on the vendor's schedule, or a cloud call that sends the image of the item out of the building and waits. A facility inference pool offers a third option: a larger model in the rack the arm can ask right now, and a way for the site to teach that model the new item itself.

The question the arm cannot answer alone

An arm's on-board model is sized for the arm. It is fast and good within its training, and it is the right place for the reflexes: closing the gripper, detecting a slip, respecting the force limit. It is the wrong place for a large perception model that can look at an unfamiliar object and say what it is, where to hold it, and whether it is fragile.

With a Czar Grid in the building, the arm sends that question over the facility's network to the pool. Substrate serves the perception model and holds a latency budget for the arm's question, which is tighter than a budget for a planning task and looser than one for a reflex. Inside that budget the arm gets an answer from a model larger than it could ever carry.

When the answer is late

A pick cell is a place where guessing is expensive. A wrong grasp drops the item, jams the conveyor, or sends the wrong product down the line. So the fallback is deliberate. If the pool does not answer in time, the arm's own controller decides, and the decision it is built to make is to stop and raise the exception, exactly as it would have without a pool.

An arm without a pool answer is an arm that halts, not one that improvises. Nothing about the arm's safety envelope, its force limits, or its stops moved to the rack. The pool added a better answer when it is available; it did not become something the arm depends on to be safe.

Learning the item once

Here is where a pool differs most from a cloud call. When the arm meets an unfamiliar item and a person resolves the exception, that moment is a capture: the image, the item, the grasp that worked. On a floor with a pool, those captures stay on the operator's hardware, and Substrate's adaptation can fine-tune the shared perception model on them, on the pool, with nothing leaving the building.

The next arm to meet that item asks the same pool and gets an answer shaped by what the floor already learned. That holds across vendors: if the site runs pick arms from more than one vendor, the item is learned once, by the pool, rather than once per vendor on each vendor's own schedule. The companion article on a new item learned once, not once per vendor goes further on that point.

What the record shows the next morning

Every exception has a trail. The pool records what it was asked, what it answered, whether the arm acted on the answer or fell back, and which captures went into adaptation. When an engineer wants to know why a cell halted on a particular item, the answer is in a record on site that a person can read, not in a support ticket to a vendor.

That record also helps the site decide what to do about the item. Some items are worth teaching the model. Some are better handled by a person or a different cell. The site can make that choice with evidence rather than with a hunch, and the choice stays local.

A floor, not a product catalog

A pick arm vendor loses nothing when a site adds a pool. The vendor keeps its arm, its controller, and its software. What the vendor's customer gains is an arm that can ask a bigger model about the box that arrived yesterday, and a floor that teaches itself the next one. For how pick arms sit alongside mobile robots and humanoids on the same pool, see the logistics solution.