An inference pool for a fulfillment center

A fulfillment center is the facility most likely to run robots from more than one vendor at the same time. Mobile robots move totes and shelves, pick arms work the stations, and on some floors humanoids are arriving for the jobs that were left to people because nothing else fit the space. Each of those machines was designed to think for itself, and each is now running into the same ceiling: the models it needs for hard perception and for language are too large to carry.

An inference pool is the building's answer to that ceiling. Element 31 installs it as Czar, a sealed node in the facility's own rack, and joins several nodes into a Czar Grid when the floor is large enough to need one pool across the whole building. The robots stay on their own on-board models for reflexes and call the pool for the rest.

What the pool does on a fulfillment floor

The perception problems in fulfillment are specific and stubborn. A pick arm has to identify an item it has never seen, in a tote with other items, under a torn polybag, with the label facing down. A mobile robot has to tell a dropped tote from a pallet jack from a person kneeling in the aisle. Those are large-model problems, and the pool is where they run.

The decision problems sit on top. Which item is the one on the order, which grasp is worth attempting, whether to re-route around a blocked aisle or wait for it to clear. The pool serves those decisions to any robot on the floor, from any vendor, inside each robot's latency budget.

Language is the newest load. Associates and supervisors are beginning to speak to the fleet, and the fleet is beginning to answer with reference to the site's own documents: the pick-path rules, the hazardous-goods list, the dock schedule. Retrieval over those documents runs on the pool and stays on the pool.

Mixed fleets and one pool

The fulfillment center is where mixed fleets stop being a slide and become a shift. Several vendors, several fleet managers, several on-board model families. The pool does not replace any of them. Each vendor's fleet software keeps managing its own robots. The Grid serves inference to all of them over the same network, and each robot keeps its own fallback.

That arrangement matters most at peak. When the floor is at full capacity and a model is suddenly needed by every robot on it at once, the pool is in the building, on the operator's hardware, with no path through a distant service. A robot that cannot get an answer in time falls back to its own model and keeps doing what it already knew how to do.

Safety stays where it was

Nothing about the pool changes the safety case of any robot on the floor. Stops, speed limits, and motion limits remain on the machine, exactly as the robot vendor set them. The Grid serves perception and decisions. If the pool fails, every robot behaves as it would have without a pool: it stops, or it continues on its own model. It never guesses on the Grid's behalf.

For the safety officer, that is the first question answered before the second one is asked. The pool adds capability without taking on authority it was not designed to hold.

Posture for a fulfillment center

Most fulfillment sites are connected sites rather than air-gapped ones. The warehouse system has to know what was picked. The robot vendors' fleet software has to see its machines. The pool connects to those systems and to nothing else: declared routes only, default-deny egress, through Local Bridge. Camera feeds, captures, adapted model weights, and the record of every decision stay on the operator's hardware inside the operator's network.

An operator that wants a stricter posture for one site can have it. Posture is chosen per site, not per appliance. The same Czar runs air-gapped, with updates carried in as signed bundles, or connected to the systems the site already secures.

What the site receives

The site receives one pool, one subscription for the configured Grid that does not move when more robots join, and a record of every decision the pool made. Site tasks authored in Agency Studio are bounded by zone, speed, and distance, approved by named roles, and replayable afterward. When a tote goes to the wrong lane, there is a record to read rather than a story to reconstruct.

For how the pool fits with the rest of a logistics operation, see Solutions for logistics. For the facility that moves parcels rather than items, see An inference pool for a parcel hub.