An inference pool for electronics assembly

An electronics assembly plant has two properties that pull against each other. It is high mix, which means the parts, boards, fixtures, and work instructions change constantly and a robot has to recognize things it was not built knowing. And it is sensitive, which means the images that would teach a model what those things look like are images of a customer's unreleased product, and they cannot go anywhere.

A pool inside the building resolves that tension. Czar serves large perception and language models to the robots on the floor from the plant's own rack, learns from the plant's own captures on the plant's own hardware, and sends nothing out. The plant gets models that know its boards. The customer whose boards they are gets a plant that kept them.

Small parts, fine perception

Assembly robots in electronics do less lifting and more looking. A mobile manipulator carrying trays between the placement line and test has to confirm that the tray holds the right boards in the right orientation. A humanoid loading a fixture has to tell one connector housing from a nearly identical one. A robot at a rework station has to find the component it was sent to look at.

These are high-resolution perception problems with a thin margin for error, and the models that handle them are too large for the robot. The pool serves them over the facility network inside each robot's latency budget. The robot keeps its on-board model for the easy cases and for reflexes, and falls back to it when the pool cannot answer in time.

Changeovers and adaptation

Every changeover on a high-mix floor is a small new-product introduction. The fixtures move, the boards are a different shape, the trays are a different color. A model that was good yesterday is uncertain today.

Substrate adapts the perception models on the pool using captures from the plant's own cameras and robots, taken on the plant's own floor during the first builds of the new product. Nothing leaves. The adapted model is evaluated against the plant's own held-out captures and recorded before it serves a robot. When the next changeover comes, the plant does it again. Over a year of products, the pool holds a lineage of adapted models, each one on record, each one built from footage that stayed in the room.

Work instructions the robot can ask about

Electronics plants run on documents: travelers, work instructions, inspection criteria, engineering change notices. Operators are starting to ask the fleet about them in plain language, and the fleet is starting to answer. Where is the current revision of this instruction, what changed in the last notice, what is the torque spec on this fastener.

The language models that answer these questions run on the pool, and retrieval runs over documents the plant has placed on the pool. The index is the plant's. The answers are produced inside the building. An engineering change notice for a customer's product is never sent to a vendor's cloud to be understood.

What the pool connects to

An electronics plant is a connected site with narrow connections. The execution system has to know what was built, and each robot vendor's fleet software needs a view of its own machines. Local Bridge permits those declared routes and denies everything else by default. Camera feeds, captures, task records, and adapted weights stay on the operator's hardware inside the operator's network.

Some customer programs ask for more. A plant building for a customer with its own requirements can run the Czar serving that program's cell air-gapped, with updates carried in as signed bundles, while the rest of the plant runs connected. Posture is per site, and in a plant with distinct programs it can be per area. It is the same appliance either way.

Safety and the record

None of this changes the safety case of any robot on the floor. Stops, speed limits, and motion limits stay on the machine as the robot vendor set them. The pool serves perception and decisions. A pool that fails leaves a robot that stops.

Tasks for the floor's robots are authored in Agency Studio, bounded by zone, speed, and distance, approved by named roles, and recorded run by run. When a board is found damaged at test, the record shows which robot handled it, what the pool saw, and what it decided. That is a record an auditor can read, and it was produced without the board ever being photographed outside the plant.

The plant subscribes to the configured Grid as one line, fixed for the term. For the broader manufacturing picture, see Solutions for manufacturing. For a plant with a heavier existing control layer, see An inference pool for an automotive plant.