Skip to content
ELEMENT 31

Camera feeds are the building's data

A robot's cameras see the whole floor: layout, goods, pace, and people. When perception runs on a pool inside the building, that sight never leaves the operator's network.

· 4 min read

IN SHORT

Camera feeds from a robot fleet are the building's data because they record the building: its layout, its goods, its pace, and the people who work in it. On a Czar Grid, perception models run on the pool inside the facility, so the feed travels only over the operator's own network and is stored only on the operator's hardware. No frame leaves.

A mobile robot moving through a warehouse carries more cameras than most of the building's own security system. They point in every direction, all shift, at close range. Together the fleet's cameras produce a continuous, detailed film of the operation, and the natural question is who that film belongs to.

Element 31 treats the answer as settled. The feed records the building, so it is the building's data, and the Grid is built so that it never has to go anywhere else to be useful.

What a camera feed on a robot floor contains

It is easy to think of a robot's camera as a navigation sensor and nothing more. Consider what is in the frame. The layout of the floor, which the operator spent effort designing. The goods on the racks, their packaging, their volumes, their turnover. The pace of the work, which is a direct measure of throughput. The people, their faces, their movements, their breaks. On a plant floor, add the fixtures, the tooling, and the process steps that distinguish one manufacturer from another. At a port, add the vessels, the containers, and the yard plan.

A feed like that is not a sensor stream. It is a record of how the business runs, refreshed continuously. Anyone holding a copy could reconstruct a great deal about the operation without ever entering it. Treating it as incidental data is a category error.

Where the feed goes when the pool serves perception

Perception is the first thing a facility pool serves. The models that recognize objects, read the scene, and judge what is in front of the robot are too large to run on the robot and too urgent to send far away. On a Czar Grid they run on the nodes in the building's own rack.

The path is short. A robot captures a frame, sends it over the facility's own network to the pool, and receives an answer over the same wires. The frame does not transit any network the site does not own. It is not stored anywhere the site does not control. Whatever remains of it afterwards is on the operator's hardware and nowhere else.

That path is the same whether the site runs air-gapped or connected only to declared systems. The connected posture adds routes to the fleet software, the warehouse system, and the PLCs. It does not add a route for video, because the site did not declare one and Local Bridge denies the rest.

Captures and adaptation

Some frames are worth keeping. A pick arm meets an unfamiliar package shape; a humanoid on an assembly line sees a part oriented in a way the model has not handled well. The fleet saves these as captures, and captures are the raw material for making the models better at this floor.

On a vendor-cloud model, captures are uploaded, pooled with other customers' captures, and used to retrain a general model that everyone receives. The operator's floor becomes training data for the operator's competitors, a few frames at a time. On the Grid, Substrate adapts the models on the pool, using the site's own captures, and keeps the resulting weights on the site's own hardware. The floor teaches only itself.

This is also the point at which workforce concerns usually surface. The people in the frame are the operator's employees, and a site that keeps its feeds inside can give them a plain answer about where their images go. The answer is nowhere.

Ports and large outdoor sites

A terminal is a different scale of building, but the principle holds. Automated vehicles, cranes, and inspection systems on a large outdoor site produce feeds that cover the yard, the quay, the containers, and the traffic. The operator has reasons of its own, and obligations to its partners, to keep that imagery on site.

A Czar Grid at a port serves those vehicles and cranes from inside the terminal's own network. Feeds stay on site. The pool is the same appliance it is in a warehouse, in a posture the terminal chooses. Ports & Terminals describes that setting in more detail.

The practical consequence

Once the feed stays inside, several hard conversations get shorter. The question of whether a vendor's privacy policy covers the workforce becomes moot, because the pool never gives the vendor the frames. The question of what a partner or a customer might learn from the operator's imagery is answered by the fact that there is no copy outside the wall. The question of what an auditor can be shown becomes a tour of the rack.

Camera feeds are the most revealing thing a robot fleet produces. Keeping them in the building is not an advanced option. It is the default behavior of a pool that lives there. For the rest of the inventory, see What leaves the building: nothing. For how the pool serves a mixed fleet, see Czar Grid.

Questions this raises

Do the robot vendor's own cameras still send video to the vendor?
That depends on what the site permits the vendor's fleet software to do, and it is a separate question from the pool. The pool does not forward feeds anywhere. On a connected site, the vendor's software reaches the pool only through a declared route, and that route does not carry camera data out.
What happens to the feed when a model needs to be improved?
The fleet saves captures of hard or new situations, and those captures stay on the pool. Substrate adapts the model by fine-tuning on them there. The improved weights also stay on the operator's hardware.
Does keeping feeds inside slow perception down?
Keeping perception close is the reason the pool exists. The models that are too large to carry on the robot run inside the building rather than in a distant cloud, and the robot keeps its on-board model for reflexes and falls back to it when the pool cannot answer in time.

KEEP READING