On premise AI for robotics: what stays on site and why

A camera above a robot cell sees your products, your packaging, your labels and, now and then, your people. The decisions made on those images describe how your plant works. Most companies would not email that to a stranger. Sending it to a cloud service is the same thing with better marketing.

Four reasons the data should stay

Confidentiality. Images of products before launch, batch codes, supplier names on labels. A log of decisions is a description of your process.

Regulation. Pharma, food and healthcare sites have rules on where data lives and who can see it. Keeping everything on site turns a compliance project into a non event.

Latency and uptime. A decision that waits for a round trip to a data centre is a decision the line waits for. A network outage should not stop a robot.

Control. When the software runs on your hardware, you decide when it updates, who logs in and when it is switched off.

What runs on site

Everything in the decision loop runs on a small box in the cell or in the plant network.

• Perception: camera and sensor data become structured facts.
• The rule engine: allowed actions and hard safety rules, as deterministic code.
• The decision model: one choice and a confidence, in a fraction of a second.
• The confidence gate and the operator console on a tablet.
• The audit log, with export to CSV and JSON.

Nothing in this list needs an internet connection to work. The box can be a compute module with Apple Silicon or an NVIDIA GPU, sized for the number of cameras.

What this asks of the model

Running on site rules out very large models. That is fine, because a decision model does not need to be large. It does not write text or hold a conversation. It chooses between a handful of allowed actions and returns a probability. That is a small, fast job, and a model built for it runs comfortably on a box the size of a book.

Being small has a second benefit: the model can be swapped. Any model that returns a choice and a confidence can take the place of the current one. You are not tied to one vendor’s model, or to one vendor at all.

Support and updates

Remote support is off by default. If you want us to look at a problem, you switch it on, for a defined scope and time, and switch it off again. Updates arrive as files. You install them when you choose, after a test on one cell if you prefer.

Questions to ask any vendor

Ask these before you sign anything, including with us.

1. Which data leaves the site, in normal operation and during support?
2. Does the system keep working when the internet does not?
3. Who can log in, and where is that recorded?
4. Can we export every decision, with its confidence and its approver, in a format we own?
5. If we stop using you, what stays on our hardware and what stops working?

If an answer takes more than a sentence, that is an answer too.